The number that gets quoted about AI agents is the gap between how many organisations are using them and how few are running them in production. Camunda's 2026 State of Agentic Orchestration and Automation, a survey of 1,150 IT leaders published in January 2026, puts it at 71% of organisations using agents in some capacity against 11% of agentic use cases reaching production in the last year.
It is worth noting that this is one survey with one sample. A lot of the versions of this statistic in circulation pair an adoption figure from one study with a production figure from another, which makes the ratio meaningless. This one does not.
The statistic gets read as "agents do not work yet". That reading is wrong, and it is expensive, because it points every stalled project at the one lever least likely to move it.
The pilot and the production system are answering different questions
A pilot asks: can the model do the task?
Production asks: can this organisation be accountable for what the model did?
Those are not the same question, and passing the first tells you almost nothing about the second. Accountability means state that survives a restart, a handoff rule a human can check, an audit trail that still makes sense next quarter, and a named owner when it goes wrong at three in the morning. A pilot with ten friendly users, watched by the team that built it, tests none of that. It cannot. There is nothing in a pilot for those things to fail against.
So the 11% is not measuring how many agents were good enough. It is measuring how many teams did the second piece of work, and most did not, because most did not know it was a separate piece of work.
The independent evidence lines up with this. Gartner's June 2025 prediction that more than 40% of agentic AI projects will be cancelled by the end of 2027 names three causes: escalating costs, unclear business value, and inadequate risk controls. Not one of those is a model capability problem. MIT's Project NANDA reached the same place from a different direction in The GenAI Divide, published July 2025 on 52 executive interviews, 153 surveyed leaders and 300 public deployments: 95% of pilots delivered no measurable profit-and-loss impact, and the separator was workflow integration and learning from feedback rather than model quality.
Camunda's own governance figures say it a third time. In the same survey, 84% cited the business risk of using AI in day-to-day processes when IT does not have appropriate controls in place, 80% pointed to a lack of transparency about how AI is used inside business processes, and 66% cited compliance concerns.
Some of those pilots were never agents
One part of the gap has nothing to do with engineering at all. Gartner calls it agent washing: rebranding existing products, including AI assistants, robotic process automation and chatbots, as agents without substantial agentic capability. Gartner estimated that only around 130 of the thousands of vendors claiming agentic AI were genuine.
A pilot built on a rebranded chatbot cannot be engineered into an agent, because the thing being piloted was never the thing on the slide. Some portion of the 89% stalled where it did because there was no path from what was bought to what was promised. Worth ruling out before spending another quarter on integration work.
What a year in production actually surfaced
We run an agent ourselves. Kavya is a WhatsApp assistant that qualifies our own inbound leads in nine languages before a human joins the conversation, and we have written up its architecture and what broke in full. It is the reason we can answer this question with something other than survey data.
The useful finding is what did not go wrong. Across a year, the two things that actually broke were:
Session state that did not survive WhatsApp's own delivery behaviour. A message arriving out of order, or a prospect resuming a conversation days later, could reset the conversation's language back to the default. Someone who opened in French got answered in English, fluently and politely and wrongly.
A disclosure that could fire after the greeting rather than inside it. The AI-system disclosure lived in a separate step, so in rare cases a templated welcome went out first. That is precisely the sequencing gap that Article 50 of the EU AI Act exists to close.
Neither is a model capability problem. One is state management and one is compliance sequencing, and a pilot would have surfaced neither, because a pilot runs in one language for a few days with everyone watching.
That is the shape of the thing. The model is rarely the binding constraint, and "try a better model" is the cheapest available response to a stalled pilot, which is exactly why it is the one teams reach for first.
The four questions to answer during the pilot, not after it
Each of these is cheap to design in and expensive to retrofit once real users have created real data you are responsible for.
