For every headline claiming AI is transforming the enterprise, there's another claiming 95 percent of pilots fail. The signal-to-noise ratio has gotten bad enough that most business owners we talk to have stopped trusting either extreme. They want numbers. Real ones.
A new analysis from Andreessen Horowitz, built on private startup revenue data and conversations with corporate buyers, finally provides some. It's the clearest picture to date of where enterprise AI is actually generating ROI — and, just as importantly, where it isn't yet.
Here is our read on what the data means for business owners trying to figure out where to place their bets in 2026.
The Headline Number: Enterprise AI Is Already Mainstream
The finding that most surprised us wasn't about a specific use case. It was about the speed of adoption.
- 29% of the Fortune 500 are now live, paying customers of a leading enterprise AI startup.
- Roughly 19% of the Global 2000 are in the same position.
- To qualify, these companies had to sign a top-down contract, convert a pilot, and go live in production — not just run a demo.
To put that in context: Fortune 500 companies are famously the slowest technology buyers on the planet. Historically, startups spent their first few years selling to other startups before they could even get a single enterprise on the phone. Landing a Fortune 500 logo was usually a five-to-seven year journey.
ChatGPT launched in November 2022. A little over three years later, nearly a third of America's largest companies have real AI deployments running in production.
The lesson for business owners: The window where "we're still evaluating AI" was a defensible position has closed. Your largest competitors are already in production. The question is no longer whether to move, but where to move first.
Where the Money Is Actually Flowing
When you look at which AI startups have generated the most enterprise revenue, three use cases dominate everything else — and one of them dwarfs the rest by nearly an order of magnitude.
1. Coding — The Breakout Category
Coding is not just the biggest category. It is so far ahead of everything else that, on the revenue chart, it looks like a mistake.
Tools like Cursor, GitHub Copilot, Claude Code, and Codex have posted growth rates that blew past even the most bullish forecasts. The majority of Fortune 500 AI tooling spend today is on code.
Why coding works so well:
- Code is data-dense. Decades of high-quality open source code gave models an unusually rich training set.
- Code is verifiable. You can run it and immediately know if it works. That tight feedback loop lets models keep getting better.
- The ROI is obvious. Portfolio companies consistently report that their best engineers have become 10 to 20 times more productive with AI coding tools. That math makes itself.
- Engineers are natural early adopters. They find the best tool and use it without waiting for a procurement committee.
- It doesn't need to be perfect. Even catching bugs and generating boilerplate is enormously valuable. There's no "100 percent or bust" test.
Why this matters to non-developers: Code is upstream of every other piece of software in your business. When AI makes writing software faster, every downstream tool you use gets cheaper and more capable. The "software cost" line in your budget should be falling, not rising.
2. Support — The Quiet Workhorse
Customer support is the opposite end of the spectrum from coding. It's the function most companies actively avoid thinking about. Most of it is outsourced to BPOs because it's considered tedious, back-office work.
That's exactly why AI has dominated it.
What makes support a perfect AI use case:
- Tasks are cleanly defined. Support teams already operate on SOPs (standard operating procedures) because they have to train new reps fast. Those SOPs translate directly into instructions for AI agents.
- Metrics are objective. Tickets resolved, CSAT scores, resolution rate — you can A/B test AI against humans and get a clear answer.
- There's a natural off-ramp. When AI can't handle something, it escalates to a human. The worst-case scenario is "no worse than today."
- The change management is easy. Most support is already outsourced. Swapping a BPO for an AI agent doesn't require restructuring your company.
- Customers don't care who's on the other end. Support is transactional. Nobody is building a relationship with the person issuing their refund.
Companies like Decagon and Sierra are growing so fast because the math is undeniable: more tickets resolved, higher satisfaction, lower cost. It's one of the rare AI decisions where the CFO and the head of customer experience actually agree.
3. Search — The Invisible Giant
Search is the third horizontal category where enterprise spend is concentrated. It's also the one most people underestimate, because a lot of its value gets baked into consumer products like ChatGPT itself.
- Horizontal search — finding information across all the systems employees use — is dominated by Glean.
- Vertical search — deep search inside a specific industry's data — is where Harvey (legal) and OpenEvidence (medical) have built massive businesses.
The pattern here is revealing: every large enterprise has the same complaint, that employees can't find what they need across Slack, email, Google Drive, Salesforce, SharePoint, and twelve other systems. AI-native search is finally solving a problem that traditional enterprise search never could.
The Industries Leading the Charge
Three industries are pulling ahead of everyone else on AI adoption. One of them is expected. Two of them are genuinely surprising.
Technology (No Surprise)
ChatGPT itself reports that 27 percent of its business users come from the tech industry. Most of the early customers of Cursor, Decagon, and Glean were tech companies. This was entirely predictable — tech always adopts tech first.
Legal (The Surprising First-Mover)
For decades, "legal software" was a punchline. Lawyers famously hated new tools. Workflow software never fit how they actually worked. Adoption timelines were measured in years.
AI broke that pattern almost overnight. Why?
- AI is excellent at parsing dense text, which is 80 percent of what lawyers do.
- AI can reason over massive document sets — contracts, filings, discovery — faster than any associate.
- AI can draft responses and summaries that a senior lawyer only needs to review and refine.
- Unlike old workflow tools, AI matches the actual shape of legal work, which is unstructured and judgment-heavy.
