AI Agent Statistics 2026: Adoption, Production, and Trust

Every AI agent vendor’s homepage now claims “enterprises are adopting agents.” Few say how many, how fast, or what happens after the pilot. We pulled the actual numbers from the analyst firms, engineering surveys, and developer polls that measured it, plus our own structured index of 94 researched tools, and grouped them by theme below. Every figure is quoted from its original source, dated, and linked; nothing here is estimated.

The short version

Enterprise intent is far ahead of enterprise deployment. Gartner expects 40% of enterprise applications to embed task-specific AI agents by the end of 2026, up from under 5% in 2025. But McKinsey’s own 2025 survey found only 23% of organizations have actually scaled an agentic AI system into even one business function, with no single function above 10% scaled use.

Among teams that got agents into production, LangChain’s 2026 survey of 1,340 practitioners found 57% now run agents live, yet quality (not cost, not latency) is the top-cited reason projects stall. Developers show the same split: Stack Overflow’s 2025 survey of 33,662 respondents found AI tool usage at an all-time high (84%) while trust in the output dropped to 3.1% “highly trust.”

Adoption, production, and trust are three different numbers, and none of them move together.

Enterprise adoption: the intent-to-production gap

Metric	Figure	Source	Date
Enterprise apps expected to embed task-specific AI agents by end of 2026 (up from <5% in 2025)	40%	Gartner press release	2025-08-26
Agentic AI’s projected share of enterprise application software revenue by 2035, in Gartner’s optimistic scenario (up from 2% in 2025)	~30% (>$450B)	Gartner press release	2025-08-26
Agentic AI projects Gartner predicts will be canceled by end of 2027, citing unclear business value and weak governance	>40%	Gartner press release	2025-06-25
Organizations that have scaled an agentic AI system into at least one business function	23%	McKinsey, The State of AI	2025 survey, published Nov 2025
Organizations experimenting with agentic AI but not yet scaling it	39%	McKinsey, The State of AI	2025 survey, published Nov 2025
Ceiling on scaled agent use within any single business function, even at organizations that report scaling somewhere	<10%	McKinsey, The State of AI	2025 survey, published Nov 2025

Read together, these numbers describe the same gap from two directions. Gartner’s forecast is about feature availability (how many enterprise applications will ship an agent capability at all), and that number is rising fast. McKinsey’s survey measures actual scaled use inside a real workflow, and that number is a fraction of the first.

The forecast and the survey are not in conflict. They are measuring adoption at different depths, and the depth is where most of the reported failure sits. Gartner’s own >40%-cancellation prediction, from Anushree Verma, Senior Director Analyst: “Most agentic AI projects right now are early-stage experiments or proof of concepts that are mostly driven by hype and are often misapplied.”

What builders report once an agent ships

LangChain’s State of Agent Engineering report surveyed 1,340 practitioners between 2025-11-18 and 2025-12-02, published 2026-06-12.

• 57% of respondents have agents running in production; a further 30.4% are actively developing agents with a concrete deployment plan.
• ~89% have implemented observability for their agents: the most widely adopted engineering practice in the survey, ahead of formal evaluation.
• 52.4% run offline evaluations against a test set; 37.3% run online evaluations in production.
• Quality, not infrastructure cost, is the top-cited barrier to shipping (roughly one-third of respondents), with latency second (~20%). At organizations with 10,000+ employees, hallucinations and output consistency are the dominant quality complaint.

Source: LangChain, State of Agent Engineering, 2026-06-12.

Developer-side reality: usage is up, trust is down

Stack Overflow’s 2025 Developer Survey drew 33,662 responses (68.7% response rate), published on its AI section.

