Notes from the finance side of the AI debate.

The pitch is always framed as a rate: the system handles 70% of the volume, so you need 70% fewer people. That is not how any of it works, and finance is usually the only function in the room positioned to say so.

Four costs get left out, and they are not rounding errors. Implementation is spent in full before a single unit is deflected. Containment ramps over months, so the first two quarters deliver a fraction of the modelled benefit. Some share of automated output comes back for a human to redo, and rework costs more than doing it right the first time. And someone owns the thing permanently — exceptions, quality, the model itself.

Put those four back and the question changes shape entirely. It stops being “how much can we automate?” and becomes “what containment rate does this have to hit before it beats just hiring?” That hurdle is computable, it is usually higher than people expect, and almost nobody quotes it.

The uncomfortable part is what it implies about pilots. A pilot running below the hurdle is not early progress toward a good outcome — on the modelled numbers it is evidence for the other decision. That is a hard thing to say out loud eighteen months into a programme, which is exactly why it should be said in month one.

Efficiency claims almost always arrive as percentages, because percentages are flattering and denominators are not. A 40% reduction in handling time is impressive until you learn handling time is 6% of the cost base, at which point you are discussing a 2.4% saving that will be consumed by the oversight headcount.

This is not scepticism for its own sake. It is the cheapest diligence available, and it is the specific thing finance is for. Everyone else in the room is incentivised to discuss the numerator.

The follow-up question is better still: what would have happened anyway? A good share of measured improvement is volume mix, seasonality, or a process change that shipped in the same quarter. If the initiative cannot be separated from the counterfactual, the number is not evidence, whatever its sign.

Never take a single-scenario model to a board, and be wary of anyone who does. One number invites a debate about whether the number is right, which is unanswerable in advance and consumes the meeting.

Three cases move the conversation somewhere useful. If all three clear the bar, the decision is made and you can spend the time on execution risk instead. If none do, likewise, and faster. The interesting case is disagreement on the sign — because that tells you precisely which assumption the decision actually rests on, and that assumption is the only thing worth arguing about.

It also changes what a wrong forecast costs you. Present one number and being wrong is a credibility event. Present a range with the driver named and being wrong is the process working — you said which variable mattered and it moved.