Between pilot and production: where mid-market AI stalls

Every mid-market operations team now has an AI pilot somewhere in the building. A document-processing trial in finance. A routing experiment in logistics. A drafting assistant in customer service. Most of them work. Very few of them ship.

The standard explanation is a technology gap: the model was not quite good enough, the integration was not quite finished. This is almost never true. The model in the pilot is the same model that would run in production. The gap is not in the technology. It is in what the two environments are actually testing.

A pilot tests capability: can the model do the work, under supervision, on clean inputs. Production tests accountability: who is supervising, at what cost, with what authority, and what happens when the model is confidently wrong at 4pm on a Friday. Passing the first test says nothing about the second. That is why so many pilots succeed and so few deployments do.

What the pilot never tests

A pilot is a controlled environment by design. Clean inputs. A bounded workflow. A motivated team watching every output. The person running the trial is usually the person who championed it, and they catch its mistakes instinctively, before anyone else sees them.

None of that survives contact with production.

Production means the data arrives late and malformed. The workflow crosses three departments, two of which did not ask for the tool. The person watching the output has forty other responsibilities and no particular affection for the project. And the mistakes no longer surface in a weekly review. They surface in a client email, a misrouted shipment, a reconciliation that does not.

The missing layer is not more AI

The standard response to a stalled pilot is to ask for a better model, a bigger budget, or a broader mandate. The organisations actually closing the gap do something less glamorous: they treat deployment as an operating decision rather than a technology purchase.

They define authority limits before they define use cases. The agent acts freely below a threshold: routine reorders, standard replies, reversible scheduling. Above the threshold, it prepares a recommendation and stops. The threshold is set in the currency the business already thinks in: dollar value, customer commitment, supplier penalty, regulatory exposure.

They run the agent in draft mode first. It recommends, a person accepts or corrects, and the organisation accumulates a record of how often the recommendation was right before anyone grants it the ability to act. The move from recommendation to execution happens on evidence, not enthusiasm.

And they keep a human in the loop deliberately, not apologetically. This is the attended-AI model: a permanent division of labour in which the agent holds throughput and a human holds judgement, and the boundary between them is managed as a design decision rather than discovered as an incident.

Judgement is the scarce resource

AI reduces routine work. What it does not reduce is the need for senior human judgement on the exceptions: the ambiguous case, the novel input, the consequential decision that falls outside the rules. Production concentrates exactly this work. The routine is absorbed by the agent; what surfaces to people is precisely the part that requires experience.

That reframes the staffing problem. In the pilot, supervision was free: the champion did it in the margins of their day. In production, supervision is a real function, and it demands enough domain seniority to know when the plausible answer is the wrong answer. The senior operator who can supervise a supply-chain agent or a finance agent is the same person the mid-market company already cannot recruit for its own openings. The talent constraint that motivated the AI programme in the first place reappears inside the oversight layer.

This creates a new bottleneck: the expertise required to supervise the exceptions.

h.work is built on the attended-AI model. Each specialist agent is paired with a named, credentialed senior expert, verified through Humanity, who supervises across a portfolio of companies. One experienced practitioner can hold the judgement layer for many deployments at once, because the work of oversight is concentrated where judgement actually lives: the escalated exception, the novel case, the decision outside the rules. Routine output is monitored continuously, and consequential decisions pause for expert review before execution.

The pilot never needed this. Production cannot run without it.

Production is an operating decision

The practical difference between the firms that deploy and the firms still piloting is not ambition or budget. It is a set of decisions made before go-live: one bounded workflow instead of a broad platform, authority limits written down, recommendation mode held until the correction record justified execution, supervision assigned as a role with a name attached, and an audit trail kept — every action logged, every correction captured — so the boundary between agent and expert can move deliberately over time.

None of this requires a better model. All of it requires deciding, in advance, who attends.

The pilot proves the model can work. Production proves the organisation can trust it. The first proof is cheap. The second is the operating model — and it is the one that pays.