• AI tool usage among developers reached 84%, up from 76% the year before.
• 30.9% of developers use AI agents specifically (not just autocomplete) daily, weekly, or monthly; 13.8% stick to copilot/autocomplete mode only; 37.9% have no plan to adopt agents at all.
• Trust moved the opposite direction from usage: only 3.1% of developers say they highly trust AI-generated code’s accuracy, and 45.7% say they distrust it (somewhat or highly). Trust has fallen from over 40% “favorable” in 2024.
• Professional developers use AI tools daily at a higher rate (51%) than the survey population overall (47.1%), and report more favorable sentiment (61%) than developers still learning to code (53%).

The pattern across both surveys is the same shape: usage keeps climbing while confidence in the output does not follow it up, consistent with LangChain’s finding that quality, not access, is what blocks a pilot from becoming production.

Our own index, measured 2026-08-27

The Agents Index held 94 published agent and agentic-tool listings on 2026-08-27, across eight categories. Every figure in this section is computed directly from that corpus on that date, and each one states its own denominator.

Those denominators are not all 94, and the difference is deliberate. Where our research could not determine a field for a listing, that listing is dropped from the denominator rather than counted as a “no”. An undetermined field is not evidence of absence, and folding it into the negative bucket is the most common way a directory statistic ends up overstating its own case.

API availability: 72 of the 81 we could determine

We could determine API availability for 81 of the 94 listings, and 72 of those 81 offer a conventional REST API on some tier. The 13 undetermined listings publish no developer-facing documentation in either direction; several of them sit in coding agents, where a product can be a very good editor extension and never say whether anything can call it.

The distinction matters for anyone quoting this. Counting the 13 undetermined as “no” would give 72 of 94, or 77%, which reads as a claim that a quarter of agent tools have no API. What the data supports is narrower and stronger: among the tools where the answer is knowable from public documentation, an API is close to standard.

This is the figure that decides whether a tool can be wired into an existing system or only used standalone, which is why it is usually the first thing a buying team checks.

Every open-source tool we could check has an API

Split the API question by licence and it stops being close to standard and becomes absolute on one side. Of the 27 open-source listings where we could determine API availability, all 27 offer one. Of the 61 closed-source listings, we could determine API status for 54: 45 offer one and 9 don’t.

Nine listings in the index have no API at all, and all nine are closed-source products: Bolt.new, Crescendo, Decagon, Gemini Code Assist, Genspark, Replit Agent, Rox, Undermind and Unify. The remaining listings are undetermined on one side or the other and are excluded from both counts above.

There is a mechanical reason for the 27 of 27, and it is worth saying out loud. An open-source tool can be called because the code is there to call, so the real question is whether the vendor documents and supports an interface for it. On that question our count is a strong signal about integration risk, and it is a floor: the six open-source listings we could not determine may well offer one too.

Open source: 33 of 94, fully determined

All 94 listings now carry a recorded open-source determination, and 33 of them are open source. MindStudio, the one listing that lacked a determination last edition, has since been resolved: it is closed source.

Our open-source taxonomy tag now matches the researched attribute exactly, at 33. The gap flagged in the prior edition, where Kilo Code and Sim carried the researched attribute without the corresponding tag, has been closed.

Licensing inside that group is not uniform. Botpress is MIT-licensed outright, while n8n ships under fair-code terms and Dify under an Apache-2.0-derived licence. Cartesia releases its Line voice-agent SDK under Apache 2.0 while keeping the underlying speech models proprietary and hosted. “Open source” in this category covers a wider range of arrangements than the label suggests.

Free tier and pricing model

Both fields are meant to be populated for every published listing. Vellum has a genuine $0/month Base tier, so it counts toward both free tier and the freemium bucket. VoiceAgent has no free tier at all, its cheapest plan starts at $12.50/month, and it prices in flat monthly tiers rather than a pure per-minute meter, so it adds to the subscription bucket only. Clay left the index this edition and its page is now a 404, so its own two values are the one point in this section we can no longer verify against source. Lead Scorer, the one new listing since the prior edition, also has a genuine free tier, a permanently free CRM at €0/month sitting under its paid MCP-driven SDR tiers, so it adds to both free tier and the freemium bucket.

