AI security as a big opportunity is fairly market consensus – a number of early companies have been acquired by the incumbents and a set of other companies have been kingmade with large rounds out the gates. But despite the innovation over the past 4 years since the launch of ChatGPT, history as a guide shows us that the defining cybersecurity company of this generation likely hasn’t been started yet.

Cybersecurity markets tend to evolve in waves, with the first few waves absorbing the chaos around new terrain and unanswered questions (what does an ideal product look like? What needs to be built? What is the right form factor?) before it ultimately produces a generational business.

Cloud Security as a Case Study

Wiz was acquired by Google last year for $32B. It was the largest cybersecurity acquisition in the history of the industry, indicative of both the magic of the Wiz product (they are wizards after all!) and the enormity of the cloud opportunity. But it’s worth remembering that Wiz was founded in 2020, ~14 years after the advent of EC2 and S3. The canonical cloud security platform emerged 14 years after the inception of the cloud. And Wiz certainly wasn’t the first to identify that cloud security would be important – a number of cloud security companies were founded in that 14 year window.

Why did it take a decade plus for the category defining company to be started, let alone crowned?

The first generation of cloud security businesses were companies like Dome9, Evident.io, and Redlock among others. They were all started in the early 2010s (Dome9 in 2010, Evident in 2013, Redlock in 2015), in an era where the shift to cloud wasn’t fully certain and, hence, the language to define cloud security wasn’t fully there yet.

If you go back and look at the messaging of these companies in the early days, it’s clear that there wasn’t a coherent vocabulary to understand what cloud security meant.

Take the Dome9 website circa 2014 for example. The cloud wasn’t mainstream enough for “cloud security” to be a surefire budget line item for CISOs, hence Dome9 had to approach messaging by porting vocabulary from a previous era. Bulletproof firewall security for cloud servers. Even when seeing how they described their product (“keep ALL administrative ports closed on your servers without losing access and control, dynamically open any port on-demand”) reveals an on-prem network security mental model of the world vs a cloud one.

Evident.io and RedLock were both steps in the right direction on a product level. When you look at the product specs for both companies, there’s a lot of similarities with Wiz (complete and continuous visibility, agentless deployment, etc). And again, it’s impossible to run the counterfactual - both Evident.io and RedLock sold to Palo Alto Networks in 2018 (Evident.io had been around for 5 years, RedLock had been around for 3), so it’s hard to know if these two companies could’ve been longstanding billion dollar businesses had they stayed private.

But aside from the overwhelming tailwind of cloud demand which catapulted Wiz when they started in 2020, both Evident and RedLock were still subject to the inherent chaotic dynamics of cloud in the mid-2010s, which influenced their positioning.

On Evident’s website circa 2016, they clearly positioned themselves as security and compliance automation. RedLock’s website in 2017 emphasized compliance in addition to secure cloud migrations. So much of this is timing dependent: this was at a time where convincing large enterprises to migrate workloads to the cloud was still a large undertaking, and hence infrastructure companies had to attach themselves to those narratives accordingly. When Palo Alto bought both companies in 2018, they were seen as complements - the compliance capabilities of Evident with the analytical capabilities of RedLock.

Much ink has been spilled on why Wiz was able to win in the 2020s. Their revenue growth curve was one of the fastest in software history, reaching $100M of ARR in 18 months (this was pre ChatGPT and pre AI!), getting to 25% of the Fortune 100 in this first 1.5 years alone. What sticks out to me:

• At a technical level, Wiz differentiated from competitors like Orca Security and Lacework by having an agent-less, graph-based approach which made deployment incredibly simple and time-to-value orders of magnitude faster than existing solutions.

• They also benefited from operating in an environment where the uncertainty around cloud as a secular trend had dissipated, and every enterprise’s focus had shifted to real visibility of these workloads as quickly as possible. Total cloud spend in 2015 was $20B annually. AWS just announced a $40B quarter.

• But so much of their magic was in the marketing and positioning too– owning the category fully and unequivocally (their tagline being “secure everything you build and run in the cloud”), brand consistency around the Wizard persona, and just messaging in a way that was crystal clear and digestible by anyone (easy to say, hard to do).

The AI security wave cycle has only just begun

AI is obviously evolving faster than the cloud did, as shown by the rapid revenue ramps of the the large model labs and a number of AI companies (both inference and app layer!). And there is a certain element of early AI security that rhymes with early cloud security: a lot of M&A.

Just in the last few years:

• Cisco bought Robust Intelligence (August 2024)

• Palo Alto Networks bought Protect AI (July 2025)

• SentinelOne bought Prompt Security (August 2025)

• Cato Networks bought Aim Security (2025)

• F5 bought Calypso AI (September 2025)

• Check Point bought Lakera (September 2025)

• Crowdstrike bought Pangea Security (September 2025)

What’s crazy is that 6/7 of these acquisitions happened in a 56 day span (July 22nd, 2025 to September 16, 2025), and 4 of them happened in September alone.

All these players approached AI security from different angles: scanning machine learning models, detecting AI tools used within an organization, establishing an MCP gateway to monitor tool invocations, and guardrails to prevent sensitive data leakage among other product functionality. Much like cloud security in 2015, the language around AI security hasn’t fully stabilized yet. What it means to secure the AI applications you’re building (and how that differs from securing the AI applications you’re using, a la CASB) isn’t wholly agreed upon.

The AI problem surface itself is in a state of rapid evolution and change.

The first generation focused on securing the model, but to stop there would be to ignore where the risk truly sits: in the agent. MCP went from developer convenience to critical infrastructure almost overnight, with the vulnerabilities following just as fast: roughly thirty MCP-related CVEs in the first sixty days of 2026. It’s hard to standardize a vocabulary for a target that is still changing shape.

If the pattern holds, then the defining AI security company is likely yet to be started. And it will likely originate from a team that will do to AI security what Wiz did to cloud: show up at the exact moment the market finally knows what it wants, ship the product that looks obvious only in hindsight, and own the category so completely that they get to name it.

The company that wins AI security will be the one that gets to define what AI security means.

We don’t know who they are yet, and that is exactly why this is the most exciting time to be looking.

My teammate Kelley Mak and I will be at Black Hat next week and are excited to be at the frontier of cybersecurity’s intersection with this AI wave. If you’ll also be there, I would love to chat! Reach out to me at proby@work-bench.com.

We’re thrilled to announce that Work-Bench led Petra Labs‘ $5.2M Seed round, along with friends from Afore Capital, Pathlight, and strategic angels.

As ChatGPT, Claude, Gemini, and other AI systems become the front door for how people find products and services, brands no longer show up in a ranked list they can optimize against. They either appear in the answer or they don’t.

That shift has spawned an entire category of visibility tools that tell brands whether they’re being mentioned by AI models. Enterprises are already running multiple versions of these tools at once and churning through them quickly, because knowing you were mentioned doesn’t tell a CMO which of the dozen things their team did last month actually caused it, or whether any of it drove revenue.

There is a structural reason for why this is hard. For fifteen years, SEO ran on a bounded, enumerable set of inputs: backlinks, on-page content, keywords, technical site health, with measurable outputs and conversions from organic search. They behaved consistently enough to be solved by agencies and fragmented tooling.

The story changes with LLMs. When LLMs are predicting the next token based on statistical relationships learned across the public internet (e.g. YouTube, Reddit, PR, documentation, reviews, and community discourse), there’s no single AEO algorithm and the permutations across models, prompts, and channels is orders of magnitude larger than anything SEO ever produced.

That’s where Petra Labs comes in.

Petra is the first AEO partner to build custom last-mile attribution models for each client, connecting AI search activity to the business outcomes a client tracks and separating AI-influenced demand from everything else.

⏱️ Why Now

ChatGPT crossed 1 billion global monthly active users in May 2026. It’s the fastest app in history to hit that milestone, outpacing Google Maps, TikTok, Instagram, and YouTube. A Bain survey found that 56% of Millennials and Gen Z say they mostly or always use chatbots or a mix of chatbot and search. Bain also found that AI overviews inside search itself now make 70% of searches “zero-click.”

Both paths, chatbots and search, are converging on the same outcome for marketers: less of the traffic and ranking signal they used to measure themselves against. At the same moment discovery is fragmenting like this, the rest of the measurement stack is collapsing. Platforms are increasingly grading their own homework, restricting third-party access, and pulling analytics in-house, leaving marketers with no equivalent of a Google Search Console or a Meta Ads Manager for AI search.

🔥 The Product

From visibility dashboards to a closed attribution loop.

Petra Labs has built three capabilities to close that loop:

1. An attribution engine. A system of record for every intervention a marketing team takes to influence AI visibility, whether it be a PR placement, a YouTube video, a Reddit thread, a page update, and connects it to changes in AI recommendations, site traffic, and downstream revenue.

2. LLM simulation and visibility intelligence. Large-scale simulations across AI models using high-intent, purchase-driven prompts, capturing which vendors get recommended, which citations get used, and how retrieval actually behaves.

3. Enterprise classification and citation dynamics. A classification layer that maps brands, products, and features in LLM outputs at the right level of abstraction, while tracking which sources are gaining or losing influence over time.

No two companies measure outcomes the same way, which is exactly why this can’t be solved with off-the-shelf software. A rise in site traffic signals a qualified buyer for one business and noise for another. Branded and organic demand carry different weight from one company to the next. And the metric that matters most, whether it’s closed revenue or qualified pipeline, differs across every account. That’s why Petra Labs builds a custom last-mile attribution model for each client rather than shipping a one-size-fits-all dashboard.

