The last mile is not the model

Every few weeks someone sends me a demo where an agent reads an invoice, decides something, and writes the result back to an ERP. The demo always works. I spend my working hours on the write-back half of that sentence, and in nearly three years of joining enterprise systems the demo has never once been the part that worried me.

The 2026 numbers make the same point in a less personal way. Gartner found that 80% of enterprise applications shipped or updated in the first quarter of 2026 embed at least one AI agent, up from 33% in 2024. In the same stretch, fewer than 10% of the companies that tried agents had scaled one far enough to show measurable value. Those two figures describe a single gap, and I think most teams look for it in the wrong place.

Where the work actually is

When teams report what blocked them, 46% name integration with existing systems as their main problem, ahead of model quality, cost, and hallucination. For most of those teams, the thing that stops the project is getting an agent safe and reliable access to the systems that already run the business.

In my experience that surprises people, and the reason is that the demo hides it. In a demo the ERP is a mock: it answers in 40 milliseconds, it returns no lock timeouts, and it holds one customer record with clean data. Production is the opposite on all three counts, and an engineer who has joined two enterprise systems before will tell you that the schema was the easy half. The hard half is what happens on the fourth retry, and on the tens of thousands of records that land the same day.

For example, consider an agent posting a journal entry. The model decides correctly, the API call times out, and now the agent cannot tell whether the entry landed. It retries, and one invoice ends up posted twice inside a day that already carries tens of thousands of records. Among engineers who have joined two systems before, this is solved ground. The fix has a name, idempotency keys, but it is dull work that no demo contains. Someone has to go and build it for each of the actions the agent is allowed to take.

The production gate

But there is a second thing the demo hides, and it is the one I keep coming back to.

An agent that can write is an agent that can write to production. My own week here was a small version of the same problem. I fixed seven one-off scripts that could not load their environment variables, and the fix worked, but those seven had been failing to connect for months, and that failure was the only thing standing between a stray command and the production database. Repairing the bug removed the accident that had been protecting me. So the next change added a guard, and all seven scripts now refuse to touch production unless the engineer running them passes --apply.

At seven scripts this is a small case, and I think it scales in an unwelcome direction. Each capability you hand an agent needs the same three questions asked of it: what does this action touch, can it be undone, and who approved this particular run? A team that cannot answer those three for one script will not answer them for forty agent tools.

Why the projects get cancelled

Gartner expects more than 40% of agentic AI projects to be cancelled before the end of 2027, and I do not read that 40% as a verdict on the models. My guess is that most of them die in the gap between a working demo and an auditable production integration, because that gap is measured in engineering months and almost nobody budgets for it. I have not seen data separating the cancelled projects by cause, so treat the guess as a hypothesis.

The industry split supports it a little. Banking and insurance run agents in production at 47%, healthcare at 18%, and government at 14%. Those three are not the industries with the best models, but the industries that already had change control, audit trails and reconciliation in place before anyone used the word agent. McKinsey puts enterprise AI use at 88% of organisations in at least one business function while only 23% are scaling agentic AI anywhere, and the 65 points between those two figures is the same distance I am describing.

What I would do first

If a team asked me where to begin, I would not begin with the model, because the 46% who named integration were not describing a model problem.

In the first month, I would pick one action instead of one workflow, make it reversible, write down what it touches, and put a person in front of the version that cannot be undone. Then I would run that one action for 30 days and count two things separately: how often the integration failed, and how often the model was wrong. My expectation is that integration failures win by a wide margin, and the ratio between those two counts is the real project plan.

The model is the part a company can buy. For the 90% who have not scaled one, the last mile is the part they have to build, and it looks a great deal like the integration work that was already there before agents arrived.

Sources

• Gartner figures on agent embedding and project cancellation — via kore.ai
• Adoption, integration-blocker and scaling figures — digitalapplied.com, prefactor.tech
• Production-deployment rates by industry — paul-okhrem.com
• McKinsey State of AI scaling figures — via writer.com