Most AI agents never reach production
Directional survey figures put agent shipping near 80 percent, production at 31, and org-wide scale at 23. The cited blockers are integration and data quality rather than model capability.
Roughly 80% of organisations have shipped an AI agent. About 31% got one into production. Around 23% scaled one across the organisation.
These figures come through secondary reporting rather than from the primary surveys, and different sources define shipped and production differently. Treat them as directional. The shape of the funnel is the useful part, not the specific percentages.
What the drop is made of
The interesting thing is what does not appear in the blocker lists: model quality.
The cited obstacles are integration, data quality, and change management. All three are the same category of problem. They are what happens when a working demo meets an organisation that already has systems, definitions, and people with habits.
Integration means the agent needs to read from and write to systems that were not designed for it, owned by teams with their own roadmaps. The demo called one API. Production calls six, three of which have undocumented behaviour and one of which is a mainframe.
Data quality means the agent is now reading the warehouse rather than the curated set it was tested against. Every ambiguity in a column definition becomes a behaviour the agent exhibits.
Change management means the analysts whose work the agent touches have to trust it. That is earned by the agent being right in the cases those analysts already know the answer to, which is a much harder bar than a demo.
Why this matches what I see
A pilot proves the model can do the task. Production asks a different question: can this run unattended against real inputs, owned by someone, with a path for when it is wrong.
Nothing in the first question prepares you for the second, and the gap between them is not closed by a better model. It is closed by the unglamorous work of integration and data quality, which is why the funnel narrows where it does.
If you are planning agent work, the useful read of these numbers is a budgeting one. The build is the cheap part now. Assume the integration and trust work costs more than the agent did, and plan for it from the start rather than discovering it at the pilot review.
Related reading
- AI engineeringNot every agent in your pipeline should be an LLM callMost agent pipelines route validation, formatting, retries and the final write through the model. Those are the steps that should never vary, and handing them to a model is what makes an agent flaky and expensive.
- Data platformBad data does not slow you down. It scales your mistakes.Teams invest in models and underinvest in the data feeding them, then wonder why outputs feel off. AI did not create the problem. It removed the human who used to catch it.
- AI engineeringAI made execution cheap. Judgement did not get cheaper.Building an autonomous agent took days instead of months. Deciding which tasks deserved automation took far longer, and that gap is not closing.