Why AI agent projects fail, and what the survivors did differently

A note on demand

This is a Track-B page. No keyword row in our research database matches "why AI agent projects fail": there is no measured search demand for the question yet. We wrote it anyway because the evidence base points to a 2027–28 disillusionment phase, and because the adjacent commercial terms are moving: "enterprise ai agents" draws about 2,900 US searches a month at a cost per click of USD 103 and grew 1,273% year on year, while "ai agents news" (1,300 a month) fell 94% (DataForSEO, retrieved 2026-11-09). People are buying, not browsing. The question on this page is the one they ask after the pilot.

The numbers: pilots versus production

What was measured	Result	Grade	Source
Organisations scaling an agentic AI system somewhere in the enterprise	23%; a further 39% experimenting; no single business function above 10% scaling	B	[1]
Organisations attributing any EBIT impact to AI	39%; about 6% report 5% or more	B	[1]
Organisations exploring autonomous agents "to a large extent"	26%, with regulatory compliance the top barrier for 38%	A	[2]
Companies planning to deploy agentic AI within two years (2026 survey)	Close to three-quarters; only 21% report a mature model for agent governance	A	[2]
CEOs who saw GenAI profitability gains in the past year	34%, against 46% who had expected them	A	[3]
GenAI pilots delivering measurable P&L impact	"95% of organizations are getting zero return"; "Just 5% of integrated AI pilots are extracting millions in value"	C	[4]
Agentic AI projects cancelled by end-2027 (forecast)	"Over 40%", for cost, unclear business value or inadequate risk controls	C	[5]

Read the grades. The McKinsey and Deloitte rows are large executive surveys; the first is read through secondary coverage because the primary page was unreachable. The MIT NANDA figure that travelled furthest in 2025 comes from a preliminary report based on 52 interviews, a survey of 153 leaders and a review of 300 public announcements, with success defined by the authors; we grade it C and quote it because it is quoted everywhere, not because it is precise. The Gartner forecast is a press-release headline we could not read behind the firewall. None of the rows contradicts the others: lots of pilots, few in production, and a governance gap that the people running them acknowledge.

The failure rate is not evidence that agents do not work. It is evidence that organisations are buying something they have not yet learned to run. Our explainer on autonomous AI agents covers what the systems can do; this page covers what goes wrong around them.

Pattern 1: the wrong task

The most common failure is choosing a task because it is visible rather than because it is suited. The field evidence is specific about suitability. In a randomised experiment with 758 consultants, GPT-4 improved speed and quality on tasks inside the model's competence and made people 19 percentage points less likely to be correct on a task designed to sit just outside it; users could not tell which side they were on.^[6] B A meta-analysis of 106 experiments found human–AI combinations underperformed the better of the two alone on average, with losses concentrated in decision tasks and gains in content creation.^[7] A The MIT NANDA interviews report budgets flowing to visible, top-line functions over the "high-ROI back office", and external partnerships succeeding at twice the rate of internal builds.^[4] C

What survivors did: they chose bounded tasks with checkable outputs (classification, drafting, data movement between systems) and left judgement tasks for later. Our evidence review of AI agents for business ranks use cases by the evidence of payoff.

Pattern 2: no evaluation harness

A demo shows one run. Agents are inconsistent across runs, and a project that cannot measure that inconsistency cannot know whether it is improving. The τ-bench authors introduced pass^k, the probability that an agent succeeds on all k attempts at the same task, precisely because single success rates hide it; in their retail domain, even state-of-the-art agents succeeded on fewer than half of tasks and pass^8 fell below 25%.^[8] B On OSWorld, agents driving real desktop software completed 12% of tasks against more than 72% for humans at publication, mostly failing at the interface rather than the reasoning.^[9] B Newer models score higher; the point is that without your own task set and repeat counts you have no number at all.

What survivors did: before the pilot, they wrote 30 to 50 representative tasks with known correct outcomes, ran the agent on each five times, and set a threshold for production. Every model or prompt change re-runs the set.

Pattern 3: cost blow-ups

An agent's cost has three parts: model tokens multiplied by steps and retries, platform fees per run or per action, and the human time spent reviewing. Projects budget the first, meter the second badly and ignore the third. An agent that loops on a failing tool call can consume a month of budget overnight on a platform that bills per action step, which is why billing units matter; the same five-step workflow costs between EUR 5 and USD 89 per thousand runs depending on whether the platform bills per run or per step. And the review line scales with the error rate: at a 50% completion rate, half the runs need correction, and the expected cost per completed task roughly doubles.

