49.7% of AI Agents Build Software. The Other 50% Is Wide Open.
Half the AI agent market is one category. The bloodbath is in software engineering. Healthcare: 1%. Legal: 0.9%. Education: 1.8%. 300 vertical AI unicorns are waiting to be built.
Anthropic published data on what people actually use AI agents for. Measured by tool calls — actual actions agents take, not conversations, not demos — the distribution is lopsided:
| Vertical | Share of AI Agent Tool Calls |
|---|---|
| Software Engineering | 49.7% |
| Education | 1.8% |
| Healthcare | 1.0% |
| Legal | 0.9% |
| Everything else (16 categories) | <5% each |
Half the AI agent market is one category. The bloodbath is in one corner. The rest of the market is empty.
Why Software Engineering Dominates
Engineers build AI tools. Engineers use AI tools. The creators and the users are the same people. The feedback loop is tight. The use cases are obvious. The tooling is mature.
This isn't because software engineering is the most valuable use case for AI. It's because it was the most natural starting point. The people building the hammers used them on their own nails first.
The Autonomy Data
Anthropic's research on agent autonomy reveals something non-obvious about the deployment gap:
Session length nearly doubled in 3 months (October 2025 to January 2026), from under 25 minutes to over 45 minutes. This happened smoothly across model releases, meaning existing models are capable of more autonomy than they exercise in practice.
Trust builds slowly: New users auto-approve 20% of AI actions. Veterans with 750+ sessions auto-approve 40%. The gap isn't capability. It's trust.
Agents are conservative: Claude Code initiates clarification requests more often than humans interrupt it. On complex tasks, agent-initiated pauses happen 2x more often than human interruptions. The agent is more cautious than the user.
300 Vertical AI Unicorns
170+ SaaS unicorns were created over decades. The thesis: vertical AI could produce 300+ unicorns in a fraction of the time.
Why? Because vertical AI is 10x larger than SaaS.
SaaS replaced filing cabinets with databases. The clerk who fetched your file now uses Workday. But you still need a person to make decisions about the data. SaaS stored information. Vertical AI acts on it.
A vertical AI company replaces BOTH the software AND the operator. Not just the Zendesk license. The customer service team that uses Zendesk. Not just the EHR system. The nurse who enters data into the EHR.
That's a fundamentally larger market. Every SaaS company captured the software budget. Vertical AI captures the labor budget. The labor budget is 10-50x the software budget in most organizations.
Where the Gaps Are
Healthcare (1% of agent activity)
Healthcare voice AI market: $468M in 2024, projected $3.18B by 2030 (37.79% CAGR). The fastest-growing vertical in voice AI after financial services.
IVF clinics generate 15-25 patient touchpoints per cycle. At 100 cycles/month, that's 1,500-2,500 calls per month that are routine and automatable. Nurses spend 3 hours/day on these calls.
The AI exists. The deployment doesn't. Not because the technology isn't ready. Because nobody has built the domain-specific orchestration, the trust ladder, the integration with clinic workflows.
Legal (0.9% of agent activity)
85% of legal work is theoretically automatable by AI. Only 10% is being automated.
When Anthropic shipped a legal plugin for Claude in February 2026, Thomson Reuters, RELX, and Wolters Kluwer lost billions in market cap overnight. Legal research crossed below the commodity line in a single product launch.
But legal practice — filing motions, negotiating contracts, managing cases — remains untouched. The tools exist. The vertical applications don't.
Education (1.8% of agent activity)
78% theoretically automatable. 8% observed coverage. The largest gap of any knowledge work category by absolute volume (millions of teachers worldwide).
Finance (not separately listed but massive)
Business & Finance: 98% theoretical, 35% observed. 63% untouched. Every financial analyst, every accountant, every compliance officer does work that AI can technically handle. Almost none of it is automated.
The Enterprise Bridge Problem
Aaron Levie (Box CEO) explains why horizontal agents don't capture vertical value:
"The last mile of making agents work in variable, hostile enterprise environments is the most valuable and hardest part."
The Pattern from Every Previous Wave
| Wave | Horizontal Platform | Vertical Winner |
|---|---|---|
| Cloud | AWS | Veeva ($40B, healthcare CRM on AWS) |
| Payments | Stripe | Vertical fintech built on Stripe |
| Mobile | iOS/Android | Uber, DoorDash, healthcare apps |
| AI | ChatGPT, Claude, Gemini | ? (You're here) |
The horizontal platform always comes first. Then vertical applications capture the real value. The platform captures infrastructure revenue. The vertical captures the industry revenue. The industry revenue is always larger.
AWS revenue: ~$90B/year. The businesses running on AWS: trillions in combined revenue. The platform is not where the value concentrates. The vertical is.
What This Means for Builders
1. Don't Build Another Coding Tool
49.7% of agents are in software engineering. The space is saturated. GitHub Copilot, Cursor, Claude Code, Devin, Replit Agent, v0, Bolt, Windsurf. Every week brings a new entrant. The competition is intense and the platforms (Anthropic, OpenAI) have inherent advantages.
2. Pick a Vertical Nobody Else Is Touching
Healthcare at 1%. Legal at 0.9%. Education at 1.8%. These aren't small markets. Healthcare is a $4 trillion industry in the US alone. 1% AI penetration means 99% is untouched.
3. Domain Expertise Is the Edge
The AI model is commodity. Anyone can call the Claude API. The domain knowledge — understanding IVF patient workflows, understanding legal case management, understanding how a supply chain actually functions — that's what makes a vertical AI product work.
4. The Window Is Finite
The gaps are real but temporary. Healthcare at 1% won't stay at 1% for five years. The early movers in each vertical will establish the integrations, earn the trust, accumulate the usage data. Late entrants will face switching costs they can't overcome.
The One Question
300 vertical AI unicorns are waiting to be built.
49.7% of builders are fighting over the same corner.
The rest of the map is empty.
Which part of the map are you on?
Data from Anthropic's Measuring AI Agent Autonomy research (Feb 2026), a16z market analysis, Aaron Levie's enterprise AI thesis, and Garry Tan's YC observations.