Petra Labs embeds directly into each customer’s data and attribution environment through a forward-deployed model, translating a company’s internal ontology (products, pages, segments, conversions, revenue) into Petra Labs’ measurement model. An embedded team of operators, powered by the Petra Labs Agents Platform, identifies the prompts a brand needs to win, tracks visibility and sentiment across every major model, and executes across earned, owned, and social media on the client’s behalf.

💪 The Team

CEO Sami Akkawi spent his career in strategy consulting at Strategy&, where he ran thousands of interviews with senior marketing leaders at the world’s largest brands and the same complaint kept surfacing: nobody had a credible way to connect marketing activity to what AI models were telling their customers. COO Ramie Khoury is an engineer-turned-sales-leader who led strategy and operations at Traba. CTO Rishi Penmetcha brings engineering experience from Palantir, Ironclad, and Traba.

Petra Labs got its start as a fully managed AEO service, with the team owning 100% of strategy, execution, and attribution for its earliest customers by hand before building the software to do it at scale. The product is opinionated the way it is because the founders spent their first few months running the process manually, against real budgets, before productizing their approaches.

The team is fully based out of New York, in the same city as many of the enterprise marketing organizations they sell to. That proximity is part of the operating model. They’re showing up in person, sitting inside a customer’s actual workflow, and bringing real outcomes.

If you’re a CMO, growth leader, or performance marketing executive stitching together visibility tools, spreadsheets, and gut instinct to make sense of how AI search is reshaping your funnel, it’s time to see what a real attribution layer looks like. Reach out to the Petra Labs team to see how they can help.

Welcome to the Work-Bench family, Petra Labs! 🚀🚀🚀

Here is what I heard and where we see the industry moving:

• Generative design is real, but requires trustworthy downstream integration. Parametric, code-compliant layout generation is advancing quickly and moving from CAD to computer-generated design. The gap forming is reliability in a high-stakes environment where a confident hallucination costs time, capital, and eventually safety. On top of that, a generated floor plan only creates value if it becomes a single source of truth for purchasing, manufacturing, framing, and sequencing. One BIM integration lead put it that design’s value is measured downstream. We’re watching for the layer that connects generated models to execution across procurement, sequencing, and field coordination, which goes beyond just the design tool itself.

• Subcontractor selection is a judgement problem. Current tooling is optimized for background checks: financial health, safety records, compliance. Those are necessary but not enough. The harder problem is matching a specific subcontractor to a specific job, site, crew, and moment, which is an assessment that lives partly in rubric and partly in experiential judgment that hasn’t been formalized. On top of this, the relevant data is fragmented across systems, email, text messages, and personal devices. Solving sub management requires decision intelligence infrastructure that can synthesize a multitude of structured and unstructured signals into contextual recommendations.

• Jobsite monitoring is abundant yet timeline intelligence isn’t there. Cameras, drones, wearables, smart glasses, exosuits, instrumented hard hats, and quadruped robots scanning hundreds of thousands of square feet are all deployed now. What remains open is the last mile of stitching that data into a coherent, real-time view of schedule risk. A scan tells you what the site looked like at one moment. What comes next is automated updates on whether the project is behind, at risk, or ahead. The opportunity we see is tying observational data back to schedule, cost, and sequencing logic.

• The build-vs-buy inflection is accelerating, except where proprietary lock-in holds. One technology leader described an internal engineering team using agentic development, running their own GPUs, and replacing legacy platforms (including Procore) with cheaper, more capable in-house tools. The switching costs in the ecosystem have dropped meaningfully. The one exception was Oracle Primavera P6 because owners contractually require schedules in P6’s proprietary format, making it essentially irreplaceable regardless of internal engineering capacity.

Wherever contract or regulatory structure creates a proprietary format lock, legacy vendors retain pricing power. Everywhere else, the barrier to custom software is falling, which threatens incumbent SaaS and opens space for these vertical infrastructure plays.

The opportunity

The demand pulling construction forward, data centers especially, has created more urgency and capital behind these problems than the industry has seen in years. The consistent pattern is that AI can generate outputs, yet construction still runs on fragmented data, unformalized judgment, and proprietary formats.

There’s clearly much opportunity ahead in the built world. We’ve been fortunate to back incredible companies at the forefront of this movement, including Nox Metals and Congrue. We’re always on the lookout for founders with a deep earned perspective in construction tech. If this is you, please reach out!

Dear Friends of Work-Bench,

What a sprint the first half of 2026 has been 🏃

We’re very excited to partner with several exceptional new Seed companies—more on those soon. In the meantime, we’re thrilled to celebrate two meaningful fundraising milestones from the portfolio:

🩺 Courier Health raises $50M Series B led by Oak HC/FT

After leading their Seed, we’re thrilled to continue backing Courier Health through its $50M Series B. Courier Health is building the operating system for patient services across biopharma, helping life sciences companies coordinate the complex journey from prescription to treatment. Up to 70% of specialty patients abandon treatment within a year—Courier helps close that gap, driving 15%+ increases in patient starts while building biopharma’s system of record for patient engagement.

🏛️ GovWell raises $25M Series A led by Insight

After leading their Seed, we’re excited to continue backing GovWell through its $25M Series A. GovWell is modernizing the software that powers local government, replacing decades-old systems with an AI-native platform that helps municipalities process permits, licenses, planning, and code enforcement dramatically faster. Today, GovWell helps municipalities across 35+ states complete processes like permitting up to 95% faster, delivering a better experience for both public servants and residents.

We’ve also spent the first half of the year doing what we love most: bringing together founders, operators, customers, and future builders across NYC.

• Continued our Best of NYC series, bringing together top operators across engineering, product, GTM, AI, and more

• Hosted Future Founder Dinners with builders we believe will create the next generation of enterprise companies

• Hosted four Customer Dinners for our portfolio companies, helping generate more than $1M in qualified pipeline

• Convened Fortune 500 leaders around topics including AI Infrastructure and AI Agents in Financial Services

As always, we so appreciate your support in sending us great b2b founders in NYC and beyond. We love meeting founders as early as possible to back their pre-seed and Seed—please keep sending them our way!

We often get asked what we’re seeing in the market; here are a few observations from the past quarter.

Is this a venture-scale business?

AI has dramatically lowered the cost of building software. More companies are getting started than at any point we can remember, and many are reaching meaningful early revenue with fewer people and less capital than ever before.

But the question we find ourselves asking founders more than ever is surprisingly simple: is this a venture-scale business?

As AI continues to commoditize software creation, we’re spending even more time underwriting durable moats. Does the customer gain a meaningful competitive advantage from buying this product? Is there proprietary data, workflow ownership, distribution, or execution that compounds over time?

The next generation of founders feels more intentional

One of the most encouraging trends we’ve noticed has come through our Future Founder Dinners.

Rather than chasing the latest AI capability, we’re meeting more founders who are spending time developing unique insights before writing a single line of code. They’re choosing problems they want to dedicate the next decade to—not simply the next fundraising cycle.

We think that’s incredibly healthy for the ecosystem—and one of the reasons we’re investing even more in building relationships long before companies are formed.

In an AI-first world, GTM may be the only moat

The biggest shift we continue to see is that AI is compressing the value of software while making go-to-market execution even more important.

This quarter we launched our Founder Sales Cohort, bringing together a small group of NYC Seed founders for an intensive working session led by David Politis, our CEO-in-Residence and former Founder & CEO of BetterCloud.

As products converge faster and categories become increasingly crowded, the founders who consistently win aren’t simply building better products—they’re building stronger brands, tighter customer communities, and more repeatable sales motions.

Founder brand, customer community, curated events, customer references, and thought leadership are increasingly becoming competitive advantages—not simply marketing tactics.

For Work-Bench, that reinforces what we’ve believed since day one: our job isn’t simply to write checks—it’s to help founders build category-defining companies.

📚 Recent Writing

🧠 H1’26 AI Snapshot — We rounded up the biggest AI developments from the first half of the year. Here’s what stood out: (1) code authorship has shifted from human to machine, with “software factories” now merging thousands of AI-written PRs each week; (2) as models commoditize, the “harness” around them is becoming the real moat; (3) tokenmaxxing is pushing the ecosystem toward a FinOps-style reckoning on AI spend; (4) OpenClaw’s breakout made agents outside the chatbox mainstream while exposing a new security attack surface; and (5) SaaS itself is being rebuilt for a world where the primary “user” may be an agent rather than a human.

⚙️ Not All Software Is Created Equal — AI is turning code into a commodity, but that doesn’t mean all software is equal. The winners won’t be the companies that ship the most code—they’ll be the ones creating the most irreplaceable business transformation with it.

🏦 The Real AI Opportunity in Financial Services — Banks spend more than $600B annually on technology, yet 43% still run on COBOL. AI is one of the hardest technologies to deploy in financial services—and one of the most valuable once it lands.

🏥 The Real AI Opportunity in Healthcare — The U.S. spends $5.3T on healthcare each year, yet nearly a quarter of every dollar goes toward administration instead of patient care. AI may finally be the force that changes that equation.

📰 Catch Us in the Wild

Wall Street Journal — Jon shared his perspective on the NYC vs. SF debate and why he believes the NYC enterprise ecosystem has never been stronger—from increasingly technical founders to deeper operator talent and unmatched access to enterprise customers.