What survivors did: hard budgets per run (tokens, steps, wall-clock time), a kill switch, and a review-minutes line in the business case from day one.

Pattern 4: data access and permissions

Agents need access to be useful and access is where they do damage. Two failure modes recur: the agent is given a human's full credentials because scoping them was slow, and the agent reaches data nobody realised was connected. Deloitte's finding that only 21% of companies planning agents have a mature governance model is this pattern in survey form.^[2] A

What survivors did: a service account per agent with the minimum scopes, read-only for the first month, draft-only for anything that sends, and a log of every action the agent took that a person can search. Our guide to managing AI agents sets out the permission boundaries and the brief format.

Pattern 5: the human in the loop is a rubber stamp

Putting a person on the approve button does not make a system safe if the person stops reading. In a controlled study, participants with an automated aid made more errors when the aid was wrong than participants with no aid, even with contradicting information on screen.^[10] B A systematic review of 74 studies documents the effect across aviation, medicine and process control, and finds that training, display design and accountability reduce it.^[11] A An integrative review in Human Factors explains complacency and bias as attentional effects of trust and workload.^[12] A A gate that is pressed 400 times a day is not a gate.

What survivors did: sampled blind review rather than approve-everything; named accountability for each output; agreement-rate tracking between the agent's proposal and the reviewer's decision, so that a reviewer who never disagrees is visible.

Pattern 6: vendor lock-in

Agent products are young, and the formats in which prompts, tools, memory and run logs are stored are proprietary. Projects that build deeply on one vendor's agent framework discover at renewal or at a price change that the work does not move. The adjacent evidence: analysts expect agentic features in a third of enterprise applications by 2028,^[13] C which means every suite vendor is building its own, and the odds that your chosen one is the one that persists are not high.

What survivors did: kept task definitions, evaluation sets and logs in their own storage in plain formats; preferred tool interfaces based on open protocols; and ran at least one task on a second provider to know the switching cost before it mattered. Self-hostable platforms such as n8n reduce one dimension of the risk and add an operations cost; our comparison of the best AI agents for work notes which products export what.

Pattern 7: compliance surprises

The regulatory barrier that 38% of Deloitte's respondents named as the top one has sharpened since.^[2] Under the EU AI Act, AI used for recruitment, task allocation or "monitoring and evaluating performance" of workers is high-risk (Annex III, point 4), and inferring emotions of people in the workplace is prohibited.^[14] B Deployers of high-risk systems must follow the provider's instructions and "assign human oversight to natural persons who have the necessary competence, training and authority"; employers must inform workers and their representatives before use.^[15] B After the 2026 Omnibus amendment these Annex III obligations apply from 2 December 2027, while the AI-literacy duty has applied since 2 February 2025.^[16] A Separately, GDPR Article 22 gives people a right not to be subject to solely automated decisions with legal or similarly significant effects, and the CJEU held in the SCHUFA case (C-634/21) that an automated score a third party "draws strongly" on is itself such a decision, with a right to human intervention.^{[17] [18]} A

An agent that triages job applications, allocates tickets to named people by predicted performance, or scores customers for credit is inside these rules even when a person clicks the last button. Projects discover this at the legal review, after the build.

What survivors did: ran the Annex III and Article 22 test on the use case before procurement, kept agents on the work rather than on the workers, and built the human-intervention route into the workflow rather than into a policy document.

A readiness checklist

Score one point per yes. Below seven, run a longer pilot; nine or more, you are ahead of most of the survey respondents above.

1. The task is bounded, repeated and has an output a person can check faster than they could produce it.
2. We have written 30 or more test tasks with known correct outcomes.
3. We have run the agent on them at least five times each and recorded the pass rate and the consistency.
4. The agent has its own account with the minimum scopes, read-only or draft-only to start.
5. Every action is logged in storage we own, in a format we can read without the vendor.
6. There are hard limits per run on tokens, steps and time, and a kill switch someone owns.
7. The business case includes reviewer minutes per output.
8. Review is sampled and blind, with a named accountable person per output, and we track the agreement rate.
9. We have checked the use case against AI Act Annex III and GDPR Article 22 and documented the result.
10. We can run the same task on a second provider and know what it would cost to switch.