1. What is the handoff condition, and can a human check it?
A handoff rule has to be a state transition, not a judgment call left to the model's discretion. Kavya moves a conversation to a human on three checkable conditions: the qualification questions are cleared, the prospect asks something outside the closed question set twice, or the prospect asks for a person. The failure mode of an uncheckable rule is an agent that either never escalates, which is a bad experience for a real customer, or escalates everything, which is an agent doing nothing. Both look fine in a pilot.
2. What is the agent allowed to not know?
"I do not know" has to be a designed, evaluated output rather than the absence of one. The failure mode being designed against is a model that would rather produce a plausible wrong answer than admit uncertainty, and a pilot audience does not push on that because a pilot audience is being nice. Anything touching capability, pricing or timeline should route to a human rather than get improvised.
3. What state has to survive a restart?
Whatever a conversation depends on being true across turns needs one owner and one source of truth, read everywhere and re-inferred nowhere. For us that was language, made sticky and explicit and durable across sessions rather than guessed independently by each component. Pilots do not restart, so pilots do not find this.
4. Who can answer what the agent did, three months from now?
This is the one that maps directly to the 80% in the Camunda data who cited a lack of transparency about how AI is used in their processes. If nobody can reconstruct why the agent did what it did, the project cannot pass a risk review even when the agent is working perfectly. That is a production blocker created entirely in the design phase.
This is not the same problem as a prototype that needs finishing
Worth separating, because the two get conflated. Our post on taking an AI-built prototype to production is about a codebase: something a tool like Lovable or v0 generated, where the gaps are auth, data modelling and observability in the artifact itself.
This post is about a programme: an agent that works, inside an organisation that cannot yet be accountable for it. Different object, different reader, different fix. A team can be stuck on both at once, but the remedies do not overlap.
Budget the production step as its own line
The last practical point. The pilot-to-production step is not a rounding error on the pilot, and treating it as one is how a project ends up cancelled for "escalating costs" when the real problem was that the second phase was never funded. Our stage-by-stage breakdown of what agents cost prices that step separately for exactly this reason, and shows why the jump is as large as it is.
A pilot that cannot be funded through to production is a sunk cost with a demo attached. Decide which one you are approving before you approve it.
Tell us where yours stalled
If you have an agent that works in a pilot and will not move, the useful next step is usually identifying which of the four questions above is the one actually blocking it, because it is rarely all four and it is rarely the model. Describe it in a paragraph and we will tell you what we think, NDA-first if you would rather not describe it in the open. See how we structure delivery, or get in touch.
Common questions
Because a pilot and a production system answer different questions. A pilot asks whether the model can do the task. Production asks whether the organisation can be accountable for what the model did, which covers state that survives a restart, a handoff rule somebody can check, an audit trail, and a defined owner. Gartner attributes the coming wave of cancellations to escalating costs, unclear business value and inadequate risk controls, and none of those three is a model capability problem.
Camunda's 2026 survey of 1,150 IT leaders found 71% of organisations using AI agents in some capacity while 11% of agentic use cases reached production in the last year. Gartner separately predicts more than 40% of agentic AI projects will be cancelled by the end of 2027. MIT's Project NANDA found 95% of generative AI pilots delivered no measurable profit-and-loss impact.
Usually not, and this is the most expensive misdiagnosis available. Teams respond to a stalled pilot by changing model, which is the cheapest lever and rarely the binding one. After a year running our own agent in production, the two things that actually broke were session state that did not survive out-of-order message delivery, and a disclosure that could fire after the greeting rather than inside it. Neither is a capability problem. Both are plumbing and accountability.
Gartner's term for rebranding existing products, such as AI assistants, robotic process automation and chatbots, as agents without substantial agentic capability. Gartner estimated only around 130 of the thousands of vendors claiming agentic AI were genuine. It matters to the production gap because some share of stalled pilots were never agent projects, so no amount of engineering would have moved them.
Four things. What the handoff condition is and whether it is checkable rather than a judgment call. What the agent is allowed to say it does not know. What state has to survive a restart. And who can answer what the agent did three months from now. Each one is cheap to design in and expensive to retrofit once real users have created real data.