"Why most AI agent projects stall before production, what Gartner, MIT, RAND and S&P Global actually measured, where the 80%, 85% and 95% figures come from, and how to choose a vendor, write a contract and decide whether to kill, fix or restart a failed AI pilot."
Key Takeaways
- 1S&P Global (May 2025, 1,006 respondents in North America and Europe): 42% of companies abandoned most of their AI initiatives before production, up from 17% a year earlier, and on average 46% of projects were scrapped between proof of concept and broad adoption.
- 2Gartner (25 June 2025) predicts over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. It is a forecast; nobody has measured the 2027 outcome yet.
- 3RAND (August 2024) cites estimates that more than 80% of AI projects fail, about twice the rate of non-AI IT projects, and names a misunderstood problem as the most common root cause.
- 4MIT NANDA (July 2025) found 95% of organizations getting zero return from generative AI. It measured organizations and generative AI in general, not AI agent projects.
- 5The "Gartner 85%" figure is a February 2018 prediction that 85% of AI projects would deliver erroneous outcomes through 2022. It was never a measured failure rate.
- 6Most failures are decided before the build starts: one narrow job, live data access, a written pass mark, human approval on risky actions and a kill date prevent most of them.
- 7A failed pilot is worth fixing when the job matters, the data exists and the gap is scope or testing; it should be stopped when the problem is too hard for current models or nobody uses it.
The short answer
Most AI agent projects fail before they ever reach production. S&P Global found 42% of companies abandoned most of their AI initiatives before production in 2025, up from 17% a year earlier. The usual causes are a vague job, data the agent cannot reach, no agreed test of good enough, and nobody owning the agent after launch.
We build custom AI agents, so weigh this page with that in mind. Every outside figure links to the page we read it on, and each was re-read on 18 October 2026. Where a widely quoted number does not say what people think it says, we quote what the source does say.
How we made this page
On 18 October 2026 we read the top 10 Google results and the People Also Ask questions for the main searches ("why do ai projects fail", "why ai agent projects fail", "ai project failure rate") in the US, UK and Australia, using DataForSEO. The same day we asked 36 buyer questions across ChatGPT, Gemini, Claude, Perplexity and Google's AI answers, and read the pages they cited. Then we opened each primary source (Gartner, MIT NANDA, RAND, S&P Global, NIST, the ICO and the OAIC) and copied the exact wording.
People type the question many ways: "why do AI projects fail", "what percentage of AI projects fail", "how many AI projects fail", "what percent of AI projects fail", "AI project failure rate", "MIT AI projects fail". When we asked ChatGPT, Claude and the others why AI agent projects fail, they went looking for three names: Gartner, MIT and RAND. Most pages that rank quote at least one of the three, and several quote them in ways the source does not support. This guide gives each figure with its exact wording, what it measured and who it sampled. Then it covers the seven causes behind most failures, the vendor-side causes nobody likes to name, a 10-point checklist for hiring, and what to do when your pilot has already stalled.
It is written for buyers at companies of 10 to 50 staff, 50 to 500 and 500+ in the US, UK and Australia who are about to hire an AI development firm in 2026, or who already did and are looking at a pilot that never shipped.
What percentage of AI projects fail? What the research says, and what it does not
Four sources do most of the work in every article on this topic. Here is what each one says, word for word where we can quote it, and what it measured.
