In 2025, 42% of companies abandoned most of their AI initiatives, up from 17% the year before. That is S&P Global Market Intelligence surveying more than 1,000 organizations across North America and Europe, and the same survey found the average company scrapped 46% of its AI proofs-of-concept before they ever reached production.
We build AI systems for clients, and we have also inherited the corpses: half-dead pilots handed to us after eighteen months and a burned budget. The failure causes are boring, repeatable, and visible before the first line of code, which is the good news. This post covers what the studies actually measured, the five causes we see up close, the situations where we tell clients not to build AI at all, and the pre-build checklist we run before taking a project. It is not an argument against AI; the surviving cohort is compounding while the abandoners write off budgets.
The real numbers, read carefully
Two studies dominated the 2025 headlines, and they measured different things.
S&P Global Market Intelligence measured abandonment: 42% of companies scrapped most of their AI initiatives in 2025, against 17% in 2024, with cost, data privacy, and security cited as the top obstacles. MIT's NANDA initiative measured impact: its State of AI in Business 2025 report, built on 52 executive interviews, 153 leader surveys, and 300 public deployments, found 95% of generative AI pilots produced no measurable P&L impact.
Read them carefully and neither says "AI does not work." They say organizations fail at AI in predictable ways. Two findings inside the MIT report matter more than the headline:
- Buying beats building, by a lot. Pilots run with specialized external vendors succeeded about 67% of the time; comparable internal builds succeeded at roughly a third of that rate.
- The money went to the wrong department. More than half of generative AI budgets went to sales and marketing tools, while the measurable ROI concentrated in unglamorous back-office automation: document processing, reconciliation, reducing outsourced work.
So the 2025 death toll is mostly a story of unowned projects chasing visible use cases with unready data. Each of those is fixable before you start.
The five causes of death
| Cause | What it looks like | The early tell |
|---|---|---|
| No owner | Pilot lives in an "innovation" team, nobody's targets depend on it | You cannot name who gets fired if it fails |
| Data not ready | Six months cleaning data that was assumed clean | Nobody has actually opened the source data |
| Wrong problem | Use case chosen because it demos well, not because it costs money | No baseline number for the current process |
| Unpriced inference | $400/month pilot becomes $15,000/month at production volume | No per-task cost model at 10x scale |
| No evaluation | "It seems pretty good" is the quality bar | No test set of real cases with known answers |
1. Nobody owned it
The strongest predictor we see. Surviving pilots have a named business owner whose annual targets move if the system works, with budget authority and the power to change the process around the tool. Dead pilots were "sponsored by leadership" in general and owned by no one in particular. If you cannot answer "whose number does this move?" in one sentence, the pilot is already dead; it just has not stopped moving yet.
2. The data was not ready
Every failed document-processing and knowledge-assistant pilot we have audited shared this: the team assumed the data was usable and discovered otherwise in month three. Duplicates, superseded policy versions, scanned PDFs that extract as noise, fields that mean different things in different systems. Data preparation is routinely 30% or more of an AI project's real cost, a breakdown we detail in our RAG implementation guide. Budget for it up front or pay for it mid-project with interest.
3. It solved the wrong problem
The MIT budget-allocation finding is the tell: money flowed to customer-facing showpieces while the returns sat in back-office work. A chatbot on the homepage demos well to the board. An agent that cuts invoice-processing time from 12 minutes to 90 seconds never gets demoed and pays for itself in a quarter. If there is no baseline measurement of the current process, the use case was chosen for optics.
4. Nobody priced inference at scale
A pilot serving 20 users costs a few hundred dollars a month, and nobody models what production volume does to that number. We have reviewed builds where per-task token costs, multiplied by real volume, exceeded the salary cost of the humans they replaced. Cost per task at 10x volume is a spreadsheet you build in week zero. The architecture choices that keep it sane, caching, model routing, and batching, are covered in our AI-powered SaaS architecture guide.
5. There was no evaluation, so there was no defense
When budget review comes and the sponsor asks "does it work?", the team without an evaluation set answers with anecdotes. Anecdotes lose to a CFO. The fix costs almost nothing: 50 to 200 real cases with known correct answers, scored on every change, tracked as a number. Pilots with an accuracy trend line survive budget season. Pilots with vibes do not.
When we tell clients not to build AI
Roughly a third of the AI briefs we receive get a "do not build this" answer, because the fastest way to join the 42% is to build something that should never have been built. Our disqualifiers:
- The process is deterministic. If the rules can be written as if-then statements, a workflow engine does it for a tenth of the price with zero hallucination risk.