Not Another CEO Podcast — Jess joined David Politis for an unfiltered conversation about how investors actually build conviction, why many investment decisions are made within the first five minutes, and what it really takes to earn a Series A in today’s market.

Tokenmaxxing is splitting the world into haves and have nots — Danny shared his perspective on how the companies spending most aggressively on R&D tokens are growing revenue much faster than those throttling it (more R&D spend → better products → more sales), and why enterprises that just speed-ran a decade of cost management and FinOps still aren't sure how to respond.

The hardest part of formal verification is knowing what to prove - Kelley shared his perspective on formal verification, drawing on conversations with practitioners across quant funds, banks, and fintechs—the exact environments where it could have the biggest impact.

Bold predictions: Ramp Labs is to Ramp what AWS is to Amazon — Proby shared his prediction that Ramp Labs will either generate more profit for Ramp than its core business within a decade or spin out entirely, pointing to AWS as a case study. AWS took just over a decade to become more profitable than Amazon's core retail business.

👋 Come Hang With Us

🎉 Cofounders.nyc — Join our curated, in-person community for NYC builders looking to meet future co-founders and early teammates.

🌱 Future Founder Dinner — Thinking about starting a company in NYC? We’d love to meet you. We host a quarterly dinner bringing together top builders to swap notes and share ideas. Reach out to Danny to learn more about joining a future dinner.

‍Join us July 1 at 1pm ET as we walk through our Work-Bench AI Snapshot report live. Sign up for the webinar here.

Four years ago, ChatGPT felt like magic. Today, software is starting to build itself. The pace of change has been staggering: AI went from answering questions, to writing code, to completing entire workflows with minimal human involvement. The gap between thinking and doing is shrinking. Agents are escaping the chatbox and becoming coworkers, operators, and increasingly, users of software themselves.

For founders, operators, and investors alike, many of the assumptions that defined the SaaS era are being rewritten in real time. As early stage investors in enterprise software, we’re constantly studying these shifts. We don’t pretend to know what AI will look like six months from now, but we can be students of the present. This report is our attempt to capture the state of AI in June 2026: what’s happening at the frontier, what patterns are emerging, and where we believe the most important opportunities and challenges are beginning to take shape.

Here are a few themes:

The Author of Code Has Shifted from Human to Machine

When early instantiations of AI coding products like GitHub Copilot came to market, it was simultaneously obvious that this would be something big yet impossible to know how fast the change would occur. The early days of AI coding was an LLM being an accelerant to a developer’s workflow; now the LLM is the developer.

This has also changed the form factor that AI products have taken. We now have “software factories”, which are a set of coding agents working asynchronously in the background to create software. The behavior change this has inspired is extraordinary! Engineers will send a Slack message to the coding agent describing a change they want to make while cooking dinner, and by the time they’re eating their food - the change would’ve been made.

While it’s still super early days, the amount of engineering activity that has begun to be routed through these software factories is astounding: 57% of Ramp’s PRs were written by their background agent Inspect, while Stripe’s Minions and Spotify’s Honk both merge over 1,000 PRs on a weekly basis.

That said, shipping faster has created real reliability questions. An AI agent destroyed one company’s production data and admitted to it. The industry is still figuring out how to have both speed and trust.

The Harness

The last few years have shown us that while the model is important, the model alone isn’t enough for great applications. The harness, which is everything outside of the model, is incredibly key too.

One can think of the relationship between the model and harness as similar to that of the semiconductor and operating system. LLMs are very much like semiconductors: horizontal pieces of technology that can be used for a wide variety of purposes. It’s the harness that truly makes it production-usable and translates those capabilities to certain tasks.

What this implies for the longer arc: when frontier models commoditize, the harness is what matters. The moat moves from the intelligence to the scaffolding around it.

The Token-Maxxing Party

There’s a really good mental framework called Goodhart’s Law. It claims that when a measure becomes a target, it ceases to be a good measure. And we’ve seen this play out with the token-maxing phenomena in AI. While it’s obvious that people who experiment with new technology early in a paradigm shift tend to be better off than those who wait on the sidelines, letting a single number (tokens) define effective AI use and having that be hyper-optimized for across an entire ecosystem is bound to have an impact.

What comes next looks a lot like what happened with cloud spend. It took a decade for cloud FinOps to fully mature: telemetry, attribution, cost governance, unit economics per workload. AI is speedrunning that same cycle.

The instrumentation stack for managing AI spend is starting to form!

OpenClaw and CUA

OpenClaw was a watershed moment for AI. Its nature of being always-on, residing wherever the user wanted to use it from (WhatsApp, Slack, iMessage, etc), and having access to a user’s machine and files made it a complete expression of what’s possible when agents broke out of the chatbox. OpenClaw became the fastest growing open source project in the history of software, now sitting at over 370K GitHub stars.

But as our friends in the cybersecurity world know, with every expansion in technological capability comes an expansion in the attack surface. With computer use agents being able to click on a screen and drive actions like a human would, without needing an API, there’s a gap in visibility that IT can’t see. It rhymes with the early days of SaaS, when business users could download software on the cloud without direct IT approval. This gap in the cloud era led to the rise of CASB, with great companies such as Netskope and Zscaler, and we expect it to produce some great AI security companies as well.

Exciting Tech at The Frontier

One theme we’re fascinated by is the emergence of continual learning. This is defined as the ability for an AI system to improve from its own experience after training ends, the way an employee gets better at a job by doing it. While continual learning has traditionally been thought of in the realm of the model, we’re beginning to see the early innings of it being applied at the harness layer as well.

Case in point, Harvey has started to run an autoresearch loop to enable its harness to self-improve. After each run, the LLM-as-a-judge system grades performance, and hypotheses around why failures happen lead to the harness being edited accordingly. This led the harness to improve in emergent ways that couldn’t have been predicted by the Harvey team.

This points toward what we think is the most interesting question in enterprise AI right now. What does it mean to own your loop? The workflows your team runs, the evals you build, the institutional memory you accumulate. That context is yours. When models commoditize, the companies that have been quietly compounding their loop will have built something that travels with them regardless of which model they use.

It’s still so early. Excited to spend the rest of 2026 continuing to study the bleeding edge of AI and learning from those who are living and breathing it.

Read the full report at our blog here.

The US has more than 90,000 local governments. Together they spend roughly $1.8 trillion per year on essential services, much of it running on software that hasn’t meaningfully changed since the early 2000s. Permitting, licensing, code enforcement, planning and zoning: these are the processes that determine whether a house gets built, a business opens its doors, a fire safety inspection happens on time. The operating system running underneath all of it is from a different era.

The dominant players, Tyler Technologies, Accela, and OpenGov, were built when on-premise deployments and monolithic architectures were the norm. Many municipalities still operate on paper-based systems or software so rigid that any configuration change requires consultants and months-long timelines. The result is a perverse irony: the institutions tasked with enabling community development are themselves constrained by the tools meant to help them do it.

Enter GovWell. Troy LeCaire saw the problem firsthand working in the U.S. Senate. Ben Cohen lived the other side, helping his father navigate the contractor’s experience of municipal bureaucracy. When they connected and started talking to hundreds of municipalities, they saw universal pain with high willingness to change.

The GovWell platform covers permitting, licensing, planning and zoning, and code enforcement on top of its AI operating system. Their AI Community Assistant gives residents and contractors instant, 24/7 answers to questions about zoning and municipal services. Their AI AutoCheck reviews applications for errors and compliance issues, catching problems early that would otherwise bounce back and forth for weeks.

The results: GovWell powers 5,000+ mission-critical processes across 35+ states. Implementations that take legacy vendors 12-18 months are going live in weeks. Hampton, Georgia’s community development director told her city council that business license processing went from 10 days to under an hour.

Today we’re announcing that Insight Partners is leading GovWell’s $25M Series A, with participation from Bienville Capital and GovTech operators including David Reeves (former President and CRO of OpenGov), Andreas Huber (Founder and CEO of First Due), and Chris Bullock (Founder and former CEO of ClearGov). We’re doubling down after leading their seed in 2024. Local government was always supposed to work for the people it serves. It just never had the software to make that real.

Troy and Ben believe that how government works matters, and that people who dedicate their careers to public service deserve tools as good as the ones in the private sector. We couldn’t be more excited to continue this journey with them.

Check out more on the GovWell blog and their launch video!

Read More Here

Every year, nearly 7 billion prescriptions are dispensed in the United States. Of the patients prescribed specialty medicines, nearly 30% never start therapy and 60-70% abandon it within the first year. The drug works, but the patient just never gets to take it. That failure is operational.

Biopharma has built sophisticated commercial organizations to support these patients: field access teams, patient services teams, marketing. But the work is performed across spreadsheets, legacy CRMs built for sales reps, hub portals, and email threads. The patient journey gets fractured across a dozen tools that don’t talk to each other.

Enter Courier Health. Danny Sigurdson spent years leading sales at ActionIQ, where he watched enterprise after enterprise struggle to unify fragmented customer data. He realized biopharma had the same problem with dramatically higher stakes: when a biopharma company loses track of a patient, that patient might lose access to a therapy they need to survive.

Courier Health is the first solution designed entirely around the patient. It ingests the data streams that matter (specialty pharmacy feeds, hub data, claims, EHR signals, field activity), stitches them into a single patient record, surfaces intelligence on at-risk patients, and orchestrates compliant action across teams.

