Commercials · 7 min read · Aug 12, 2026

Measure the process, not the pilot.

Ask why an AI project cannot prove its value and the answer is almost never that the value was absent. It is that nobody wrote down what things looked like before. This is a measurement design problem, and it is solvable — but only in advance.

The baseline is the whole game

The single most common failure is starting without a measured baseline. Once a system is live, the pre-system state becomes a matter of recollection, and recollection is generous in whichever direction the person answering prefers.

Four steps in order: measure the baseline before build, agree the pass mark in writing, run in parallel with the same inputs through both paths, then track the process metric rather than model accuracy. The baseline step is highlighted as unrecoverable if missed.
The first step is the one that cannot be done later.

Capture it before anything is deployed, and capture more than the headline: how long the process takes end to end, how many items go through it, the error and rework rate, how much of it is exception handling, and how the volume varies across a month. That last one prevents a comparison against an unrepresentative week.

A baseline captured after go-live is not a baseline. It is a reconstruction, and it will be argued with by whoever does not like the conclusion.

Time saved is not money saved

The standard calculation multiplies hours saved by an hourly rate and produces a large number that finance correctly refuses to believe. The reason it is wrong is straightforward: unless the hours convert into either fewer people or more output, nothing has been saved. It has been redistributed.

Distributed savings are real but they are not bankable. Twenty people saving fifteen minutes a day is a genuine improvement in working life and it is not five headcount. Claiming it as five headcount is how AI programmes lose credibility with the finance function permanently.

Bankable outcomes are capacity absorbed without new headcount, an avoided planned hire, cancelled external spend and reduced error cost. Not bankable are minutes saved across many people, lower frustration, faster ad-hoc answers and better morale.
Both columns are real. Only one of them survives a finance review.

What is bankable is specific:

  • Capacity absorbed. Volume grew and headcount did not — the clearest and most defensible form.
  • Cost avoided. A hire that was planned, budgeted and did not happen.
  • External spend removed. An outsourced service or licence cancelled.
  • Error cost reduced. Fewer credit notes, penalties, reworks — measurable if you were counting before.
  • Cycle time. Faster close, faster quote, faster response, where speed has a demonstrable commercial effect.

Count the whole cost, including the permanent part

The cost side is usually understated by omitting everything after launch. A defensible figure includes build, integration, running cost across inference, infrastructure and integration — and supervision.

Supervision is the line that gets treated as temporary and is not. Someone reviews the exception queue, samples auto-approved output, and re-tests after model changes for as long as the system runs. Business cases that model oversight as a first-year transition cost systematically overstate payback.

Measure the process, not the tool

"Did the AI work?" is unanswerable and invites anecdote. "Did invoice processing get faster, cheaper or more accurate?" is answerable, and it correctly captures the cases where the model performs well and the process does not improve — because the bottleneck was approval, or because exception handling now takes longer than the manual path it replaced.

That distinction matters because process-level measurement is the only kind that catches partial automation making things worse. A system that handles 70% of cases and hands over a badly-framed 30% can reduce total throughput while showing excellent model accuracy.

A measurement design that survives review

WhenWhat to do
Before buildMeasure the baseline across a full cycle, not a week
At scopingAgree the metric and the pass mark in writing
At go-liveRun in parallel where feasible — same inputs, both paths
OngoingTrack the process metric, not model accuracy
QuarterlyRe-measure; behaviour drifts after model upgrades

The second row is the one that decides whether the exercise concludes. Agreeing the pass mark before build removes the possibility of the target moving to fit the result, in either direction — and it is the same discipline that separates a pilot which concludes from one that runs indefinitely without a verdict.

Parallel running is underrated and cheaper than it sounds. Same inputs through both the old path and the new one for a bounded period gives you a direct comparison on identical work, which no amount of before-and-after reasoning can match. It also surfaces disagreements while there is still a manual path to fall back on.

Dealing with this in your own group?

We answer scoping questions before there's a contract in sight — including the ones about cost and data handling.

Questions

Short answers,
in full.

The questions this article gets asked most, answered so each one stands on its own.

Talk to us
Why can't we prove our AI project's ROI?

Almost always because no baseline was measured before deployment. Once the system is live the previous state becomes a matter of recollection, which is unreliable and disputed. Capture process duration, volume, error and rework rates, exception share and monthly variation before anything is built — afterwards it is a reconstruction, not a baseline.

Is time saved a valid measure of AI ROI?

Not on its own. Hours multiplied by an hourly rate produces a number finance will reject, because unless the time converts into fewer people or more output it has been redistributed rather than saved. Bankable outcomes are capacity absorbed without new headcount, an avoided planned hire, cancelled external spend, reduced error costs, and cycle-time gains with a demonstrable commercial effect.

What costs should be included in an AI business case?

Build, integration, running cost across inference, infrastructure and integration, and supervision. Supervision is the commonly omitted line and it is permanent, not transitional — someone reviews exceptions, samples auto-approved output and re-tests after model changes for as long as the system runs. Treating oversight as a first-year cost systematically overstates payback.

Should we measure model accuracy or process improvement?

Process improvement. Model accuracy can look excellent while the process gets worse — a system that handles 70% of cases and hands over a badly-framed remainder can reduce total throughput. Measuring at the process level is the only way to catch partial automation that shifts work rather than removing it.