Measurement & governance

If nobody measured the before, there is no ROI to report

The most common question in an executive meeting about AI is what it returned. Most organizations cannot answer it — not because there was no value, but because nobody measured the process before it changed. That is a week of work at the start of a rollout and close to impossible to reconstruct at the end.

Baseline before the rolloutTwo metrics, not twelveFull cost, not just licensesOne page leadership can read

Why the question goes unanswered

Four patterns behind every unprovable AI result

There is no starting point

Nobody recorded how long the process took before, or at what quality. There is nothing to compare the new number against, and it cannot be reconstructed afterwards.

Usage is measured, not value

Seat counts, message volume, and weekly active users all go up. None of them say what got faster, cheaper, or better — which is the only thing the question was about.

Savings nobody believes

A round hours-saved figure assembled after the fact and multiplied by an average salary. No finance team signs off on that, and asking them to is how AI budgets get frozen.

Only the license is counted

The subscription is in the number. Implementation time, training, prompt and workflow maintenance, and the internal hours spent making it work are not.

What we do

Numbers you can defend

We would rather hand over one metric leadership trusts than a dashboard nobody believes.

Baseline the process first

Before anything is automated, we measure the process as it runs today: how long it takes, how often it comes back for rework, and what it costs to run. This is the step almost everyone skips, and without it there is no ROI discussion later — only estimates.

Two metrics, not twelve

One time or cost metric and one quality or throughput metric, both tied directly to the process being changed, and both collectable from systems you already have. A dashboard with forty tiles gets built once and trusted by nobody.

Total cost, not just the subscription

Licenses, implementation, training, and ongoing maintenance, including your own team’s hours. An honest cost line is what makes the benefit line credible — a flattering number invites the wrong kind of scrutiny.

Reporting that survives the meeting

One page: what was measured, against what baseline, over what period, and what has not been proven yet. We include what did not work, because that is what makes the rest of the page believable.

What you end up with

Four artifacts, not a measurement program

Baseline record

The current-state numbers for every process in scope, with how each was collected, so the comparison later is defensible rather than remembered.

Metric definitions

Written definitions of the two metrics per use case: what counts, what does not, where the data comes from, and who owns the number.

Cost model

A simple model covering licenses, build, enablement, and maintenance, so cost per process is a number rather than an argument.

60-day report

A one-page result against the baseline with a clear recommendation: scale it, change it, or stop. The same report we run on AI training and rollout engagements.

This is how measurement ran on our AI training and engineering enablement engagement: licenses were scattered across Copilot, ChatGPT, and Claude with no usage data, so a baseline plus a 60-day metrics report was what made the result legible to the owners.

When this comes up

Almost always right after the first rollout

The question lands at the moment someone asks for the next tranche of budget. If any of these sound familiar, measurement is the gap:

  • Leadership asked what AI returned and there is no backed answer
  • You want to expand the rollout and need a budget justification
  • There is plenty of AI activity and not one agreed metric
  • A previous pilot ended without a clear conclusion either way
  • A board or PE owner wants adoption numbers, not anecdotes
  • Licenses are spread across ChatGPT, Claude, and Copilot with no usage picture

FAQ

AI ROI questions, answered

Can we just measure time saved?

It is a reasonable first metric and not a sufficient one, because saved time only becomes value if it goes somewhere — more output, shorter cycle times, or work that stops being outsourced. We pair the time metric with a quality or throughput metric: how much came out, at what standard, and how often it came back for rework.

What if the measurement shows it did not return anything?

Then you have saved the budget for the next phase, which is a good outcome. Not every process is worth automating, and honest measurement is what lets you stop early instead of funding a second year of something that never worked.

How much work is this on our side?

Less than people expect, if it happens at the start. Baselining a single process is usually a short questionnaire, a handful of reports you already produce, and a few hours of observation — not a measurement project. Reconstructing a baseline after the rollout is the expensive version, which is the argument for doing it first.

Can you measure a rollout that already happened?

Partly. Where the pre-AI state survives in your systems — ticket timestamps, close dates, invoice volumes, rework logs — we can rebuild a usable baseline from that history. Where it does not, we say so, and baseline the next phase properly rather than back-filling a number that will not hold up.

Who needs to be involved?

The person who owns the process, whoever owns the data it runs on, and someone from finance for the cost side. Keeping finance in from the start is what turns the final report from a vendor claim into an internal number.

How does this fit with the rest of the work?

It is built into our training and rollout engagements rather than sold as a separate analytics project: a usage baseline, license consolidation, and a 60-day metrics report alongside the build. It also works as a standalone starting point when AI is already live and nobody can say what it did.

Where this leads

The pages that continue from here

What would you need to prove, and to whom?

A short call, no deck. We will map what you already have, which processes are worth baselining first, and say honestly whether there is a number here worth measuring.