Most AI agent pilots that die do not die because the model was weak. They die at the next budget review, because nobody wrote down what "working" meant, or the agent could not reach clean live data, or it sent off-brand output to a customer with no human check and no record of what happened. Those failure modes are predictable. Each one maps to a single readiness control you can put in place before you deploy. Get the controls in, and the failure stops being an accident and becomes a decision you make on evidence.
A note on the number first, because you will see it repeated. A widely-circulated figure claims about 29% of agent projects are abandoned within 90 days, split neatly across three causes. We could not trace it to any survey with a disclosed method or sample size. It comes from a single marketing-vendor statistics roundup, quoted second-hand, with no underlying study named. So we are not going to treat it as fact, and we do not need to. The named analysts and academics who publish their method all tell the same story, and the mechanism carries the argument on its own.
What the credible evidence actually says
Two sources, both published on or before this article's date, both with an attributed method, describe the same pattern.
Gartner predicts that over 40% of agentic AI projects will be cancelled by the end of 2027, in a press release dated 25 June 2025. Note that word: predicts. This is a forecast, not a count of history. Gartner names the drivers plainly, and they are the point: escalating costs, unclear business value, and inadequate risk controls. Gartner also flags "agent washing," where ordinary automation and chatbots get rebranded as agents on a slide. None of those drivers is a model-quality problem.
The wider backdrop is worse. MIT's NANDA initiative, in "The GenAI Divide: State of AI in Business 2025" (August 2025), found that 95% of enterprise generative-AI pilots stall and never reach scaled use. Its method is disclosed: more than 300 publicly documented AI initiatives plus interviews with 52 organisations, gathered between January and June 2025. That figure is about generative-AI pilots broadly, not agents specifically, so treat it as the pond the agent projects are swimming in. The cause MIT points to is a friction and learning gap, not a capability gap. In plain terms, the tools could not reach the right data or fit the real workflow.
The pattern is consistent across independent, method-disclosed sources: most agentic pilots do not reach production, and a large share get cancelled outright.
You do not need a precise abandonment percentage after reading those three. A high cancellation forecast from the analysts, a near-total pilot-stall rate from the academics, and a single-digit production rate all point the same way, and they point at readiness, not at the model.
The failure modes are predictable, so the fixes can be pre-committed
The useful move is to name the failure modes and then bind each to a control you can check off before launch. Here is the mapping we work from.
| Failure mode | Weight in the failures | The readiness control that prevents it |
|---|---|---|
| Unclear success criteria | Largest, per Gartner's "unclear business value" driver | Define one measurable success metric plus a numeric kill-or-keep threshold, written before day one |
| Poor data and tool access | Major, per MIT's friction and disconnected-data finding | One source of truth, typed API access, and a live server-rendered record with state drift under a minute |
| Brand-voice drift on customer-facing output | Recurring compliance-shaped failure | A confirm step before any outbound customer action, plus bounded permissions |
| No accountability trail | Underlies audit and dispute failures | An immutable audit trail: agent ID, prompt ID, timestamp, and reviewer, retained at least 90 days |
The weight column is qualitative on purpose. We are ranking the modes by how the credible sources emphasise their drivers, not importing an invented split.
The unclear-success-criteria trap, and how to close it
The single most common way a pilot dies is that no one ever defined "done." The demo impresses a room, the project runs for a quarter, and at the budget review nobody can say whether it worked. So it gets cut. Gartner names this as unclear business value, and it is the driver most within your control.
The fix is unglamorous. Before day one, write one measurable success metric and a numeric threshold for keeping or killing the project. "Resolves 60% of tier-1 tickets without escalation, or we stop" is a real gate. "Improves customer experience" is not. Put the review date on the calendar at launch, not after the money runs out. That turns the roughly-90-day death from an accident into a governed decision, and it lets you defend the spend or cut it on evidence rather than a feeling in the room.
The data-reach trap, and the confirm step
The second killer is that the agent cannot reach clean, current data through one gateway. It quotes a stale price, invents a stock level, or cites a policy that changed last month. MIT's finding is that this friction, not model weakness, is what stalls pilots. The fix is architectural, not a bigger model: a single source of truth, typed API access so the agent asks for a named field rather than guessing, and a server-rendered record that reflects a database write in under a minute. You can test this before launch by handing the agent an out-of-scope token and confirming the request is denied.
The third killer is output. An agent that can send text or take an action for a customer with no human check will, sooner or later, send something off-brand or off-policy, and there will be no record of who approved it. Two controls answer this: a confirm step on any liability-bearing outbound action, and bounded permissions so the agent's token reaches only what its job needs. The confirm step only works if the human reads it, which is the failure mode hiding inside the failure mode. So instrument the human-override rate. Over 30 days, a 0% modify-or-reject rate is not a well-behaved agent, it is evidence the confirm step has decayed into a rubber stamp.
Where Origin Pi stands
The market's named analysts and academics independently say what we have said all along. Pilots do not die from a weak model. They die from no defined success metric, no clean live-data reach, and unbounded output to customers with no confirm step and no audit trail. That is the whole case, and it holds without a precise abandonment number.
Readiness is the difference between a pilot that dies quietly and one that ships. Our posture is a short, executable pre-deploy checklist, and every item is pass or fail before launch:
- A measurable success metric and a numeric kill-or-keep threshold, written before day one.
- Live data reach through one gateway, with state-drift latency under a minute.
- Bounded permissions, verified with an out-of-scope token that must be denied.
- A confirm step on liability-bearing and customer-facing actions, with the human-override rate tracked.
- An immutable audit trail retained at least 90 days.
- A 90-day kill-or-keep review already on the calendar at launch.
The last item is the quiet one. A round 90-day figure is partly an artefact of the quarterly budget cycle, not a property of the technology. Schedule the review before you launch and the deadline stops ambushing you. You decide on the metric you set, with the trail to back it. That is what agent readiness buys, and it is the foundation the agentic business layer and sound AI governance are built on.