The results: The company quadrupled the number of therapies on its platform in 2025. Customers are achieving an over 15% increase in patient starts, 10-12% increase in 6-month persistency rates, and 5-7x increase in team productivity. At Denali Therapeutics, the platform shifted the focus from rigid process design to patient outcomes. At SetPoint Medical, their VP of Market Access called it “the next generation of life science patient navigation technology.”

Today we’re announcing that Oak HC/FT is leading Courier Health’s $50M Series B. We’re tripling down on our investment, having led their seed round in 2021. We’ve long believed at Work-Bench that the best enterprise software companies reveal a better operating model that was always theoretically possible but practically unreachable without the right infrastructure. Courier Health is doing exactly that. The patient journey was always supposed to be coordinated and patient-centric. It just never had the software to make that real.

Danny and the Courier Health team are building something that quite literally changes and saves lives. We couldn’t be more proud to continue backing them.

Read More Here

Before committing to the startup journey, the most valuable work a founder can do is understanding a market - how it works, who the stakeholders are, and where real problems exist.

Historically, mapping a market, understanding its stakeholders, and identifying real pain could take weeks of manual research. Today, AI can dramatically accelerate the process - if you know how to use them properly.

Join us for a Work-Bench Masterclass with our Venture Partner Diego Oppenheimer, where he’ll walk through a practical framework for using AI to research markets and evaluate startup ideas. Diego will also share the exact AI prompts that he uses for market research.

In this session, Diego will cover:

• How to choose the right industry “haystack” to explore

• How to use AI to map a market and identify stakeholders

• How to uncover real problems worth solving

• How to test early solution hypotheses before writing code

This session is designed for:

• Operators exploring startup ideas

• Builders researching potential companies

• Founders validating a new market

• Anyone curious about structured startup discovery

  ---

🗓️ Wednesday, March 18th
⏰ 1:30-2:30PM ET
💻 On Zoom

Register

A Bit About Diego Oppenheimer

Diego previously founded Algorithmia, an AI infrastructure company acquired by DataRobot.

He is now a Venture Partner at Work-Bench and an Executive Fellow at Stanford GSB, where he assists in teaching AI for Research-Driven Inspiration, a course focused on systematic startup discovery.

Developers spin up Claude Code instances like they’re opening new browser tabs. Marketing teams write everything in ChatGPT first. If you don’t know Cursor, you’re behind. And these are just a handful of examples across the enterprise.

The data backs it up: 57% of companies have AI agents in production, according to G2’s 2025 survey. Another 22% are piloting them, while the rest are still in the experimental phases before that.

But it’s still not that simple: 65% of leaders cite agentic system complexity as the top barrier (for two consecutive quarters), according to KPMG’s Q4 2025 AI Pulse Survey. Gartner thinks 40% of agentic AI projects will get canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls.

While enterprises have no shortage of ambition, they lack the right infrastructure.

We’ve been calling this the agent runtime: the infrastructure layer being built to safely execute, constrain, observe, and improve agent work at scale. Here’s how we’re thinking about it.

Why Now

If agents have been technically possible for a while, what changed in 2025 to make the infrastructure gap undeniable?

In short, agents graduated from demos to daily drivers:

• Claude Code went from research preview to general availability.

• Cursor documented agents operating autonomously for close to a week, writing over a million lines of code across 1,000 files, to create a web browser from scratch.

• OpenAI’s Codex sustained 24+ hour autonomous sessions.

• Gartner predicts 40% of enterprise applications will embed task-specific AI agents by the end of 2026, which is up from less than 5% in 2025.

Our use of agents has increased dramatically without proportional adaptations in the infrastructure needed to support it.

Then there’s the cost. A mid-sized e-commerce firm building an agentic supply chain optimizer, for example, saw infrastructure costs jump from $5K/month in prototyping to $50K/month in staging (a 10x increase), due to unoptimized RAG queries fetching far more context than needed. Budget constraints aren’t intuitive to an LLM; they need to be enforced architecturally.

And on top of all that, security incidents make the risks concrete. CyberArk Labs demonstrated how a financial services company’s procurement agent was weaponized through a malicious prompt embedded in a shipping address field, where the agent used its invoicing tool access to compromise sensitive data. Palo Alto Networks’ 2026 predictions put it bluntly: with “a single well-crafted prompt injection or by exploiting a ‘tool misuse’ vulnerability,” adversaries now “have an autonomous insider at their command.”

The fundamental issue is that agents are nondeterministic. Both the input and output space for an LLM are infinite, and the same input doesn’t necessarily guarantee an identical output. They execute untrusted code determined at runtime, so there’s no build-deploy-test cycle. Attribution is blurry when agents complete tasks across systems with multiple tools.

And they can fail invisibly: An agent that confidently returns a wrong answer, loses context mid-conversation, or takes an action misaligned with user intent doesn’t throw errors because from the infrastructure perspective, everything worked.

Early adoption comes with risk. Agents demand a new runtime to make that risk manageable.

The 4 Pillars of the Agent Runtime Stack

• Execute. The systems (sandboxes, skills) that enable agents to take safely action.

• Constrain. The guardrails (permissions, identity) that keep agents operating within defined boundaries.

• Observe. The visibility and monitoring infrastructure that tracks what agents are doing.

• Improve. The feedback loops and learning mechanisms that make agents better over time.

Continue reading on our blog.

We are thrilled to announce that Work-Bench led Flock’s $6M Seed round. The round includes participation from January Ventures, Red Bike Capital, Outlander VC, and strategic angels.

The round comes as retail brands face a fundamental shift: ecommerce success now depends on delivering personalized, high-quality visual storytelling at scale, something traditional photoshoots and first-generation AI tools were never designed to support.

Traditional photo shoots are slow, expensive, and fundamentally limit how much content brands can produce. They were built for seasonal campaigns, not continuous content generation across massive catalogs.

At the same time, generic AI image tools have failed to earn enterprise trust. While they can generate images quickly, they lack product accuracy, brand safety, and the workflow integration creative teams require. For large brands, “almost right” isn’t good enough.

That’s where Flock comes in.

Flock is building the AI-native content engine for retail, replacing traditional photo shoots with brand-accurate imagery and video at scale, while fitting directly into how creative and merchandising teams already work.

Acting as a creative team’s co-pilot, Flock allows brands to generate tailored, brand-accurate product imagery for every customer, channel, and context, without expanding headcount or budgets. Using reinforcement learning and brand-specific AI models, Flock enables one-to-one ecommerce experiences that were previously impossible, while freeing creative teams to focus on the tasks that move the needle.

Why Now

Retail content is hitting a structural breaking point.

Product catalogs are expanding. Channels are fragmenting. Personalization expectations are rising. Yet content production capacity remains constrained by cost, timelines, and headcount.

Retailers are under margin pressure and being asked to do more with less. They need higher content velocity without sacrificing brand integrity or increasing spend.

First-generation AI image tools educated the market on what’s possible - but also revealed their limits. Generic outputs, brand inconsistency, and lack of workflow integration left enterprise retailers searching for vertical, production-grade solutions.

Creative teams are willing to adopt AI when it doesn’t ask them to compromise quality, control, or process.

This convergence of pressure, maturity, and readiness is what makes this moment different.

The Product

From photo shoots to content infrastructure.

Flock is not an AI image studio or a standalone creative tool. It is a full-stack content system designed to replace the parts of retail content workflows that no longer scale.

At the core of the platform is Flock’s brand DNA dataset.

Every creative approval feeds into more than 200 brand-specific parameters that encode a company’s visual language, style rules, and constraints. Approved assets are pushed directly back into existing DAM and ecommerce systems.

Unlike static AI tools, Flock uses reinforcement learning to learn from creative team feedback, approval patterns, and real customer conversion data. Over time, this creates a compounding advantage where output quality, brand alignment, and performance improve with every asset generated.

This allows teams to:

• Generate brand-accurate ecommerce, editorial, and video content at scale

• Maintain control over every element lighting, pose, background, fabric texture, and detail down to every stitch line and zipper

• Reduce production timelines from months to days without changing creative workflows

• Produce inclusive, realistic imagery across skin tones, body types, ages, and textures

By turning content creation into a repeatable production pipeline, Flock enables a shift from one-off assets to continuous content infrastructure.

Traction

From pilots to real production impact.

Flock is already working with billion-dollar fashion and beauty brands with rapid expansion across product catalogs and channels.

What stood out to us is how quickly the platform moves from experimentation to impact.

Customers report up to 90% cost savings compared to traditional photo shoots, alongside 30%+ conversion lift driven by better coverage, faster iteration, and more relevant content.

Because Flock plugs directly into existing ecommerce and DAM workflows, teams can adopt it without re-architecting how they work. That combination of immediate ROI and deep workflow fit is what gives Flock a path to becoming durable infrastructure rather than an optional tool.

The Team

Built by insiders solving their own problem.

CEO Vita Mallela spent years leading merchandising and creative operations at Bloomingdale’s and Jet.com/Walmart, managing multi-billion-dollar P&Ls and experiencing firsthand how slow and expensive traditional photo shoots can be.

She’s building the platform she wished she had.

CTO Malavika Reddy brings over a decade of AI and computer vision experience from Microsoft and Brightcom, where she built and scaled production machine learning systems.

Together, they bring a rare combination of operator empathy and technical rigor. They understand exactly what creative and merchandising teams need because they’ve lived the problem themselves.

Why Work-Bench Is Excited

This is the kind of investment we get excited about: a team solving a mission-critical problem they’ve lived firsthand that sits directly on the path to revenue and brand trust.