That puts free tier at 58 or 59 of the 94, and the pricing-model split at freemium 42 or 43, usage-based 23 or 24, subscription 17 or 18, and free-with-no-paid-tier 9 or 10 — in each pair, the lower figure is what stands if Clay held that bucket, and exactly one bucket is one lower than the other three. Every other figure in this section is a fresh count against the current corpus; this is the one place a departed listing leaves an unresolved bit rather than a clean number.

Capability tags: 48 autonomous, 35 multi-agent, as of the last full recount

These two figures are from our last full tag recount, 2026-08-19 against a 92-listing corpus, and were not independently re-tallied against the current 94 for this edition: unlike the researched fields above, we do not yet have Vellum’s, VoiceAgent’s, Lead Scorer’s or the departed Clay’s tag assignments confirmed against source for this pass. They should be read as what our taxonomy asserted as of that date rather than as a fresh census, an untagged listing has not been judged non-autonomous, it has simply not been tagged, and we hold these to a lower evidentiary bar than the researched fields above.

Botpress carries the autonomous tag on the strength of its Autonomous Engine, which plans and executes multi-step actions from a single instruction rather than following a fixed decision tree. CrewAI sits in the multi-agent slice, coordinating several agents rather than running one.

Category depth

Coding agents is the deepest category we track at 25 of 94, ahead of agent platforms (16), frameworks (14), voice agents (12), research agents (11), support agents (9), sales and marketing agents (8) and agent tools (2).

Those add to 97 rather than 94 because three listings, Botsify, Sierra and ElevenLabs Conversational AI, each sit in two categories. Category counts on this index are not a partition of the corpus and should not be summed to a total.

What this means

Three numbers matter more than the rest if you’re deciding whether to build on an agent today.

First, “adoption” without a qualifier is close to meaningless: ask whether the figure means feature availability (Gartner’s 40%), piloting (McKinsey’s 39%), or scaled production use (McKinsey’s 23%, capped under 10% per function), because vendors quote whichever is largest.

Second, the practices that correlate with getting past pilot (observability first, 89% in LangChain’s survey; evaluation second) are cheap relative to the agent itself, and worth doing before scaling headcount on top of an unmeasured system.

Third, developer trust is not a rounding error. A 3.1% “highly trust” rate on code output is a real adoption ceiling for autonomous tools specifically, distinct from the 84% usage figure vendors lead with.

There is a fourth point that applies to every statistic on this page, including ours. Ask what the denominator was. Our own API figure moves from 89% to 77% depending on whether undetermined listings are counted as absent, and that is a single index being careful with itself. A survey that does not tell you how many respondents skipped the question is doing the same arithmetic invisibly.

Methodology & sources

This is a sourced compilation, not original survey research. Its value is putting several independently-conducted studies next to our own directly-computed index numbers, with every figure dated and attributed to where it came from. Nothing here is estimated, extrapolated, or averaged across sources. Where two sources measure a related but different thing (Gartner’s forecast versus McKinsey’s measured scaling), we say so rather than blend them into one number.

The own-index figures are a dated measurement, not a live counter. They were computed on 2026-08-27 against the 94 listings published on The Agents Index at that moment, counting only listings in Published status. Population, open-source status, API availability and category membership were freshly cross-checked against source for this edition. Capability-tag counts were last fully checked on 2026-08-19 against 92 listings and are flagged as such above rather than presented as current. Free tier and pricing-model bucket are fresh for every listing except the departed Clay, whose page no longer resolves; that single unresolved value is stated as a range rather than guessed.

Each figure’s denominator is the number of listings for which that specific field could be determined, and it is stated alongside the figure. Listings with an undetermined field are excluded from that figure rather than counted in its negative bucket.

The corpus grows, so a later reading will differ, and any citation of these numbers should carry the 2026-08-27 date.

Primary sources: Gartner, Aug 2025 press release; Gartner, Jun 2025 press release; McKinsey, The State of AI; LangChain, State of Agent Engineering; Stack Overflow 2025 Developer Survey, AI section.