A printable version of the checklist is in preparation; it is not downloadable yet. The Productivity Signal briefing announces it when it is ready and subscribers receive it first. The ten points above are the whole content.

Prediction log: what we expect by 2028

We commit to checkable statements with dates. Each is logged in our forecast model's prediction file. We will check them by hand at the dates below and publish the results, right or wrong, on the Future Demand Radar.

#	Prediction (logged 2026-11-09)	Check date	Baseline expectation	What would falsify it
P1	US search demand for the "AI agents and delegated work" category (seven tracked terms, 13,530 searches a month in November 2026) reaches 1.73 times today's level	2028-12-31	Band 1.47–2.04 under scenario B	A multiplier below 1.47 (scenario A) or above 2.04 (scenario C) on the same DataForSEO terms
P2	The share of organisations "scaling" an agentic system in McKinsey's annual survey rises from 23% to between 30% and 40%	2027-12-31	Baseline scenario expects steady scaling without a collapse	Below 28% (disillusionment deeper than expected) or above 45% (acceleration)
P3	The Google Trends share of "agentic ai" within our future-skills group, which rose 4.3 points in the past twelve months, falls for at least two consecutive quarters before end-2028	2028-12-31	Our baseline scenario expects a 2027–28 trough consistent with the cancellation forecast	Continuous share growth through 2028
P4	At least one EU data-protection authority decision or court ruling applies GDPR Article 22 to a decision taken by an LLM-based agent in an employment or customer context	2027-12-31	Baseline: likely, given the SCHUFA reasoning	No such decision by the date

In our radar, "AI agents and delegated work" scores 59.7 out of 100 in the baseline scenario (p10–p90 band 55.2–64.1) and is classed "build now"; under scenario A it scores 56.0 and stays "build now", under scenario C 61.7. Our baseline scenario expects demand roughly 1.7 times today's by 2028 and 3 times by 2030 (band 2.3–3.9). Forecasts are model outputs with stated uncertainty, not facts, and the point of the table is that they can be wrong in public.

What to do next

Run the checklist on your pilot before the next budget cycle. If it scores below seven, the cheapest fix is usually pattern 2: write the test tasks. Then use the Tool Finder to shortlist agent platforms by the permissions model, billing unit and data location that the checklist asks about; it returns a shortlist with prices dated to the day we read them. The briefing tracks the survey numbers in the first table and the four predictions above, and tells you when either moves.

Frequently asked questions

What share of AI agent projects fail?

It depends on the definition. McKinsey's 2025 survey finds 23% of organisations scaling an agentic system and 39% experimenting;^[1] B a preliminary MIT NANDA report says 95% of organisations see no measurable return from GenAI pilots, based on 52 interviews and 153 survey responses, which we grade C;^[4] Gartner forecasts over 40% of agentic projects cancelled by end-2027, a press-release headline we also grade C.^[5] The consistent picture is many pilots, few in production.

What is the most common reason AI agent pilots fail?

Task choice. Projects pick visible tasks rather than suitable ones. The field evidence shows gains inside the model's competence and losses just outside it,^[6] B and a meta-analysis finds human–AI combinations underperform the better of the two on decision tasks.^[7] A Bounded, checkable, repeated tasks survive; judgement tasks do not, yet.

How do I evaluate an AI agent before deploying it?

Write 30 to 50 representative tasks with known correct outcomes, run the agent on each at least five times, and record both the pass rate and the consistency across runs (the τ-bench pass^k idea).^[8] B Set a threshold for production and re-run the set after every model or prompt change.

Does the EU AI Act apply to AI agents at work?

If the agent is used for recruitment, task allocation or monitoring and evaluating worker performance, it is high-risk under Annex III, with deployer duties including competent human oversight and informing workers, applying from 2 December 2027; emotion inference at work is prohibited already.^{[14] [15] [16]} B An agent that summarises meetings or moves data is not high-risk; one that ranks people is heading that way. GDPR Article 22 applies to solely automated decisions regardless.

Why does a human approver not prevent agent errors?

Because of automation bias: people with an automated aid make more errors when the aid is wrong than people with no aid, as they stop checking.^{[10] [11]} A A button pressed hundreds of times a day is not oversight. Sampled blind review, named accountability and tracking how often the reviewer disagrees with the agent are the mitigators the evidence supports.