| Source | What it says | What it measured | Use it as |
|---|---|---|---|
| S&P Global, Voice of the Enterprise: AI & ML Use Cases 2025 (30 May 2025) | Companies "abandoning the majority of their AI initiatives before they reach production" rose "from 17% to 42% year over year"; on average "46% of projects are scrapped between proof of concept and broad adoption". | A survey of 1,006 midlevel and senior IT and line-of-business professionals across North America and Europe. | The best measured AI project failure rate. Name the sample region. |
| Gartner press release (25 June 2025) | "Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls." | A prediction. The 3,412-person poll in the same release measured how much firms had invested, not cancellations. | A forecast about agentic AI. It measures no failure rate. |
| Gartner article by Arun Chandrasekaran (26 January 2026) | "At least 50% of generative AI projects were abandoned after proof of concept due to poor data quality, inadequate risk controls, escalating costs or unclear business value." | Gartner says it analyzed hundreds of generative AI implementations. It does not publish the sample. | Generative AI after proof of concept, as of the end of 2025. |
| MIT NANDA, The GenAI Divide: State of AI in Business 2025 (July 2025) | "95% of organizations are getting zero return." External partnerships reached deployment "~67% of the time, compared to ~33% for internally built tools". | 52 interviews, 153 survey responses and 300+ public initiatives. Success meant deployment beyond pilot with measurable KPIs, 6 months after the pilot. The authors call the figures "directionally accurate". | Generative AI in general, counted by organization. Not "95% of pilots fail". |
| RAND, RR-A2680-1 (13 August 2024) | RAND writes that, by some estimates, more than 80 percent of AI projects fail, which it puts at twice the failure rate of IT projects without AI. | 65 interviews with data scientists and engineers with at least five years of experience. The 80% is other sources' estimate, footnoted. | Five root causes, measured. Write "RAND cites estimates of more than 80%". |
So what is the real AI failure rate? The honest answer for 2026 is a range of measured and predicted figures, each about a different thing. If you need one number for the percentage of AI projects that fail, for a board paper, use S&P Global's: 46% of projects scrapped between proof of concept and broad adoption, from 1,006 respondents. It was measured, it says who was asked, and it describes the stage where most agent projects die.
Three figures people quote that the sources do not support
- "Gartner says 85% of AI projects fail." The only Gartner source we found is a press release of 13 February 2018, which predicted that through 2022, 85 percent of AI projects would deliver erroneous outcomes due to bias in data, algorithms or the teams managing them. It forecast wrong outputs, five years out. It did not count failed projects.
- "MIT says 95% of AI pilots fail."MIT NANDA says 95% of organizations are getting zero return from generative AI. It counts organizations, it is about generative AI broadly, and success meant measurable P&L impact six months after the pilot.
- "RAND found 80% of AI projects fail." RAND reports that estimate from other sources and footnotes it. What RAND measured is why projects fail, from 65 interviews.
Why do most AI projects fail? Seven causes behind AI agent failures
The first five come from RAND's root causes, which held up across its 65 interviews. The last two are specific to AI agents, which act in your systems rather than only answering questions. Each one is decided before or during the build, which is the useful news: the buyer can prevent most of them.
- The problem is fuzzy.This is the most common reason for AI project failure. RAND says misunderstandings and miscommunications about the intent and purpose of a project are the most common reasons AI projects fail. An agent briefed to "handle customer service" has no finish line. An agent briefed to answer "where is my order" from Shopify and the carrier's tracking data, with a target share of tickets closed without a person, does.
- The data is missing, messy or locked away.RAND's second cause is lacking the data to train an effective model. For agents the usual version is access: the stock level lives in NetSuite, the price rules in a spreadsheet, the customer history in HubSpot, and the agent can reach only one of them. Gartner's 2026 article lists data that is not ready as failure point number two.
- The technology came first.RAND's third cause is focusing on the latest technology over the user's problem. Gartner makes the same point about agents: "Many use cases positioned as agentic today don't require agentic implementations." If a fixed workflow or a simple integration does the job, an agent adds risk without adding value.
- There is nowhere to run it.RAND's fourth cause is weak infrastructure to manage data and deploy models. For an agent that means logging, access control, secrets management, a staging copy of your systems to test against, and somebody watching it in production.
- The task is too hard for current models.RAND's fifth cause. Some jobs need judgment a model cannot give reliably: a disputed insurance claim, a contract negotiation, a safety decision on a factory line. An agent can prepare the work. A person should decide.