Across diligence, the feedback from large retailers was remarkably consistent. A multi-brand apparel company told us the results versus their prior solution were “night and day.” A global luxury retailer that evaluated nearly a dozen vendors said Flock was “the only platform we tested where the images actually looked real.”

After testing dozens of pilots, and in side-by-side evaluations, teams were especially struck by how consistently Flock handled the details that broke other tools. Competing solutions struggled with prints, adding buttons that didn’t exist or losing pockets entirely. Flock was the only platform customers tested that reliably produced correct creases, folds, and shadows, handled complex patterns and prints, and maintained model consistency without creating exact duplicates. In one case, the output quality was so realistic that a marketing team initially didn’t believe the images were AI-generated.

That quality is already translating into real production impact. At one enterprise retailer, teams generated “5,000-6,000 images in a single month” to launch a new shopping experience, and saw a “roughly 3 percentage point lift in conversion since launch.” Another large brand told us they are now replacing meaningful portions of traditional photography rather than treating AI as a side experiment.

Execution matters just as much as output. Multiple customers emphasized that Flock works within existing creative and merchandising workflows and comes with hands-on support when edge cases arise. As one team put it, “it’s not just a tool, we need someone who understands our workflows and can troubleshoot in real time.”

With a natural path from imagery into video, personalization, and broader commerce workflows, we believe Flock is becoming foundational infrastructure for modern retail.

What’s Next

If you’re excited about building AI-native infrastructure for modern retail, now is the time to join. Reach out to the Flock team to learn more about open roles and how you can be part of what they’re building.

We’re excited to support Flock as they grow the team behind the future of retail content.

Welcome to the Work-Bench family, Flock! You can check out their press coverage in Vogue Business and PR Newswire.

This raise comes less than a year after Goodfire’s Series A and will fuel frontier research, the next generation of the core product, and expanded partnerships across AI agents and the life sciences.

Our investment in Goodfire at its heart is rooted in community & long-term conviction. We’ve known Eric Ho, co-founder and CEO, for more than eight years, first backing him at his prior company, RippleMatch, where he was a co-founder and CTO. It was there that Eric met Daniel Balsam, now his co-founder at Goodfire. In addition to Eric and Daniel, Goodfire rounded out the founding team by adding Tom McGrath a brilliant research scientist and cofounder of Google DeepMind’s Interpretability team.

Everything that the Goodfire team has built in such a short time is remarkable. In the earliest days, Eric and Danny from our team were whiteboarding ideas, exploring opportunities, and debating what the future of interpretability, AI infrastructure, and agents could look like. Watching those early conversations evolve into what Goodfire is today has been incredibly rewarding.

Less than two years later, Goodfire has assembled a world-class team drawn from frontier labs and top research universities, and partnered with industry leaders including Arc Institute, Mayo Clinic, and Microsoft. Along the way, they’ve developed novel techniques to decompose model internals, cut LLM hallucinations in half using interpretability-informed training, and even identified a new class of Alzheimer’s biomarkers by reverse-engineering a biological model.

This round is a major milestone, but still just one step on a much longer journey. We’re incredibly proud to be early partners and can’t wait to see what the team builds next.

You can read more about Goodfire’s Series B on their blog, as well as in the accompanying press coverage from Bloomberg News.

Building the AI System of Record for In-House Legal Teams

We are thrilled to announce that we led Chamelio’s $5.7M Seed round, following their pre-seed led by Emerge, bringing their total funding raised to $10M.

In-house legal teams sit at the center of modern businesses. They’re asked to accelerate revenue, manage risk, and scale operations while operating across thousands of contracts, policies, and workflows that were never designed to be analyzed, connected, or automated.

Still, most legal teams rely on a fragmented tech stack that includes legacy CLMs that behave like storage systems, generic AI tools that lack context, and manual workflows stitched together with spreadsheets and email.

That’s where Chamelio comes in.

Chamelio is building the AI context layer for in-house legal teams with a platform that transforms contracts, policies, and negotiation history into a living intelligence layer to power end-to-end legal workflows.

Why Now

Legal teams are drowning in data but starved for intelligence.

For the last decade, contract lifecycle management (CLM) has been treated as the core system for legal teams, but in reality, most CLMs are just glorified document repositories. They’re expensive storage layers wrapped in workflow theater.

They store contracts, they don’t understand them, and they certainly don’t learn from them.

At the same time, the rise of generative AI has produced a wave of “legal copilots” that can draft clauses and summarize contracts, but they are fundamentally stateless. They’re disconnected from a company’s actual legal history, negotiation patterns, and risk posture.

The result is a paradox:

• CLMs have data but no intelligence.

• AI tools have intelligence but no data.

• Legal teams are stuck stitching together brittle workflows in between.

That model is breaking.

Modern in-house legal teams are under unprecedented pressure. Contract volume is exploding. Deal cycles are accelerating. The business expects legal not just to reduce risk, but to enable growth, without adding headcount.

For the first time, AI is powerful enough to change this. But only if it’s embedded inside a system that understands a company’s legal memory: what it has signed, how it negotiates, where it bends, and where it doesn’t.

That’s the shift Chamelio is betting on.

Not another CLM. Not another chatbot. But a new category: an AI-native legal operating system that turns contracts from static documents into a living intelligence layer.

We believe this is the moment the CLM category begins to unbundle - and the legal OS category begins to emerge.

The Product

From document storage to legal intelligence.

Chamelio addresses the full spectrum of legal work from intake and drafting to negotiation, signing, and post-signature operations, all while keeping legal firmly in control.

At its core, Chamelio unifies three core capabilities that have historically lived in separate tools:

1) Best-in-class contract redlining.
Chamelio integrates directly into Microsoft Word, enabling legal teams to run playbook-based reviews on complex agreements, refine standards independently, and scale negotiation without adding headcount.

2) Low-risk contract automation.
For simpler agreements like NDAs and vendor contracts, Chamelio’s AI agent drafts addenda, clauses, and revisions faster than generic LLMs while maintaining guardrails that legal teams trust.

3) Stateful legal intelligence.
Beyond redlining, Chamelio allows teams to query their entire contract repository and generate structured outputs: issue trackers, stakeholder summaries, diligence reports, and risk analyses across thousands of documents.

On top of these capabilities, Chamelio automates workflows: routing approvals, flagging renewals, generating agreements, and integrating with the systems legal teams already use - Slack, Teams, Salesforce, and document repositories.

The result is a shift from static document management to a living legal operating system.

Traction

From pilot to essential infrastructure.

Chamelio’s market pull was one of the strongest signals we’ve seen in legal tech.

Today, Chamelio supports ~100 in-house legal teams, including Wiz, Socure, Fiverr, Global-e, Cellebrite, Lightricks, and more.

One example that stood out:

• A company we spoke with ran a 6-month bake-off against 4-5 competitors and selected Chamelio.

• Since adopting the platform, their legal team has driven a ~15% increase in completed contracts with flat turnaround times and no additional headcount.

• The team estimates $10K+ in monthly savings, with a clear path to enterprise-scale ACVs as integrations deepen.

We heard similar stories from other customers: massive repository analyses completed in minutes, due diligence automated across thousands of contracts, and insights that previously required weeks of manual work.

The Team

Deep domain expertise meets AI-native execution.

Chamelio was founded by operators who have lived the realities of legal and enterprise workflows. Alex, Gil, and Gal are not just pedigreed, but hungry. The team ships quickly, internalizes customer feedback, and iterates relentlessly.

One moment that stuck with us: when the team flew in for their final IC pitch, they passed on dinner with us - not because they were tired, but because they were hosting a prospect dinner with 12 legal leaders, including two existing customers, that evening.

A week later, they were back in New York, carving out time for dinner with us between days spent working the floor at a legal tech conference (with Danny rescuing their booth after a last-minute monitor failure).

It was a reminder that great enterprise companies aren’t built in pitch decks. They’re built in rooms with customers.

Why Work-Bench Is Excited

Chamelio is a system of action that we believe defines the next generation of enterprise software:

• Mission-critical workflows: Legal sits on the critical path of revenue, risk, and compliance.

• Broken legacy systems: CLMs were built for document storage, not AI-native workflows.

• A platform wedge: By becoming the context layer for legal knowledge, Chamelio has a path to expand across procurement, diligence, analytics, and cross-functional workflows.

• Real customer ROI: When legal teams can increase throughput without adding headcount, the value proposition becomes undeniable.

We believe Chamelio has the potential to become the definitive context layer for in-house legal operations replacing fragmented tooling with a unified, AI-native platform built for how modern legal teams actually work.

What’s Next

If you’re a GC or legal ops leader still stitching together spreadsheets, legacy CLMs, and generic LLMs to keep up with contract volume, it’s time to see what an AI-native legal platform feels like. Reach out to the Chamelio team to see how they can help!

Welcome to the Work-Bench family, Chamelio 🦎

Congratulations to the entire Chamelio team! You can check out their press coverage in Law360.

Huge congratulations to Robert Ross (aka “Bobby Tables”), Dylan Nielsen, and the entire FireHydrant team on joining Freshworks!

When we invest at seed, we’re not just underwriting an outcome. We’re underwriting people, conviction, and the willingness to endure the full startup sinusoidal journey. FireHydrant was exactly that, from day one.