- The volume is too low. A task a human does 40 times a month in 5 minutes each is not worth $60k of automation. The math has to clear payback inside 18 months.
- Errors are catastrophic and unreviewable. If a wrong output means legal exposure and no human checkpoint fits the workflow, current models do not belong there.
- The data does not exist. No history of the documents, decisions, or outcomes means nothing to ground the system in and nothing to evaluate against.
- An off-the-shelf tool already does it. If a $200-a-month product covers 90% of the need, we say so and lose the build.
The pre-build checklist
We run this before any AI engagement. Every item is answerable in under two weeks, and any "no" that stays a "no" kills the project.
- A named owner whose performance targets move if the system works.
- A baseline measurement of the current process: cost, time, error rate.
- One workflow, defined narrowly. Not "AI for operations," but "extract these 6 fields from these 3 invoice formats."
- An evaluation set of 50+ real cases with known correct answers, built before choosing any vendor or model.
- A data audit: someone has opened the actual source data and reported what state it is in.
- A per-task cost model at 10x pilot volume.
- A human fallback path for the outputs the system gets wrong, because some outputs will be wrong.
- Kill criteria in writing: the metric and date at which the pilot is stopped. Killing a pilot on schedule is a success, not a failure; zombie pilots are what push companies into the 42%.
What the survivors do differently
The surviving cohort is not smarter about models. It is narrower and more disciplined about deployment. The patterns that repeat: one deep workflow instead of ten shallow demos, back-office before front-office, partnering with teams who have shipped the pattern before rather than building from scratch (the MIT data puts external partnerships at roughly twice the success rate of internal builds), and an evaluation harness from day one. We keep a running list of the deployment patterns that actually produce returns in what works in AI agents for business in 2026, and the honest summary is that the winners look boring: documents get processed, tickets get resolved, reconciliation gets done, and the number moves.
That discipline, narrow scope, evaluation-first, priced inference, is the entire difference between the 42% and the rest. None of it requires frontier research. All of it requires saying no to the impressive demo.
Frequently asked questions
Why do most enterprise AI pilots fail?
Not because the models are weak. The documented causes are organizational: no accountable owner, data that was never audited, use cases chosen for optics instead of measurable cost, inference costs that were never modeled at production volume, and no evaluation set to prove the system works. All five are detectable before the build starts.
Is the 95% failure figure real?
It comes from MIT's NANDA State of AI in Business 2025 report and measures something specific: the share of generative AI pilots with no measurable P&L impact, based on interviews, surveys, and 300 public deployments. It does not mean the technology fails 95% of the time; the same report found vendor-partnered deployments succeeding about 67% of the time. The gap is execution, not capability.
Should we build AI in-house or buy from a vendor?
The MIT data is blunt: external partnerships succeeded at roughly twice the rate of internal builds. Build in-house only when the workflow is genuinely proprietary and you have the engineering depth to maintain evaluation, monitoring, and model updates for years. For most companies the right split is buying the commodity layers and custom-building only the workflow logic that differentiates them.
How long should an AI pilot run before we kill it?
Write the kill criteria before you start: a target metric and a review date, typically 60 to 90 days after the pilot reaches real users. If the metric is not trending toward target by then, stop, and treat the stop as the checklist working. The expensive failure mode is not the pilot that dies at day 90; it is the one that limps along for 18 months because nobody defined what failure looks like.
Key takeaways
- 42% of companies abandoned most AI initiatives in 2025, up from 17% in 2024, and the average firm scrapped 46% of its proofs-of-concept.
- The five causes are organizational, not technical: no owner, unready data, wrong problem, unpriced inference, no evaluation.
- Back-office automation produces the measurable returns; most budgets went to sales and marketing showpieces instead.
- Vendor partnerships succeed at roughly twice the rate of internal builds.
- Write kill criteria before the pilot starts. A pilot killed on schedule is cheap; a zombie pilot is how you join the 42%.
Sources
- CIO Dive: AI project failure rates are on the rise (S&P Global Market Intelligence survey)
- Forbes: MIT finds 95% of GenAI pilots fail
- Mind the Product: Key findings from MIT's 2025 AI report
What to do next
If you are planning an AI build for 2026, run the checklist above before you talk to any vendor, including us. Then, if you want a second opinion, send us the workflow you want to automate and our AI development team will send back a written assessment: the payback math, the data risks, and a build/buy/don't-build call. About a third of the time that answer is "don't build this," and we will tell you so before you spend anything.