- No human approval step on risky actions.An agent that can issue a refund, post a journal entry or email a customer can do damage at speed. OWASP's 2025 Top 10 for LLM applications calls this Excessive Agency and recommends human-in-the-loop control so a person approves high-impact actions before they happen. Projects without that step tend to be shut down after the first expensive mistake.
- No evaluation set, so nobody can tell if a change helped. An evaluation set is a fixed list of real past cases with the right answer for each. Without one, every prompt tweak or model upgrade is a guess, quality drifts, and trust drains away until users quietly stop using the agent.
Other ranking pages group causes differently. IBM's October 2025 piece by Neil Dharcalls it the science experiment trap and cites IBM's 2025 CEO Study: just 16% of AI initiatives have achieved scale at the enterprise level. The University of Queensland's interview with Dr Evan Shellshearnames the two biggest causes as not building a need for the project inside the organisation, and not having the right data in the right place at the right quality. Informatica, LexisNexis and Cutter also rank for "why do AI projects fail" in the US with their own lists. The causes overlap more than the numbering suggests.
Which causes sit with the vendor?
Research reports describe failure from the buyer's side. When a firm is hired to build the agent, four more causes show up, and each is visible before you sign.
| Vendor-side cause | What it looks like | What to ask for instead |
|---|---|---|
| Demo on sample data only | A polished demo on the vendor's own examples; your data first appears after the contract | A demo on a sample of your own emails, invoices or tickets before you sign |
| Nobody owns production | The build team leaves at launch; no named person watches errors or user feedback | A named owner after launch and a written support option, priced per change |
| Keys and code in the vendor's accounts | The repository, model API keys and cloud accounts sit in the vendor's name | Code, prompts and accounts in your name from day one |
| Agent washing | A chatbot or RPA script sold as an agent | A live run where the agent reads your data, decides and acts, with logs |
Agent washing is Gartner's term. Its June 2025 release defines it as the rebranding of existing products, such as AI assistants, robotic process automation and chatbots, without substantial agentic capabilities, and estimates that only about 130 of the thousands of agentic AI vendors are real. Our guide to questions to ask an AI development company turns these four causes into interview questions.
See a demo on your own data before you sign
Send us the one job you want an agent to do and a handful of real examples. We will show it working on your own data before you sign anything, and FactoryJet quotes a fixed price in writing after a short scoping call. Founder replies within 24 hours.
See a demo on my dataHow do you avoid AI project failure when you hire? A 10-point checklist
S&P Global found that organizations with lower failure rates were more likely to weigh compliance, risk and data availability when choosing projects. This checklist turns that into ten things to settle in writing before the build starts.
- One job. Name a single task with a start and an end, such as "match supplier invoices to purchase orders in Xero".
- Your data in the demo. The vendor runs the agent on a sample of your real records before you sign.
- A written pass mark. For example: 90% of the test cases handled correctly, the rest handed to a person with a reason.
- An evaluation set. 50 to 200 real past cases with the right outcome for each, rerun after every change.
- Human approval on money and customer actions. Refunds, payments, journal entries and outbound emails wait for a person to approve.
- Logs of every action. What the agent read, what it decided and what it did, kept where your team can read them.
- Phased fixed prices. Each phase scoped, priced and signed on its own, so you can stop after any phase.
- Code and keys in your name. Repository, prompts, model API keys and cloud accounts belong to you from day one, which settles data ownership.
- A named owner after launch. One person on your side and one on the vendor's side, with a support option you can choose or decline.
- An exit clause and a kill date. A handover of everything built if you part ways, and a date when the pilot is judged against the pass mark.
Points 7 to 10 belong in the contract itself. Our AI agent development contract checklist covers the clauses. For the governance side, our AI agent governance policy guide and AI agent security risks and controls go deeper. If you have not yet checked whether your data and team are ready, start with the AI readiness assessment checklist.
What a written pass mark looks like
A pass mark is only useful if anyone can check it. Write it as three numbers: the share of test cases the agent must handle correctly, the share it may hand to a person, and the share of wrong actions you will accept (for anything touching money, zero, because a person approves first). Run it on the evaluation set before launch and again every time the prompt, the model or a connected system changes. NIST's AI Risk Management Framework, released 26 January 2023, and its Generative AI Profile (NIST-AI-600-1, 26 July 2024) give a free structure for this testing and governance work that US, UK and Australian firms can all use.
