First 90 days of an AI Operator retainer: what actually ships.

The AI Operator engagement is the smallest of the three tiers KJ Capital offers and the one most firms understand least. It runs from £8k a month, it is not a build engagement, and it is not consulting. It is the running of AI systems inside a firm that has bought or built systems and does not yet have the in-house capability to operate them at the quality a regulator or a customer expects.

This note is the honest first-90-days log from three separate Operator engagements over the last year, aggregated and anonymised. It is what actually shipped, week by week, at the £8k/month price point. It is deliberately unvarnished — the goal is to describe what the retainer buys, not to sell it.

Week 1 — the situation report

Every Operator engagement begins with a situation report. Five days on site or over video, mapping every AI system in the firm, its ownership, its data flows, its compliance posture, its incident history, its cost, and its business owner. Most firms discover during this week that they have between two and four AI systems in production that leadership had lost track of. Every one of those is either an opportunity or a liability.

By end of week one, we have a shared document — the AI operating map — that names every system, every owner, every metric, and every gap. This is the artefact leadership refers to for the rest of the engagement. Nothing else ships in week one, deliberately.

Weeks 2–3 — the safety pass

The first thing that ships is a safety pass on every system on the operating map. Failure-mode analysis, guardrail audit, evals coverage check, incident-log review, compliance-envelope inspection. Anything below the acceptable safety line gets a mitigation plan and a shipping date.

The output of the safety pass is a short document of P0 fixes — the things that must ship before anything else. Usually two to five items. Common examples: an evals dashboard that stopped updating four months ago; a policy rule that was disabled during an incident and never re-enabled; a retrieval source that has drifted out of version; a monitoring alert that has been muted since June.

By end of week three, every P0 has either shipped or has a named owner and a shipping date. This alone is often worth the first quarter of the retainer.

Weeks 4–6 — the metric layer

Once safety is stable, the next thing to ship is the metric layer. Every system on the operating map gets a business-facing metric — conversion, retention, time-to-decision, cost per interaction, whatever the business owner most cares about — connected to the engineering metrics the system already produces.

By week six, leadership has a single dashboard showing every AI system, its business metric, its safety metric, and its trend. Most firms have never had this. It is the artefact that turns AI from a set of experiments into a set of assets. Every subsequent decision uses this dashboard as its evidence base.

Weeks 7–10 — the first improvement

With safety and measurement in place, the first structural improvement ships. Which one depends on the operating map and the metric dashboard. In one engagement it was upgrading a foundation model that had been unchanged for eight months; in another it was rebuilding a retrieval index that had accumulated four generations of superseded content; in a third it was introducing human-in-the-loop routing on the outputs with the lowest eval scores.

The pattern is that the first improvement is data-selected, not opinion-selected. The measurement layer surfaces the highest-return change, and the Operator ships it. In every case, the business metric moves within the ninety-day window.

Weeks 11–13 — handover discipline

The last month of the first quarter is not about shipping new work. It is about ensuring the firm’s existing team can run what has been put in place. Runbooks for the operating map. Runbooks for the metric dashboard. Runbooks for the top-five most likely incidents in the operating environment. A named on-call handover if the firm has an in-house engineering team.

The Operator engagement is designed to be either continued indefinitely or handed over cleanly. Both need this discipline in month three. Firms that skip it end up dependent on the Operator, which is a legitimate but expensive posture. Firms that invest in this month build the in-house muscle that makes the Operator optional over time — and those firms usually keep the Operator on anyway, for the strategic layer rather than the operational one.

What the £8k/month buys, honestly

By end of day 90, the aggregate output across the three engagements we drew from looked like this.

• A live operating map of every AI system in the firm, updated weekly.
• A safety pass that closed two to five P0 issues per firm.
• A metric dashboard connecting engineering and business metrics for every system.
• One material improvement that moved a business metric in the measurement window.
• A set of runbooks the in-house team can operate without the Operator present.
• A named Operator on retainer for the following quarter, with a defined scope of continued work.

What it does not buy

The Operator retainer does not build new AI systems from scratch. That is what the £75k+ AI Build engagement is for. If the diagnostic shows the firm needs a system it does not have, we do not build it inside the Operator retainer — we scope a Build alongside it.

The Operator retainer also does not replace a Chief Data Officer or a Head of AI. It is a fractional engineering leadership function, not an executive one. Some firms use it as a bridge until they can hire; others keep it in place indefinitely because the marginal cost of hiring internally is higher than the retainer.