Put the AI inside the record, not beside it

A common pattern in applied AI goes like this. A team has an operational system — tickets, inspections, orders, documents. Someone adds an AI model that analyses the data and produces insights. The insights go into a separate dashboard. Six months later, nobody looks at the dashboard.

The model was not wrong. It was in the wrong place.

The system of record is where work happens

Every operational domain has a system of record: the place where the authoritative version of the work lives, where decisions are made and where someone is accountable for them. In inspection and document control, that is the inspection record and the controlled document. In support, it is the ticket. In commerce, it is the order.

People do their jobs inside that system. Anything outside it requires them to remember to go and look, interpret what they see, and copy the result back in by hand. That is three chances for it not to happen.

Beside versus inside

AI beside the record produces an alert, a report or a score in a separate place. Acting on it requires a human to bridge the gap.

AI inside the record writes its output into the same record people already work in — as a proposed finding, a flagged field, a suggested classification, a draft — with clear attribution that it came from a model, and a way for a person to accept, correct or reject it.

The second is harder to build. It needs a data model that can represent machine-generated suggestions, a review workflow, and an audit trail. It is also the only version that reliably gets used.

Why the audit trail matters

In domains where the record is evidence — inspections, compliance, anything that might be disputed later — it matters enormously who, or what, said what.

A good design keeps these distinct:

• What the model proposed, with the model version and the input it saw.
• What a person decided, and when.
• What changed as a result.

This protects everyone. When the model is right, it gets credit. When it is wrong, the record shows that a person reviewed it — or that nobody did, which is its own useful finding.

Computer vision is no exception

Monitoring with computer vision is especially prone to the "beside" pattern: a video feed, a detection model, a wall of alerts. The alerts only become useful when a detection creates or updates an item in the operational record, attached to the right location, asset and time, where the people responsible for it already look.

The takeaway

Before adding a model to an operational system, ask one question: where does its output land? If the answer is anywhere other than the record people already work in, the integration is not finished. The model is the easy part; getting its output into the flow of accountable work is the product.