Your AI pilot already failed: should you kill it, fix it or restart?
AI pilot failure is common enough that S&P Global could average it: 46% of projects scrapped between proof of concept and broad adoption. A stalled pilot needs a decision. Answer these five questions honestly and the answer usually picks itself.
| Question | If yes | If no |
|---|---|---|
| Was the job clear and is it still worth doing? | Keep going to the next question | Kill it. No amount of engineering fixes a job nobody needs |
| Can the agent reach the live data it needs? | Fix: the data path works | Fix if a connection or API can be built; kill if the data does not exist |
| Is there a test set and a written pass mark? | Fix: you can measure the gap | Fix: build the evaluation set first, then judge the pilot against it |
| Did the intended users adopt it? | Fix the accuracy and widen it | Restart with the users in the room, starting from their real workflow |
| Is fixing cheaper than starting over? | Fix with the current code | Restart, and keep the evaluation set and lessons from the first attempt |
Two signs point straight to a restart with someone new: the original vendor holds your code or keys and will not hand them over, or nobody can explain why the pilot missed. Two signs point to stopping altogether: RAND's fifth cause (the task is too hard for current models) and Gartner's point that many use cases positioned as agentic do not need an agent at all.
Who fixes AI projects that never made it to production?
When we asked the assistants who rescues a failed AI pilot, few of the pages they cited dealt with rescue at all; tkxel.com's page on failed AI pilot recovery was one. The work itself follows a set order, whoever does it:
- Collect what exists: the repository, prompts, model and cloud accounts, test results and user feedback.
- Rerun the agent on a fresh sample of your real cases and record where it fails.
- Sort each failure into scope, data, testing or adoption, using the table above.
- Agree one narrow first phase with a written pass mark and a kill date.
- Fix or rebuild that phase, test it against the evaluation set, then widen the scope.
A rescue with FactoryJet starts the same way as a new build. FactoryJet quotes a fixed price in writing after a short scoping call, each phase is scoped, priced and signed on its own, and we show working software on your own data before you sign. Our AI consulting service covers the review, and our guide to taking an AI proof of concept to production covers scoping the next attempt.
Where do AI agents fail by department?
The same seven causes show up differently in each team. These are the patterns to check for in the four places mid-sized companies most often start.
| Department | Typical failure | What prevents it |
|---|---|---|
| Customer service | The agent answers from an out-of-date help center, or cannot hand a frustrated customer to a person with context | A live link to orders and tracking, a clear handoff rule, and test cases drawn from real tickets in Zendesk, Gorgias or your help desk |
| Finance and accounts payable | The agent posts entries or approves invoices with no person checking, and one mismatch erodes trust | Human approval before anything posts to Xero, NetSuite or SAP Business One, and logs of every match |
| Sales and order entry | Purchase orders arrive as PDFs, emails and EDI, and the agent was only tested on one format | An evaluation set covering every format customers send, and a person approving orders over a set value |
| Operations and procurement | Stock, prices and supplier data live in an ERP the agent cannot reach, so it guesses | A read connection to Odoo, NetSuite or ERPNext first, and write access only after the reads pass testing |
For worked examples by team, see our guides to AI agents for customer service, AI agents for finance and accounting teams and AI agents for procurement and supply chain.
Why do multi-agent LLM systems fail?
Some vendors pitch a team of agents that hand work to each other. A paper led by University of California, Berkeley researchers, Why Do Multi-Agent LLM Systems Fail? (Cemri and 12 others, arXiv, revised 26 October 2025), annotated more than 1,600 traces across 7 popular frameworks and sorted the failures into 14 modes in 3 groups: system design issues, misalignment between agents, and weak task verification. In plain words, agents get unclear roles, talk past each other, and nobody checks the final result. Start with one agent and one job unless the work splits into separate jobs. Our guide to multi-agent workflows with LangGraph and n8n covers when more than one is worth it.
