Cost, honestly

AI projects have a characteristic financial shape: cheap to start, cheap to pilot, and considerably more expensive to run at scale than the pilot implied. Almost nobody is being dishonest. The costs simply arrive in categories that pilots do not exercise.

The four cost lines

Licences and usage. Per seat, per volume, or per unit of text processed. Predictable, and it grows exactly in proportion to your success — which is a strange thing to budget for and catches teams that modelled a flat cost.

Internal implementation time. The line that never appears in the business case. Kavita's eight-week pilot consumed 60 hours of her team's time, plus 15 of her own. At loaded cost that is a real number, and it was larger than the licence fee for the pilot period.

Ongoing review time. The permanent one. If verification adds three minutes to a task done 400 times a week, that is 20 hours a week, forever. It may still be an excellent trade against the time saved — but it is a cost, and reporting the saving without it is how a project gets approved on numbers that do not hold.

Maintenance and drift. Vendors update models. Behaviour changes. Your prompts and configurations need revisiting, your quality measurements need re-running, and someone owns that. Budget a day a month per significant deployment and treat a rising number as a signal.

Why pilots mislead about cost

Volume. A pilot on one team at 400 files a week is not a linear predictor of five teams at 3,000, because pricing tiers, rate limits, and support requirements all step rather than scale.

The easy cases went first. You piloted on the sub-process most likely to work. The next one is harder, and the marginal value falls while the marginal cost does not.

Integration was deferred. The pilot ran with people copying things between systems. Making it not require that is the expensive part, and it was outside the pilot's scope by design.

Enthusiasts need less support. The pilot group figured things out. The wider rollout generates training and support demand that the pilot never revealed.

A reasonable planning heuristic: the fully-loaded cost of running something at scale is typically several times the licence line, and the multiplier is mostly your own people's time.

What it is worth

Against the costs, be equally disciplined about the benefit, because this is where business cases get inflated.

Time saved is only money if it is redeployed. Twenty minutes saved per person per day either becomes more work done, or it does not. Both outcomes are legitimate — capacity relief in an overloaded team is genuinely valuable — but they are different claims and a CFO will distinguish them.

Quality improvements are real and hard to price. Fewer errors, faster response, more consistent output. Use whatever your organisation already uses to value these; do not invent a method.

Capacity is the honest benefit for most teams. Kavita's team was at ninety per cent for two years. The summarisation deployment did not reduce headcount and was never going to. It absorbed a 22% volume increase without adding people, which is the thing her CEO actually needed and is a cleaner claim than a synthetic cost saving.

The uncomfortable question

Someone will ask whether this reduces headcount. Have an answer before you are asked, and make it the true one.

For most operations teams in the current state of the technology, the honest answer is: it changes the mix of work rather than removing roles. The routine drafting shrinks; the checking, the exceptions, and the difficult conversations remain and become a larger share of everyone's day. Whether that is a better job depends on the person, and pretending the question is not being asked is the fastest way to lose the team's cooperation — which lesson 11 is about.

Where the answer genuinely is headcount, say so to your leadership honestly and plan the human side properly. What does not work is being vague to your team while being specific to your board, because those two audiences overlap more than managers expect.

Do this today: take your pilot's costs and model them at full scale, including review time at your real volume. If the number surprises you, better now than in the second budget cycle.