The results are not subtle. Harvey reported around $200 million in annualized recurring revenue within three years of founding. Eve, which specializes in plaintiff law, has more than 450 customers and a $1 billion valuation. These are not normal numbers for legal tech.
Healthcare (Breaking the Legacy Gate)
Healthcare was similarly hostile to traditional software, for two reasons: the work is complex and high-stakes, and the market is locked down by system-of-record EHRs like Epic that squeeze out new vendors.
AI found a way around the gate. Instead of trying to replace the EHR, the best AI companies in healthcare focused on discrete, high-pain use cases that sit alongside it:
- Medical scribing (Abridge, Ambience) — AI listens to patient visits and writes the clinical note.
- Medical search (OpenEvidence) — doctors query the medical literature in plain English.
- Back-office automation (Tennr) — AI handles the byzantine administrative work around referrals, authorizations, and billing.
None of these required rip-and-replace. They just quietly took over the work nobody wanted to do, and scaled from there.
Why These Categories Are Winning
When you look at the winning use cases side by side — code, support, search, legal work, medical documentation — they share a very specific profile. This is the pattern to memorize:
- Text-based work. Models are still best at reading, writing, and reasoning over language.
- Repetitive and rule-driven. Work that already has playbooks, SOPs, or templates.
- Tight feedback loops. You can tell immediately whether the output is right.
- A human in the loop. A person reviews before anything ships, which makes adoption low-risk.
- Light regulation. Or at least, regulation that doesn't block pilots.
- Clear, objective metrics. Tickets resolved, contracts reviewed, bugs caught, notes generated.
The categories where AI adoption is slower usually fail one or more of these tests. Physical-world work, relationship-driven sales, heavily regulated finance, and multi-stakeholder coordination are all harder. That doesn't mean AI won't get there. It means the order of arrival is predictable — and if your business operates in a "tough" category, you still have a window.
The Capabilities Are Climbing Fast
The most underappreciated finding in the data is how quickly model capabilities are improving in areas where adoption hasn't caught up yet.
OpenAI's GDPval benchmark — which measures AI performance against human experts on real, economically valuable tasks — has jumped dramatically in just the last few months:
- Accounting and auditing improved by nearly 20 percentage points.
- Police and detective work (yes, really) improved by nearly 30 points.
- Spreadsheets, financial workflows, and "computer use" are all top research priorities at the model labs right now.
What this tells us: the categories that haven't broken out yet in enterprise revenue are not held back by technology. They're held back by the fact that no startup has yet built the right wrapper, the right workflow, or the right trust layer around the model. That gap is closing fast.
The practical translation: If you run an accounting firm, a financial services business, or an operations-heavy company, the tools that will transform your work are being built right now. The question isn't whether they'll arrive. It's whether you'll be ready to use them when they do.
What This Means for Your Business
This research was written for venture capitalists thinking about where to invest. We think the more interesting audience is business owners trying to figure out where to spend — because the map the VCs are using to pick startups is the same map you should be using to pick tools.
Here's how we'd translate the data into action:
- If you haven't deployed AI in engineering, do that first. It's the single highest-ROI move available, and the tools are mature. Every week you wait is a productivity gap against competitors that widens on its own.
- If you haven't deployed AI in customer support, do that second. The ROI math is unambiguous. Pick a vendor, run a pilot on a narrow slice of tickets, and let the numbers make the case.
- If internal search is painful, solve it. Every business we work with has the same complaint — employees can't find what they need. The category is mature enough now that this is a solved problem.
- If you're in legal, healthcare, or professional services, you are in an advantaged position. Your industry is one of the fastest movers. The tools for your work exist. Use them.
- If you're in a "slower" category, start building the foundation now. Get your data organized. Document your processes. When the right tool lands in your industry — and the GDPval numbers suggest that window is months, not years — you want to be positioned to adopt it in weeks, not quarters.
The common thread in every winning deployment is that the business owner moved before they felt certain. The data is now clear enough that waiting is the expensive decision.
Key Takeaways
Enterprise AI adoption is already mainstream. Nearly a third of Fortune 500 companies have real AI in production just over three years after ChatGPT launched. "We're still evaluating" is no longer a defensible position.
Three use cases dominate the spend: coding, support, and search. Coding is larger than every other category combined. These are the areas where ROI is proven, vendors are mature, and the adoption path is well understood.
Tech, legal, and healthcare are the breakout industries. Legal and healthcare are especially notable because they were historically among the slowest technology adopters. AI changed the shape of the value proposition, and they responded.
Winning categories share a clear profile. Text-based, repetitive, verifiable, with a human in the loop and objective metrics. Use this pattern to identify where AI will work in your own business first.
The slower categories are not blocked by technology. Model capabilities in accounting, finance, and operational work are climbing fast. The tools for these industries are being built right now, and the businesses that prepare their data and workflows today will be ready to deploy them in weeks, not years.
The highest-ROI move for most businesses right now is the same one the Fortune 500 is making: AI tools for engineers and AI agents for customer support. Start there, prove the model inside your organization, and use the wins to fund the next workflow redesign.
This article synthesizes and interprets research originally published by Kimberly Tan of Andreessen Horowitz in the article "AI Adoption by the Numbers: Where Enterprise AI is Actually Working." We're grateful for the primary data, and we've built on it with our own perspective from working with mid-market businesses across professional services, healthcare, and e-commerce.