What changes for UK and Australian buyers?
The failure research is mostly American and European. S&P Global sampled North America and Europe. RAND wrote for the US Department of Defense and others. We could not find a measured AI project failure rate for the UK or Australia alone. The University of Queensland's 80% comes from a literature review in a book, and Canon Business Services Australia's page ranked first in Google Australia for "why ai projects fail" on 18 October 2026.
What does change is the regulator. In the UK, the ICO's guidance on accountability and governance in AIsays the accountability principle makes you responsible for complying with data protection law, and for demonstrating it, in any AI system that processes personal data; it calls a data protection impact assessment (DPIA) an ideal way to show that under UK GDPR. In Australia, the OAIC's guidance on commercially available AI products (published 21 October 2024, updated 17 January 2025) says organisations should do due diligence, including how human oversight can be embedded and who will have access to personal information. Both line up with points 5, 6 and 8 of the checklist above. When the AI services and servers run in your own cloud account, you choose the region, including the UK or the EU. See our UK AI agents and Australian AI agents pages.
What else ranks for "why do AI projects fail"
These pages ranked on Google's first page or were cited by AI assistants for this topic on 18 October 2026 (DataForSEO and our own assistant runs). We list them so you can read them alongside this guide.
| Page | Where it showed up | What it covers |
|---|---|---|
| rand.org, The Root Causes of Failure for Artificial Intelligence Projects | Number 1 in Google US for "why do ai projects fail" | Five root causes from 65 interviews |
| business.canon.com.au, Why AI projects fail | Number 1 in Google Australia for "why ai projects fail" | Written for Australian businesses by Canon Business Services Australia |
| hbr.org, Why Agentic AI Projects Fail (Anushree Verma, 21 October 2025) | Number 1 in Google US for "why ai agent projects fail" | Written by the Gartner analyst quoted in the 40% release |
| gartner.com, Why 50% of GenAI Projects Fail | Top 4 in Google Australia | Five failure points for generative AI |
| ibm.com, the science experiment trap | Top 10 in Google US and Australia | Silos and scale, with IBM CEO Study figures |
| softermii.com, beam.ai, tkxel.com | Cited by AI assistants for agent failure and rescue questions | Agent failure patterns, partner checklists and pilot recovery |
What this guide adds: each headline figure checked against its primary source on the day, the origin of the Gartner 85% figure, a kill, fix or restart decision table, and the contract terms that stop the vendor-side causes.
How FactoryJet keeps agent projects out of the 42%
We design, build and support custom AI agents for companies of 10 to 50 staff, 50 to 500 and 500+, and connect them to the systems they work in: Odoo, NetSuite, SAP Business One, ERPNext and custom ERPs, Salesforce, HubSpot, Xero, EDI and WhatsApp. The terms below exist because of the causes above.
- We show working software on your own data before you sign.
- You own the code and the repository from day one.
- The AI and cloud accounts can run in your own name, at cost, with no markup.
- Defects are fixed inside the build price for a cover period agreed before work starts.
- Each phase is scoped, priced and signed on its own.
- Support is optional: each change is quoted and approved, with no monthly minimum.
See AI agent development and AI agent monitoring. If you are weighing the budget, our enterprise AI agent cost guide covers market build ranges, and Copilot Studio vs custom AI agents covers when a Microsoft tool is the smaller project.
More 2026 decision guides
- AI MVP development cost in 2026: what a first AI product costs to build, by scope.
- AI agency vs in-house engineer vs freelancer: which to pick for your company size.
- 15 questions to ask an AI development company: and the red flags in the answers.
- AI agent development contract checklist: 14 clauses on data, ownership and support.
- Copilot Studio vs a custom AI agent: pricing and when each wins.
- Enterprise AI agent cost in 2026: build ranges and what drives them.
Sources
Every page below was opened and read on 18 October 2026.
