AI Agent Statistics 2026: 70+ Verified Numbers on Adoption, Capability and Reliability

Enterprise adoption

Claimed adoption is high; scaled, production-grade deployment is rare. The gap between the two is the single most important number in this page.

1. ~9 in 10 organisations report regular AI use in at least one business function [1]
2. 44% report AI scaling across the enterprise, up from 38% a year earlier [1]
3. 40% of large organisations (>$1B revenue) are scaling AI agents in at least one function, up from 27% the prior year [1]
4. 22% of smaller organisations are scaling agents — flat year over year [1]
5. ~2 in 10 respondents overall are scaling agents (31% at larger enterprises) [1]
6. 88% organisational AI adoption in 2025 (78% in 2024, 55% in 2023) [8]
7. 2% of organisations have deployed AI agents at scale; 12% at partial scale, 23% in pilots, 61% exploring [4]
8. 15% of business processes expected to reach semi- or full autonomy within 12 months [4]
9. <20% of organisations report high maturity in the data/technology infrastructure agentic AI needs [4]
10. 19% / 42% / 8% made significant / conservative / no investment in agentic AI; 31% waiting (poll of 3,412 webinar attendees, Jan 2025) [2]
11. 35% of adopters use agents broadly; 17% in almost all workflows and functions [5]
12. 68% of adopters say half or fewer of their employees interact with agents in everyday work [5]
13. 57% / 54% / 53% using or planning agents within six months in customer service / sales & marketing / IT & cybersecurity [5]
14. 46% of leaders say their company uses agents to fully automate workflows or processes [6]
15. 81% of leaders expect agents to be moderately or extensively integrated into AI strategy within 12–18 months [6]
16. 9.3% of surveyed leaders (844 of 9,037) meet Microsoft's 'Frontier Firm' criteria (org-wide AI plus agent use) [6]
17. 99% of enterprise developers surveyed are exploring or developing AI agents [7]
18. 1 in 5 companies have a mature governance model for autonomous AI agents [3]

ROI, cost and value

Individual productivity gains are near-universal in self-reports; enterprise-level financial impact is flat. Token cost has become a measurable constraint.

1. 80% of respondents say AI improved their individual productivity; 50% say it helps them make better decisions [1]
2. 37% report any positive EBIT contribution from AI — unchanged from 2025 [1]
3. 6% are 'AI high performers' (≥5% of EBIT attributed to AI) — flat year over year [1]
4. 32% decided against buying at least one software product because they could build it in-house with agentic coding tools (nearly 50% among high performers) [1]
5. ~20% say AI operating costs, including token costs, constrained their AI use [1]
6. 28% spend more than 10% of their enterprise ICT budget on AI; 60% expect to increase AI investment next year [1]
7. 66% / 40% / 20% of organisations report productivity gains / reduced costs / increased revenue from AI [3]
8. 74% hope to grow revenue through AI in future, versus 20% already doing so [3]
9. 88% of US executives plan to increase AI budgets in the next 12 months because of agentic AI; over a quarter by 26% or more [5]
10. $450B potential economic value (revenue growth + cost savings) from AI agents by 2028 across surveyed markets [4]
11. $2B Cursor annual recurring revenue, doubled in three months (Mar 2026) [21]
12. 1.4–2× median self-reported change in value of work from AI tools among 349 technical workers (early 2026) [13]

Developers and engineering teams

Developer usage is bimodal: a minority run agents daily, a plurality has no plans to. Teams that ship agents overwhelmingly instrument them.

1. 57.3% of respondents have agents in production (51% in the 2024 survey); a further 30.4% are actively developing with plans to deploy [10]
2. 67% vs 50% production agent rate at 10,000+ employee organisations vs under-100 employee organisations [10]
3. 26.5% / 24.4% / 18% primary use case: customer service / research & data analysis / internal workflow automation [10]
4. 32% cite output quality as the top barrier to production; latency is second at 20% [10]
5. 89% have some form of agent observability; 62% trace individual steps and tool calls (94% and 71.5% among teams in production) [10]
6. 52.4% / 37.3% run offline / online evaluations; 29.5% run no evaluations at all [10]
7. 59.8% / 53.3% of evaluating teams use human review / LLM-as-judge [10]
8. >75% use multiple models across production and development; 57% do no fine-tuning [10]
9. 84% of developers use or plan to use AI tools (76% in 2024); 47.1% use them daily [11]
10. 14.1% use AI agents daily, 9% weekly, 7.8% monthly; 37.9% have no plans to [11]
11. 52% say agents or AI tools changed how they work in the past year (16.3% 'to a great extent') [11]
12. 83.5% of developer agent use is in software engineering; 24.9% data/analytics, 17.6% business-process automation [11]
13. 34.4% of developers building agents use MCP servers; 28.1% use multi-agent orchestration [11]
14. 87% / 81% are concerned about agent accuracy / security & privacy [11]
15. 46% vs 33% distrust vs trust the accuracy of AI tools; only 3.1% highly trust [11]
16. 81.7% / 67.9% / 40.8% of developers using out-of-the-box agents use ChatGPT / GitHub Copilot / Claude Code [11]
17. 27%→39% share of Claude.ai conversations that are 'directive' (user delegates the whole task), late 2024 to mid 2025 [19]
18. 77% of business API usage follows automation patterns, versus ~50% on Claude.ai [19]
19. 986M code pushes on GitHub in 2025; a new developer joins every second [20]

Capability and benchmarks

Benchmark scores move fast, but reliability is what production needs. Use the 50%-horizon and pass^k framing rather than single-shot leaderboard scores.