I still remember first meeting Robert through Namely’s VP Eng in late 2018. One of those introductions where, within minutes, you can tell you’re talking to someone who deeply understands the problem, not just intellectually, but viscerally. The scars from years as an SRE were already there. The obsession was obvious. We quickly pounced to lead FireHydrant’s Seed round in Dec 2018.

As first time technical founders, they made many common mistakes you’d expect. But we were there each step of the way to coach, guide, and support.

We still laugh about the first board meeting 7 years ago. The deck was rough around the edges, to put it kindly, but the thinking underneath it was razor sharp. The product vision, the customer pain, the clarity around why incident response was fundamentally broken. The ingredients were all there. The potential was undeniable.

A tradition started - unintentionally at first, but very deliberately over time. For every single FireHydrant board meeting over the past seven years, I showed up wearing a FireHydrant sweatshirt or t-shirt. No matter the agenda. No matter the market cycle. No matter how hard the conversation was going to be. It was a small thing, but it symbolized something bigger: belief, alignment, and being all-in alongside the team.

Not long after, Robert gave a dev lunch talk for the Work-Bench community. That was a turning point. You could feel the room light up with founders, operators, and engineers all leaning in. Soon after he co-headlined a NY Enterprise Tech Meetup alongside Alexis Lê-Quôc of Datadog. He didn’t just deliver a product pitch, it was a manifesto. People walked out buzzing, convinced this was a company that had to exist.

From there, it was a blur of milestones and moments that brought me closer to the entire team: speaking at SKOs, sharing the stage at Firehose events, karaoke nights, meals squeezed between meetings, holiday parties, conference hangs, and so much more.

One of my favorite memories over the years was seeing Death Cab for Cutie and The Postal Service with Robert. We both share a fondness for emo music and jumped at the opportunity to see this special 20th anniversary show together at MSG after Robert spoke at our Annual LP meeting. It’s a perfect snapshot of how this relationship always blurred the line between work and life in the best way possible.

Of course, it wasn’t all highs. Scaling a company during ZIRP and the fallout afterwards, there came the whiplash that so many teams endured: growth-at-all-costs…and then the market literally turning upside down a week later. Hard decisions followed. A RIF that no founder ever wants to make. Deeply human moments, handled with care and integrity after lots of planning and deliberation. During those moments you can tell a founder’s true colors, and Robert consistently communicated with transparency and empathy with the board and his employees.

There was the true bet-the-company moment: going all in on Alerting. A bold call. A scary one given the company’s limited engineering capacity. I’ll never forget the late night Zooms and the lunch conversation right before Robert pulled the trigger. He knew it was the correct call and was proven right when competitors started copying the feature and customers started flocking to FireHydrant in droves. That’s how you know you’ve struck a nerve. And then, in the midst of more turbulence, the conviction to buy a competitor, Blameless. Calm, deliberate leadership when it mattered most.

Along the way, we’ve had the privilege of spending real time with so many incredible people:

• Trow, the company’s VP Sales who deftly handled the role of player-coach, closing strategic deals himself while scaling and leveling up his org. I’ll fondly remember our end of quarter recaps over drinks.

• Danielle, building with clarity and strength as CTO while enhancing the company’s enterprise offerings.

• Colin, a FireHydrant OG who was a sales machine during his tenure and always gave me the real talk on deals during our coffee chats in Chicago.

• And Scott, who joined in the earliest days when everything was still fragile and helped secure our first large customers.

These are the moments seed investors actually sign up for.

When we say “we have your back,” we mean it. We’re not here just for the up-and-to-the-right slides. We sign up for the uncertainty, the doubt, the hard calls, and the quiet wins. FireHydrant was the full journey - and then some.

Robert, Dylan, and team: we couldn’t be prouder. You built something real, something durable, something that mattered deeply to customers - and you did it with heart.

We’re so pumped for this next chapter at Freshworks and have no doubt you’re going to make a massive impact.

Cheers to what you built and to everything that’s still ahead. 🍻

We’re thrilled to announce our latest investment as we co-led AIR Platforms’ $6.1M Seed round with Lerer Hippeau. This comes after we previously led their Pre-Seed raise.

We’re excited to partner with AIR Platforms, backing their mission to build the first truly AI-native credit intelligence system - real-time, transparent, and built for the complexity of modern credit markets.

Financial Institutions are making credit decisions in an environment where the volume of available data has grown more than 10x in the last decade, yet most systems haven’t materially evolved. AIR is closing that gap. Their credit intelligence layer increases model transparency, accelerates analysis, and can reduce manual review time by 30–50%.

The market needs infrastructure that enables participants to make faster, more accurate, and more explainable decisions. We expect AIR to become core to how modern institutions assess and price risk.

Why Credit Risk Intelligence Needs a Reckoning - and Why Now

• Credit markets have exploded in size, structure, and complexity. Private credit alone is now a multitrillion-dollar asset class. As public and private credit blur, institutions face a flood of new borrowers, structures, and risk vectors - traditional models weren’t built for this.

• Information overload with no standardized framework for analysis. Firms now juggle financial statements, market data, news, covenants, sector-trends and more - often manually, often inconsistently. That forces slow underwriting cycles, human error, and lack of comparability across portfolios.

• Legacy credit-rating methods are slow, opaque, and static. Ratings from traditional agencies come with lag, limited coverage (especially in private credit), and black-box methodologies. In fast-moving markets, that lack of transparency and agility can be disastrous.

• AI + modern data infrastructure changed the math. AIR uses machine-learning and generative-AI to ingest huge volumes of structured and unstructured data - financials, market signals, news, events - and deliver forward-looking, explainable, real-time credit risk scores. That allows firms to act on signal, not lag.

In short: the convergence of a huge, under-served market + data explosion + outdated workflows + AI breakthroughs makes this the moment for AIR.

The Product - Next-Gen Credit Intelligence

AIR isn’t “just another rating agency.” It offers a full-stack, AI-powered credit-intelligence platform that:

• Produces real-time, independent credit ratings and credit risk scores for private and public entities. Scores reflect underlying financials, sector context, event-driven signals, and are updated as new data arrives.

• Delivers transparency: drivers, not black boxes. For each entity, AIR surfaces the core risk factors (leverage, profitability, cash flow, sector-specific inputs), with their relative weights - so analysis is intuitive and auditable.

• Generates automated credit memos, diligence packages, and scenario analyses. Users can quickly produce standardized, comparable credit packages - dramatically reducing manual labor, turnaround time, and error risk.

• Provides early-warning signals and continuous monitoring. As new financials, events, or market shifts occur, AIR updates scores - giving firms foresight, not hindsight.

• Scales across any credit universe. From publicly-traded corporates to private firms and structured credit vehicles - AIR supports broad coverage, making it “one standard across credit, leadership, and stakeholders.”

In practice: AIR turns what used to take weeks of manual modeling and subjective grading into seconds - with more rigor, consistency, and comparability across portfolios.

The Team - The Right Combination of Credit + AI + Operational Rigor

AIR was founded by a group of seasoned credit-risk and AI veterans who have each lived the pain of legacy workflows. Together, they combine domain expertise, AI-first DNA, and an uncompromising mission to bring clarity and speed to credit risk, something legacy systems have failed to deliver for decades. They most recently worked together at DataRobot before joining forces to launch AIR.

CEO Glenn Carvajal brings former experience at Moody’s, where he co-led Moody’s Analytics Research, Data & Analytics Credit Products business across the Americas, CTO Arbi Abeshi brings a deep ML/engineering background building financial infrastructure at firms like Goldman Sachs and Morgan Stanley, and Chief AI Officer Joe Burdis is a credit-modeling veteran with a track record in AI/ML + structured finance from SAS and Deloitte. The trio worked together at DataRobot and has since reunited to build the next generation of AI-powered credit infrastructure redefining how institutions detect and manage risk.

There’s nothing better than when a deal comes together and our NYC enterprise tentacles work their magic. Glenn came to us via our friend Lindsay (who previously worked at a WB portfolio company) from their DataRobot days. We then called our Fund II founder Diego Oppenheimer - who’d worked with Joe and Arbi post-Algorithmia acquisition - for the inside scoop. And in true “small world” fashion, our portfolio company FireHydrant’s VP of Sales had teamed with Glenn back when Brian ran US East at DataRobot.

This is the NYC enterprise community at its best: fast signals from people we trust, all pointing to the same conclusion - this is a team we want to be in business with.

Why Work-Bench Is All-In

AIR sits squarely at the intersection of what we believe matters most right now to rethink enterprise workflows:

• Mission-critical workflow. Credit risk decisions sit at the heart of institutional risk, fund allocations, and regulatory compliance - changes here ripple across entire balance sheets.

• Manual, redundant, costly processes. Until now, underwriting and monitoring have relied on human labor, static reports, and slow review cycles. That’s inefficient, opaque, and dangerous - especially in credit markets.

• AI + data = step-function improvement. AIR shows that when you apply modern ML + data infrastructure, you don’t just get slight gains - you transform the workflow.

• Real demand and early validation. With early customers and design partners across CLOs, credit funds, institutional investors, banks, and Fortune 500 companies, AIR is already proving its product-market fit.

We believe AIR has the potential to become the standard credit-intelligence layer for institutions, replacing legacy ratings and manual workflows with an AI-first, audit-ready, scalable infrastructure.

If you’re an investor, credit manager, or risk team relying on spreadsheet-based diligence or opaque rating agencies - the future is here.

Welcome to the Work-Bench family, AIR. Looking forward to building the future of credit intelligence together. ⚡️

Congratulations to the entire AIR team! You can check out their press coverage in Banking Dive.