- S&P Global Market Intelligence, AI experiences rapid adoption, but with mixed outcomes: Highlights from VotE: AI & Machine Learning (30 May 2025)
- Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 (25 June 2025)
- Gartner, Why 50% of GenAI Projects Fail, Arun Chandrasekaran (26 January 2026)
- Gartner, Gartner Says Nearly Half of CIOs Are Planning to Deploy Artificial Intelligence (13 February 2018)
- MIT NANDA, The GenAI Divide: State of AI in Business 2025 (July 2025, PDF copy)
- RAND, The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed, RR-A2680-1 (13 August 2024)
- IBM, Why AI projects fail: The science experiment trap, Neil Dhar (23 October 2025)
- University of Queensland Business School, Question the hype: why 80% of AI projects fail
- Harvard Business Review, Why Agentic AI Projects Fail, Anushree Verma (21 October 2025; title, author and date only)
- Cemri and others, Why Do Multi-Agent LLM Systems Fail?, arXiv 2503.13657 (v3, 26 October 2025)
- OWASP Gen AI Security Project, LLM06:2025 Excessive Agency
- NIST, AI Risk Management Framework
- ICO, What are the accountability and governance implications of AI?
- OAIC, Guidance on privacy and the use of commercially available AI products (updated 17 January 2025)
- Our own measurement: Google results, People Also Ask and search volumes in the US, UK and Australia (DataForSEO), and 36 buyer questions asked across five AI assistants, 18 October 2026
New agent or stalled pilot
See a demo on your own data before you sign
Tell us the job, the systems behind it and, if there was a first attempt, what happened. We will show the agent working on your own data before you commit, and FactoryJet quotes a fixed price in writing after a short scoping call. Founder replies within 24 hours.
Frequently Asked Questions
Why do most AI agent projects fail?
What percentage of AI projects fail?
Is it true that 80% of AI projects fail?
Is it true that 95% of AI projects don't make money?
Where does the "Gartner 85% of AI projects fail" figure come from?
What did Gartner say about agentic AI projects being canceled?
What is a common reason for AI project failure?
Why does AI fail so much?
What is agent washing?
Do AI projects fail more often than normal IT projects?
Is it safer to buy an AI tool or have one built?
Why do AI pilots work in the demo and fail in production?
How do I know if my AI pilot has failed?
Can a failed AI pilot be rescued?
Who fixes AI projects that never made it to production?
Should we restart with a different vendor?
How long should an AI pilot run before we decide?
Why do multi-agent LLM systems fail?
How do I choose an AI vendor that won't fail?
What should the contract say to protect us from a failed AI project?
How do you test an AI agent so it does not fail after launch?
Do small and mid-sized companies fail at AI more than big ones?
Are AI failure rates different in the UK or Australia?
What is the AI project failure rate in 2026?
Can we see the agent work on our data before we commit?
Bhavesh Barot
Founder & CEO
Founder & CEO of FactoryJet, an ecommerce and AI services company that has served 500+ businesses since 2014. Writes about ecommerce builds, AI agents, and AI search.
Related reading
- Salesforce Agentforce Pricing vs a Custom AI Agent: What It Costs and When Each Wins (2026)Salesforce Agentforce pricing read on 18 October 2026: Flex Credits, per-action costs, add-ons, Max Edition and the $5 user license, with Salesforce's own worked examples, how to estimate your bill, Einstein vs Agentforce, and when a custom agent or a custom connection is the better fit for a mid-sized company.
- AI Agent Architecture for Manufacturing: Automated RFQ Quoting & ERP Sync (2026)Technical blueprint for manufacturing AI agents: PDF CAD drawing parsing, multi-tier BOM costing, and bi-directional NetSuite and SAP ERP synchronization.
- Build vs. Buy for AI Agents: 5 Builders Compared and an Honest Framework (2026)Should you buy an AI agent tool or build a custom one? Five builders compared, 2026 market cost ranges from named sources, and a plain framework for when each path makes sense.