1. ≈7 months doubling time of the length of tasks (in human time) frontier agents complete at 50% reliability, consistent 2019–2025 [12]
2. ~100% / <10% model success on tasks taking humans under 4 minutes / over ~4 hours (Mar 2025) [12]
3. <3 months doubling time measured on SWE-bench Verified alone [12]
4. 2h 17m GPT-5 50%-time horizon on METR's software task suite (Aug 2025) [13]
5. ⅓ / ⅓ / ⅓ on 90-min–3-hour tasks GPT-5 succeeds always / never / inconsistently — roughly one third each [13]
6. 16 h ceiling above which METR's current task suite cannot reliably measure horizons [13]
7. 12%→~66% OSWorld computer-use success in one year; agents still fail ~1 in 3 attempts [8]
8. 60%→~100% SWE-bench Verified top score across 2025 [8]
9. 67.3 pts one-year gain on SWE-bench, 2023→2024 — the largest jump of any tracked benchmark [9]
10. 12.47% of the 2,294 SWE-bench issues resolved by SWE-agent at launch (Mar 2024) [16]
11. 65% of SWE-bench Verified resolved by mini-SWE-agent, an agent implemented in 100 lines of Python (Jul 2025) [16]
12. 72.36% of OSWorld's 369 real computer tasks completed by humans, vs 12.24% by the best agent at launch [15]
13. 30% of consequential real-world work tasks completed autonomously by the best agent on TheAgentCompany [17]
14. 362 documented AI incidents in 2025, up from 233 in 2024 [8]
15. 50.1% accuracy of the top model at reading an analog clock — the 'jagged frontier' [8]

Reliability, trust and governance

Single-run success overstates what an agent will do in production. Consistency across repeated runs and human trust both decay faster than headline scores suggest.

1. <50% average single-attempt (pass^1) success of the best of 12 agents (GPT-4o) across τ-bench retail and airline domains [14]
2. ~25% GPT-4o pass^8 on τ-retail — probability of solving the same task 8 times in a row; a ~60% relative drop from pass^1 [14]
3. 43%→27% trust in fully autonomous AI agents, one-year decline [4]
4. >40% of agentic AI projects will be cancelled by end of 2027 due to cost, unclear value or inadequate risk controls [2]
5. 15% of day-to-day work decisions will be made autonomously by agentic AI by 2028, from 0% in 2024 [2]
6. 33% of enterprise software applications will include agentic AI by 2028, from under 1% in 2024 [2]
7. 20% of executives trust agents with financial transactions; 22% with autonomous employee interactions [5]
8. 45% / 42% of adopters are fundamentally rethinking operating models / redesigning processes around agents [5]
9. 67% vs 40% leaders vs employees who are familiar with AI agents [6]
10. 24.9% of 2,000+ employee enterprises rank security as their second-largest agent concern [10]
11. 2,562 real-world MCP servers analysed across 23 categories; 1,438 call network APIs and 1,237 system-resource APIs [18]
12. 38% of organisations expect AI agents to be team members inside human teams by 2028 [4]

Workforce and labour

Expected job losses consistently exceed realised ones. Hiring for agent-specific roles is growing faster than displacement.

1. 39% vs 14% expect AI to reduce headcount next year vs those reporting an actual AI-driven decline in the past year [1]
2. 13% of respondents say AI makes them anxious about their career prospects [1]
3. 47% vs 31% of mid-level staff vs executives report experiencing a negative effect of AI at work [1]
4. 82% of leaders are confident they will use digital labour to expand workforce capacity within 12–18 months [6]
5. 32% of managers plan to hire AI agent specialists in the next 12–18 months; 28% plan to hire AI workforce managers for human-agent teams [6]
6. 78% / 95% of leaders / Frontier Firm leaders are considering hiring AI-specific roles [6]
7. 67% of executives agree agents will drastically transform existing roles within 12 months; 48% expect headcount to increase as a result [5]
8. 53% / 48% / 36% of organisations respond with workforce AI-fluency education / upskilling / hiring specialised AI talent [3]
9. 40% of US employees use AI at work, up from 20% in 2023 [19]
10. 73% vs 23% AI experts vs the public expecting a positive impact of AI on jobs [8]

Formulas and derived metrics

Derived quantities computed from the primary figures above. Use them to reason about agents rather than just quote them.

Frequently asked questions

Methodology

Primary sources are used wherever they exist: the organisation that ran the survey, benchmark or measurement. Secondary aggregators, vendor landing pages without a named methodology, and market-size forecasts without a disclosed model are excluded; the few secondary reports kept are labelled as such in the source list. Each figure is reproduced as stated by the source, with sample size and fielding dates where the source discloses them. Survey figures are self-reported by respondents and are not independently verified; benchmark figures depend on the harness and prompt used and should be compared only within a benchmark. The page is re-checked and updated as new editions of each report are published.

Sources