We are thrilled to announce our latest investment as we lead Built AI’s $6M Seed round, with participation from Lerer Hippeau and Timber Grove Ventures.

Commercial real estate moves hundreds of millions of dollars per day. And yet, one of the world’s largest and most consequential asset classes still runs on brittle spreadsheets, outdated modeling tools, and PDFs buried in inboxes.

That’s where Built AI enters. Their thesis is clear: real estate, one of the largest and most critical asset classes in the world, deserves modern tooling. Not just spreadsheets dressed up in a UI, but a purpose-built, AI-powered engine that lets investment teams underwrite smarter, faster, and more consistently.

In a world where “missed deals” is often just a function of bandwidth or speed, Built AI wants to make opportunity cost explicit, and reclaim time for what really matters: strategic deal analysis and key drivers that occur outside of spreadsheets.

Why Now?

Legacy players - Argus, Yardi, and a constellation of point solutions - weren’t built for the complexity, speed, or data fragmentation of today’s market.

Prior attempts to displace Argus failed because they ran into the same wall: data extraction wasn’t accurate enough. GenAI changed that. High-quality extraction from PDFs, spreadsheets, rent rolls, and deal docs is now finally fast, reliable, automatable, and commercially viable. This unlocks a completely new workflow: instant financial models generated directly from raw deal materials.

Combine that with leaner teams, rising transaction complexity, and industry-wide pressure for auditability, and the timing is perfect for an AI-native operating system for real-estate investing - exactly what Built AI is building.

The Product

Built AI is an AI-powered financial modeling and decision-support platform that analyzes investment opportunities in minutes instead of weeks - with transparent assumptions, line-by-line auditability, and a user-friendly interface that’s quick to learn.

Their wedge is simple but powerful: Turn fragmented, unstructured documents and deal materials into instant, accurate underwriting - and then unify everything a firm knows about a property, portfolio, or market into one end-to-end platform.

• Automated screening & underwriting at speed. Built AI can generate a full cash-flow model - including complex leasing scenarios and hold-period assumptions - in minutes, based on the user’s inputs and documents uploaded. That means fewer late nights in Argus and Excel, and fewer oversights from manual entry.

• Instant portfolio aggregation and analysis: C-suite need to make decisions at the portfolio level and Built AI automatically rolls up existing portfolios and any potential acquisitions. That means instant deep-dive analysis and faster decisions.

• Collaboration & shareability. Built for teamwork: analyses can be shared with partners, sponsors, or brokers in a few clicks - removing email chains, lost attachments, and miscommunication.

One major broker recently cut a two-week underwriting exercise that previously required multiple analysts down to a few minutes using Built AI.

In short: Built AI doesn’t just optimize underwriting - it raises the floor of intelligence across a firm’s entire investment stack.

The Team

Built AI is built by real estate people for real estate people.

It was co-founded by investment banking and real estate private equity veteran Natan Lempert, who previously held roles at Goldman Sachs, Citi and Benson Elliott, and Firoz Noordeen, who previously launched and scaled fintech ventures as Director at Natwest Ventures.

In our view, the future of software requires three essential attributes: domain expertise, AI skills, and ingenuity. These founders have all three and seek to fundamentally reimagine the real estate investment process.

By automating away manual tasks, unlocking organizations’ historical data, and overlaying market data, real estate investment professionals can rapidly screen deals, run various scenarios, and spend time laser focused on key diligence areas outside of their computer screen to transact with more speed and confidence than ever before.

Traction

The quality of the team and product paired with the market demand has led to strong early traction and proof points.

Built AI’s customers include US publicly traded developer Howard Hughes and global real estate firm CBRE, FTSE 100-listed REIT LondonMetric, and Knight Frank Investment Management, among many others.

Howard Hughes’ CFO remarked that “Built AI is helping us reimagine asset management by automating the analysis of rent rolls, cashflows, and business plans - which will allow us to do in minutes what it currently takes days. The ability to run multiple scenarios at both the property and portfolio level in real time is invaluable. We’re excited to be a design partner as they continue to expand and enhance their platform.”

We’re proud to back Built AI as they turn real-estate investing into something modern, agile, and scalable. If you’re a real estate investment professional looking to improve your deal analyses, book a demo with Built AI today ⚡️

PS - This investment has to be my wildest sourcing story ever. It involved a mohel, a Miami connection, and a consulting skill that got me through the 2008 financial crisis. Watch this ~2 minute clip to hear the whole story.

Congratulations to the entire Built AI team! You can check out their press coverage in the Commercial Observer.

🤖 Reinforcement Learning’s Last Mile Problem

We are thrilled to announce our latest investment as we lead Runhouse’s $5M Seed round, with participation from Hetz Ventures and Fathom Capital.

As AI labs increasingly rely on synthetic data to improve their models for high-value enterprise tasks, leading Fortune 500s are recognizing the immense value of their proprietary data for training powerful, domain-adapted AI agents (“post-training”). However, while reinforcement learning (RL) methods for agentic training have matured, the biggest barrier to mobilizing data is lagging ML infrastructure. Many organizations still rely on outdated tools like AWS SageMaker and Kubeflow, which are insufficient for the speed, customization, and fault tolerance required by modern Machine Learning and Reinforcement Learning workloads.

Runhouse is bringing AI post-training out of the lab and into production for enterprises and AI-native teams. Their flagship product, Kubetorch, allows ML teams to build and scale magically better models with proprietary data to power the next generation of AI use cases.

💥 Why Now

Enterprises are reaching a turning point in how they build and adopt AI. Frontier labs like OpenAI still dominate large-scale pretraining, creating foundation models with hundreds of thousands of GPUs. However, real differentiation is shifting to post-training, where models are adapted with proprietary data and simulation environments to excel at high-value agentic tasks.

DeepSeek released its groundbreaking R1 model in January 2025, with novel reinforcement learning (RL) training methods. Subsequently, the accuracy of AI code agents more than doubled. Now, post-training is being applied to power the next wave of innovation across domains like law, finance, healthcare, life sciences, and mechanical engineering, where even the largest models still struggle with production-level accuracy.

However, unlike coding models or general-purpose language tasks, the data and workflows in specialized enterprise domains are entirely proprietary. Frontier AI labs try to bridge this gap by buying synthetic datasets from providers like Scale AI or Mercor, but this approach is costly and limited. Meanwhile, enterprises already sit on the data needed to bring agentic AI to production quality. Investment banks, for example, have decades of ground-truth financial models, portfolios, and back-office tasks to train “Cursor for Finance.” As RL methods become simpler and more reliable, the best path forward is training first-party models, not praying for an AI lab miracle.

🌦️ The Problem

Even as reinforcement learning shows enormous promise, a critical barrier remains: there is no developer platform robust enough to support the post-training methods ML and data science teams want to use. For instance, RL is a heterogeneous workload, combining distributed training, inference, and evaluation, something most existing ML platforms, whether cloud-based like SageMaker or Vertex AI, or open-source like Kubeflow, are not designed to handle.

Teams also need a rapid research feedback loop. Today’s ML infrastructure slows this process to a crawl. Every small code change can require a 30-minute wait to test, thanks to slow CI pipelines and long worker restarts. Tools like ML Notebooks accelerate iteration, but cannot handle scale or workload heterogeneity, and create a gap between research and production. No single system currently enables both interactive experimentation and scalable, reproducible execution.

As a result, even as RL research methods become battle-tested and new libraries simplify implementation, infrastructure remains a stubborn bottleneck. A proof of concept might emerge in a day, but it can take an entire quarter for that same training to reach production and prove itself at scale.

🏃‍♀️🏠 The Product

Runhouse solves this bottleneck with Kubetorch, a powerful distributed framework for Kubernetes, purpose-built for modern enterprise ML and reinforcement learning. Kubetorch makes building and running complex distributed workloads on cloud hardware effortless. Regular Python code runs with instant iteration, perfect reproducibility, and unlimited scale.

Built for enterprise security, Kubetorch can be deployed in just 15 minutes entirely within VPC, whether on an on-prem cluster, cloud GPUs, or even multiple different Kubernetes clusters. For ML researchers, complex Kubernetes interactions are fully abstracted behind intuitive Pythonic APIs. For platform teams, Kubetorch’s Kubernetes-native architecture and familiar primitives make it plug-and-play with existing observability and management stacks.

Kubetorch is already powering production deployments at Fortune 500 companies, frontier AI labs, and AI-native startups. These teams are leveraging both proprietary data and valuable user interaction data from live agentic AI deployments, creating a continuous learning loop that drives ever-more performant first-party models.

🏆 The Team

Runhouse was founded by Donny Greenberg and Josh Lewittes, two engineers who have built their careers tackling the practical challenges of machine learning systems.

Donny led product for PyTorch at Meta, where he supported hundreds of engineers building on the framework. Before that, he worked at IBM as a tech lead in quantum computing research and at Google as a product manager in large-scale ad infrastructure. Across these roles, he saw a common problem: breakthrough models and platforms were exciting, but engineers struggled when iteration slowed down at the infrastructure layer.

Josh’s path has been rooted in applied machine learning. He built fraud detection models at SecuredTouch, developed backend ML systems at KPMG, and most recently worked as a machine learning engineer at Gloat. In each of these roles he ran into the same bottleneck from a practitioner’s perspective: ideas and data were plentiful, but getting models into production was slow and cumbersome.

