Three reports published this year by AI platform vendors reach the same conclusion from different directions. The technology works. What stalls is the work around it.
What the research says
Camunda's report, The AI Process Gap, published in September 2026, surveyed 1,000 process decision-makers and 5,000 employees at organisations with more than 1,000 staff in the US, the UK, Germany and France. The headline numbers:
- 72% of decision-makers say process-related challenges have caused AI initiatives to fail, at an average cost of $1.55 million per organisation.
- 78% say their organisation needs to fundamentally redesign its processes to stay competitive, rather than layer AI on top. Yet 79% admit that bolting AI onto existing processes creates less internal resistance than redesigning them.
- 37% have allocated 10% or less of their AI spend to redesigning processes and managing the change that supports it.
The Dataiku and Neptune reports describe the same pattern in their own terms: AI deployed as a separate layer rather than inside the systems where the work happens, data that the workflow cannot reach, and governance added late.
Why bolting AI on feels safer, and costs more
Adding AI to one step of an existing process feels low-risk. Nobody's job changes, no approval chain moves, and the pilot can be switched off quietly. The trouble is that a process designed for people is full of steps that exist only because people had to move information by hand: rekeying, checking, forwarding, chasing. Put AI inside one of those steps and it waits at the same handoffs a person did.
The result is a pilot that works in a demo and saves little in daily use. It is easy to approve and hard to defend twelve months later.
The people doing the work see a different result
The Camunda survey asked leaders and employees the same question. 89% of leaders said AI is making their teams more productive. 64% of employees agreed.
The gap has a cause. 69% of employees said they were not fully consulted about how AI would be used in their day-to-day work, and 44% said they had manually overridden AI outputs because the underlying process was not set up correctly. Those overrides are invisible to leadership unless someone counts them.
What the pilots that reach daily use have in common
Across the three reports, and in our own work, the AI projects that survive share five traits:
- One workflow, with a named owner. Someone who would lose something real if it were switched off.
- A baseline agreed before anything is built. One number, its calculation method and its data source.
- The workflow mapped with the people who run it. The documented process is rarely the real one.
- Control points written down in advance. What AI does alone, what it drafts for a person to approve, and what it must never do.
- Compliance in week one, and every action logged. The number is reported monthly, with overrides and exceptions reviewed alongside it.
None of these is a technology choice. All of them are design choices made before the first line of code.
Where to start
Pick one workflow where the volume is high, the rules are clear and a mistake is visible. Finance operations, customer service and regulated client work are good candidates, and each has a short guide below with three worked use cases.
If you want a second opinion on your last pilot, the self-check on our homepage takes two minutes.