Research brief · AI workforce · October 2026

Who sells AI
agents that stick

We started from a public list of more than a hundred companies selling AI agents to business, added the ones with real traction, and separated the winners from the agent-washing. Here is who is actually making money, how they price, and why most agent projects still fail.

By Christos Papadimitriou · House of 47 Research
Last updated:

See who wins
~40%/yrAI agent market growth, to ~$139B by 2034 (reported)
16%of enterprise AI deployments are true agents, Menlo 2025
>40%of agentic AI projects scrapped by 2027, Gartner
~130self-described agentic vendors Gartner counts as real
Abstract view of networked data and AI systems
The market · 2024 → 2030

Big money, thin reality

Enterprise AI spending tripled in a year, yet real agents are still a sliver of it and most pilots do not reach production. The money is concentrating fast in a handful of winners while the rest get acquired or quietly shelved.

2024 · The base

Spend starts

$11.5B

Enterprise generative-AI spend in 2024, by Menlo Ventures' count. Agents were barely a line item, mostly copilots and chatbots.

2025 · The surge

3.2x in a year

$37B

Enterprise AI spend in 2025. But true agents were only about $750M of it, roughly a tenth of app spend, and just 16% of deployments plan and act on their own.

2026 · Now

Consolidation

Buyouts

The winners pull away and the field thins. Intercom's Fin went to Salesforce, Cognigy to NICE, Forethought to Zendesk, Paradox to Workday. A cull of the weak is next.

2030 · The base case

Half the model spend

~50%

Gartner sees agentic AI driving about half of generative-AI model revenue by 2030, up from roughly 5% in 2025. The direction is clear; the casualty list is long.

Two markets worth the climb

AI agent software market, $B

Who is actually winning · latest reported valuation or ARR

The scorecard

The standouts, by headline number. A handful clear real revenue; the rest trade on momentum. Figures are the latest reported valuation unless marked ARR, and private numbers are company or press reported, not audited.

CompanyWhat it sellsHeadline

Sierra, Decagon, Harvey, Glean, Clay, Cursor, Abridge, Hippocratic, Replit, Mercor figures from 2025-26 funding reports and Sacra estimates. Claude Code, Agentforce and Fin are ARR, company or press reported. The seed list of 100+ sellers is the awesome-ai-employees dataset, extended with our own research.

The pattern

Valuation is running well ahead of revenue. Sierra, Decagon and Clay carry multi-billion prices on disclosed ARR in the tens to low hundreds of millions. The clearest real revenue sits in coding, where Anthropic's Claude Code reportedly passed $1B ARR within months and Cursor is reported near $1B to $2B.

The shakeout

Customer support is consolidating first. Four names from the seed list have already been bought: Intercom's Fin by Salesforce, Cognigy by NICE for about $955M, Forethought by Zendesk, Qualified by Salesforce. In recruiting, Paradox went to Workday and Moonhub was absorbed by Salesforce.

The field · 100+ sellers from the seed list, by category

From the list, by category

The seed was a public directory of more than a hundred companies selling AI agents to business. Every company from that list is here, grouped into its ten categories and linked to its site, together with our own theagency47. Names in bold are the clear leaders by traction.

Full list and categories from the awesome-ai-employees dataset. A handful of names in the scorecard above (Mercor, Glean, Clay, Abridge, Anthropic, Salesforce) sit outside this directory but belong on any map of the market, so we added them.

Cross-check · Menlo, McKinsey, Gartner, company filings

What the winners share

  • Win
    They price on outcomes, not seats

    Sierra, Zendesk at about $1.50 a resolution, and Intercom's Fin at $0.99 charge when the agent finishes the job, not per login. Around 39% of buyers now expect outcome-based pricing.

  • Win
    One job, done fully

    The winners own a narrow job end to end: a resolved support ticket, a merged pull request, a booked meeting. A general assistant that does a bit of everything rarely sells.

  • Win
    Deep integration into the stack

    Menlo and the system integrators agree the hard part is plumbing, not the model. The agents that stick plug into the customer's own tools, data and approvals.

  • Win
    A number the buyer can check

    Resolution rate, calls handled, PRs shipped. Demos do not close enterprise deals. A measured result the buyer can audit against a human baseline does.

  • Most
    Real logos, not pilots

    Decagon names Chime and Block, Harvey about half the Am Law 100, Hyro eight named health systems. Named production customers separate the winners from the demo-ware.

  • Most
    Own the workflow, rent the model

    Almost all run on Claude or GPT. The moat is the data, the integration and the domain, not a home-grown model. Vertical agents in law, health and finance win on trust. How fast that cost is falling is in The collapsing cost of intelligence, and which models they run is in Best AI models 2026, compared.

Where the field splits

How they charge

The clearest signal of a serious seller. The winners have moved off per-seat pricing.

Per seatPer outcome

Hype versus revenue

Gartner says only about 130 of thousands of self-described agentic vendors are real.

Agent-washingReal revenue

How you get one

From a rented product to a custom agent you own. Most mid-market teams land in the middle.

Rent a productBuild and own

Menlo Ventures 2025 State of Generative AI ($37B spend, $750M agents, 16% true agents) · McKinsey State of AI 2025 (88% use AI, 23% scaled agents) · Gartner 2025-26 (40% of agentic projects scrapped by 2027, ~130 real vendors, ~50% of model revenue by 2030) · outcome-pricing from Sierra, Zendesk, Intercom and buyer surveys.

Probability of outcome · House of 47 synthesis

The odds

Synthesis estimates across the sources above, not market forecasts. The through-line: the technology works in narrow jobs, the money is real but concentrated, and the failure mode is integration and value, not model quality.

Two ways to buy · product, in-house, or built and owned

How you actually get one

The market data points the same way twice: buyers are shifting to outcome pricing, and most projects fail on integration and value, not on the model. That is a problem of delivery, not technology, and it is exactly where the choice of how you buy decides the result.

Buy a productSaaS agent, per outcome or per seat

Fastest to start and proven in one lane, from Sierra in support to Harvey in law. But you rent it, it only does its lane, and outcome fees can become a billing black box you do not control. Rent

Build in-houseyour own team, your own stack

Full control and a real asset if the agent is core to your edge. The catch is talent and time: learning how to build an AI agent, then doing it, runs many months, and most mid-market teams do not have the people to spare. Own, slow

Build and own, with a partnercustom agents, your tools, your files

The pragmatic middle, and the fastest-growing way mid-market teams get agents, is what an AI agent agency does: build custom agents on frontier models, integrate them into your own stack, measure them on KPIs, and hand you the files to keep. Own, fast

Where this points

For a mid-market operator who cannot justify a Sierra-scale contract or an eighteen-month in-house build, a built-and-owned agent is usually the right answer. The competitor can buy the same off-the-shelf tool. It cannot buy the agent that knows your data and runs inside your workflow.

Our own bet

theagency47's AI agent services, House of 47's AI workforce company, sits exactly in this gap. It builds custom AI employees on Claude, integrated into a client's own tools and measured by KPIs, with public pricing from €2,500, EU-based delivery and a 14-day money-back guarantee. Not a rented seat, and not a year-long build. A working, owned agent in weeks. The ROI calculator shows the payback on that build.

Indicative figures for how much an AI agent costs, from 2025-26 market sources: SMB retainers roughly €500 to €3,500 a month, mid-market builds commonly €20,000 to €250,000 once integration and compliance are included. theagency47 figures from theagency47.com.

The model is a commodity. The winners sell a result, own the integration, and get paid when it works.

Sources · figures checked against primary reports and funding announcements

Sources

Changelog · what changed in this brief, newest first

Changelog