Together, Donny and Josh bring the platform and infrastructure expertise alongside hands-on ML engineering experience needed to solve reinforcement learning’s last mile. Runhouse reflects their conviction that enterprise teams should be able to iterate on RL models with the same ease as running code locally.

💃🏻 Congratulations to the Runhouse Team

We are proud to back Donny, Josh, and the Runhouse team as they tackle one of the most important infrastructure challenges for reinforcement learning. By removing the friction between code and compute, they are making it possible for enterprises to move from experimentation to production at the pace that modern software demands.

Runhouse is currently in open beta, and you can sign up for access at run.house. Read more about their launch on the Runhouse blog.

We’re excited to announce that Work-Bench led Paratus’ $2.5M pre-seed round alongside Bling Capital, Teamworthy Ventures, and a group of defense and startup angels.

From our very first conversation with the Paratus team, we were struck by the scale of their vision and how they wanted to fundamentally rethink military workflows and help usher in a new era of generating defense readiness. The team’s ability to navigate the complexities of defense procurement while approaching a historically overlooked category with fresh, innovative thinking was nothing short of inspiring.

Military operations are high-stakes and highly coordinated, but the systems behind them? Not so much.

Today, over 3 million U.S. service members are tasked with meeting compliance requirements across areas like weapons training and medical screening. Yet, much of this still depends on outdated, disconnected systems. The result? Inefficient processes, excessive admin work, and critical gaps in visibility. All of which compound and ultimately limit a soldier's chances of being deployed on missions.

Paratus is rethinking “readiness” for military personnel that can’t afford to be unprepared. Its AI-native platform streamlines and automates critical readiness workflows from firearms qualification and gas mask checks to immunizations, mental health screenings, and beyond. Paratus’ leadership dashboard provides commanders with granular insights about which troops are ready to be deployed and cuts admin time by 90%. It’s a bold vision for how mission-critical teams should operate, with the same kind of intuitive, modern tools many of us take for granted.

The team brings early GTM hustle and a deep empathy for the friction these units face, and they’re already earning trust from some of the military’s highest tempo units. Enterprise sales is hard. Selling into defense is even harder - and why we’ve been both impressed and energized to partner with Luke Sommer, Roberto Medrano, and Stuart Baker as they redefine work management for frontline service members.

Congratulations to the entire team!

Paratus is growing their team in SF and actively hiring founding engineers. If you know someone who’s passionate about this shared mission, send them over to Paratus.

🚀 Building and selling in 2025 looks different.

🗽 If you're a technical founder or operator in NYC, don't miss our July NY Enterprise Tech Meetup for a candid conversation with the founders of Graphite and Superblocks. We will discuss how to build, adapt, and find your footing in a market that’s evolving as quickly as AI is reshaping it.

🤖 We’ll dig into what it takes to build a company when user expectations are shifting, iteration cycles are compressing, and the lines between product and go-to-market are increasingly blurred. Whether your company started before or after ChatGPT, you’re now operating in a different world.

✍🏽 We’ll talk about how to stay close to real customer problems, and how to evolve your product and positioning when the category itself is still being formed. Founders need to stay close to both product and customer. In a market moving this quickly, that’s the only way to build the right thing.

🧨 Merrill and Brad will share hard-earned lessons from building through uncertainty, iterating quickly, and staying focused on the fundamentals as the market shifts around them.

🔥 Fireside chat with:

• Merrill Lutsky, Founder & CEO of Graphite

• Brad Menezes, Founder & CEO of Superblocks

🧠 Graphite recently raised a $52M Series B and launched Diamond, its AI-powered code review agent. Diamond delivers instant, codebase-aware feedback on every pull request, helping developers catch bugs, logic errors, and style issues in seconds. The platform also supports customizable rules, stacked PRs, and self-healing CI. Engineering teams at companies like Shopify, Snowflake, and Figma use Graphite to ship high-quality code faster.

🛠️ Superblocks recently introduced Clark, an AI agent that helps developers build production-grade internal apps from natural language prompts. Clark combines AI-generated code, visual editing, and full control with React, while enforcing enterprise standards like audit logging, SSO, RBAC, and compliance. With around $60M in funding, Superblocks is already supporting teams at companies like Instacart to accelerate internal tooling development.

✨ If you're a technical founder or early operator navigating this new terrain, this conversation will leave you with real, hard-earned insight into what it means to build right now.

The NY Enterprise Tech Meetup (NYETM) is the only in-person forum exclusively dedicated to enterprise software technology. Now over 10 years old, with 10,000+ members and 100+ past events, the meetup brings together attendees across NYC's “suits and hoodies” community of enterprise startup founders and operators, Fortune 500 executives, VCs and more.

​​This NYETM is brought to you by:

​​​Work-Bench: We are an early stage enterprise software VC fund based in New York City. We lead $3M - $6M Seed rounds in SaaS and infrastructure startups, focus on supporting companies’ 0 to 1 go-to-market journeys, and have built a dynamic enterprise tech community here in NYC. We’re investing out of our $160M Fund IV, and notable portfolio companies include Cockroach Labs, Socure, Spring Health, and Dialpad.

​​Wells Fargo’s Technology Banking Group has delivered specialized insights and recommendations to technology companies for more than 25 years, serving as a service provider, buyer, and investor. The team provides scalable support to clients in all stages of the business lifecycle – early, growth, and maturity – and combines industry experience with the strength and resources of one of the largest U.S. commercial banks. Primary subsector focus includes Software, Fintech, E-commerce, Semiconductor, Business & Technology Services, and Sustainable Tech.

We're thrilled to announce that CloudQuery has raised a $16M funding round led by Partech.

CloudQuery is tackling the growing complexity of enterprise cloud visibility by centralizing cloud assets, security metadata, and cost data into a single, queryable platform. As cloud estates become increasingly sprawling and fragmented, existing tools remain siloed. There's one for security, another for cost, yet another for asset inventory. This makes it nearly impossible to get a unified view across domains.

CloudQuery takes a developer-first approach to cloud governance, pulling data from 60+ sources including AWS, GCP, Azure, Okta, and Wiz into a centralized data warehouse that's accessible through intuitive SQL queries. Built on high-performance technologies like ClickHouse and Apache Arrow, their platform continuously syncs configuration, security, and cost metadata to ensure teams work with the most accurate, up-to-date information.

Over the past year, they saw 4x ARR growth, 4x customer expansion (including 4 Fortune 500 institutions), doubled the team size, and acquired Tailwarden to bolster their platform. The company is serving some of the biggest companies including enterprises like Fastly, Zendesk, Reddit, Palo Alto Networks, Ridgeline, and more.

As founder Yevgeny Pats notes: "When you put developers first, with accessible data, flexible APIs and native language like SQL, you empower them to move faster, catch issues earlier and build more securely."

We're proud to have backed CloudQuery from day 1 and excited to continue to support them as they scale their mission to bring transparency and control to the complex world of multi-cloud infrastructure! 🚀

Read More

Check out Kelley and Yevgeny enjoying borscht 😂

Join us for a special edition of the NY Enterprise Tech Meetup focused on AI-powered hypergrowth in healthcare with the founders of Adonis and Tennr, two of the most exciting healthcare startups here in NYC.

🎢 We’ll go deep into their founding stories: how they broke into one of the most complex industries, secured early customers, and scaled through messy realities and major inflection points.

This is a rare chance to learn firsthand from founders who’ve navigated it all: product-market fit, go-to-market motion, fundraising, and building with AI. If you’re building in healthcare or thinking about it here in NYC, you won’t want to miss this!

​​​🔥 Our fireside chat with:
- Akash Magoon, Co-Founder & CEO at Adonis
- Trey Holterman, Co-Founder & CEO at Tennr

​Adonis is a revenue intelligence and automation platform helping healthcare providers recover lost revenue. With over $50M raised, the company processes over 100 million claims annually, partnering with 20,000+ providers across various specialties.

​Tennr automates healthcare documentation, streamlining patient referrals and reducing administrative burdens. The company has raised over $60M in funding, including a recent Series B led by Lightspeed Ventures, and is rapidly expanding its impact across the U.S. healthcare system.

The NY Enterprise Tech Meetup (NYETM) is the only in-person forum exclusively dedicated to enterprise software technology. Now over 10 years old, with 10,000+ members and 100+ past events, the meetup brings together attendees across NYC's “suits and hoodies” community of enterprise startup founders and operators, Fortune 500 executives, VCs and more.

This NYETM is brought to you by:

​Work-Bench: We are an early stage enterprise software VC fund based in New York City. We lead $3M - $6M Seed rounds in SaaS and infrastructure startups, focus on supporting companies’ 0 to 1 go-to-market journeys, and have built a dynamic enterprise tech community here in NYC. We’re investing out of our $160M Fund IV, and notable portfolio companies include Cockroach Labs, Socure, Spring Health, and Dialpad.

Opal Ventures: We are thesis-driven investors backing health tech startups working to transform healthcare in the US at pre-seed.

Brex: We transform finance teams from reactive no-sayers to proactive growth drivers. With Brex, companies spend with confidence by empowering employees to make smarter financial decisions from anywhere. Brex provides corporate cards, banking, and global payments, plus intuitive software for travel and expenses, that make it easy to plan and track all company spend in one place, in real time. Tens of thousands of companies from startups to global enterprises — including DoorDash, Flexport, and Compass — use Brex to proactively control spend, reduce costs, and increase efficiency on a global scale.