Sources and evidence

1. The state of AI in 2025: Agents, innovation, and transformation (as reported by ppc.land and idm.net.au) — McKinsey & Company (QuantumBlack) (2025), report B ppc.land/most-companies-still-pilot-ai-programs-despite-widespread-ado (retrieved 2026-11-09)
2. The State of Generative AI in the Enterprise (Q4 report, Jan 2025) and State of AI in the Enterprise 2026 (press release, Jan 2026) — Deloitte (2026), report B www.deloitte.com/us/en/about/press-room/state-of-ai-report-2026.html (retrieved 2026-11-09)
3. 28th Annual Global CEO Survey (press release) — PwC (2025), report B www.pwc.com/gx/en/news-room/press-releases/2025/pwc-2025-global-ceo-su (retrieved 2026-11-09)
4. The GenAI Divide: State of AI in Business 2025 (preliminary findings; 52 interviews, 153 survey responses, 300+ public initiatives) — Challapally, Aditya; Pease, Chris; Raskar, Ramesh; Chari, Pradyumna (MIT NANDA) (2025), report C mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Business_2025_Report. (retrieved 2026-11-09)
5. Gartner predicts over 40% of agentic AI projects will be canceled by end of 2027 (press release; headline verified via search only, page returned 403) — Gartner (2025), report C www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts (retrieved 2026-11-09)
6. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — Dell'Acqua, Fabrizio; McFowland III, Edward; Mollick, Ethan R.; et al. (2023), RCT B doi.org/10.2139/ssrn.4573321 (retrieved 2026-11-09)
7. When combinations of humans and AI are useful: A systematic review and meta-analysis — Vaccaro, Michelle; Almaatouq, Abdullah; Malone, Thomas (2024), meta-analysis A doi.org/10.1038/s41562-024-02024-1 (retrieved 2026-11-09)
8. τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains — Yao, Shunyu; Shinn, Noah; Razavi, Pedram; Narasimhan, Karthik (2024), report B doi.org/10.48550/arXiv.2406.12045 (retrieved 2026-11-09)
9. OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments — Xie, Tianbao; Zhang, Danyang; Chen, Jixuan; et al. (2024), report B doi.org/10.48550/arXiv.2404.07972 (retrieved 2026-11-09)
10. Does automation bias decision-making? — Skitka, Linda J.; Mosier, Kathleen L.; Burdick, Mark (1999), RCT B doi.org/10.1006/ijhc.1999.0252 (retrieved 2026-11-09)
11. Automation bias: a systematic review of frequency, effect mediators, and mitigators — Goddard, Kate; Roudsari, Abdul; Wyatt, Jeremy C. (2012), systematic review A doi.org/10.1136/amiajnl-2011-000089 (retrieved 2026-11-09)
12. Complacency and Bias in Human Use of Automation: An Attentional Integration — Parasuraman, Raja; Manzey, Dietrich H. (2010), systematic review A doi.org/10.1177/0018720810376055 (retrieved 2026-11-09)
13. Gartner: 33% of enterprise software applications will include agentic AI by 2028 (press release; search snippet only) — Gartner (2025), report C www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts (retrieved 2026-11-09)
14. EU AI Act, Article 5 and Annex III (prohibited practices; high-risk employment systems) — Regulation (EU) 2024/1689 (text via artificialintelligenceact.eu) (2024), report B artificialintelligenceact.eu/annex/3/ (retrieved 2026-11-09)
15. EU AI Act, Article 26: Obligations of deployers of high-risk AI systems — Regulation (EU) 2024/1689 (text via artificialintelligenceact.eu) (2024), report B artificialintelligenceact.eu/article/26/ (retrieved 2026-11-09)
16. Regulatory framework for AI: application timeline (as amended 2026) — European Commission (2026), report A digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai (retrieved 2026-11-09)
17. Regulation (EU) 2016/679 (GDPR), Article 22 — European Parliament and Council (text via gdpr-info.eu) (2016), report B gdpr-info.eu/art-22-gdpr/ (retrieved 2026-11-09)
18. Judgment of 7 December 2023, OQ v Land Hessen (SCHUFA Holding), Case C-634/21 — Court of Justice of the European Union (2023), report A collab.dpa.gr/wp-content/uploads/2023/12/CJEU-C-634-21.en_.pdf (retrieved 2026-11-09)