How to Measure the ROI of an AI Consultation

Why most engagements never prove anything

Fewer than half of organisations have a clear methodology for measuring AI return. The consequence is measurable: organisations without a defined ROI framework are around 2.5 times more likely to cancel their AI programmes. Not because the programmes performed worse — because nobody could demonstrate that they performed at all.

When the budget conversation comes and the answer to "what did we get" is a description rather than a number, the programme loses to whatever else is competing for that money. The measurement is not bureaucracy. It is what keeps the work alive.

The baseline is the whole thing

Everything else in this post is arithmetic. The part that actually determines whether you can prove anything is a measurement taken before any work begins, while the process is still bad.

Once automation is running, the old number is unrecoverable. Nobody remembers how long invoice matching used to take, and estimates made afterwards are systematically flattering. A baseline reconstructed from memory is not evidence and will not survive a finance review.

• Time: hours per week spent on the process, counted rather than estimated.
• Volume: how many times it runs in a month.
• Error rate: how often it goes wrong, and what fixing it costs.
• Cycle time: how long from trigger to completion, including waiting.
• Cost per unit: the loaded labour cost of doing it once.

Baselines are taken during the diagnostic stage of an AI consultation — which is the only point at which they can still be captured honestly.

The formula

The standard shape is straightforward. Over a defined window — three years is common for consulting engagements, though a year is more honest for a first automation:

ROI% = ((total value − total cost) ÷ total cost) × 100

The discipline is in what goes into each side. Cost is not just the invoice: it includes the tooling subscriptions, the internal hours spent in discovery and testing, and any infrastructure. Value is where most calculations quietly inflate themselves, which the next section is about.

A worked example

Take a process consuming 20 hours a week at a loaded cost of $40 an hour. Loaded means salary plus employer taxes, benefits, and overhead — not the hourly rate, which understates it by a third or more.

20 × $40 × 52 ≈ $41,600 a year in recovered capacity. Against a $12,000 engagement plus $3,000 of annual tooling, that is a first-year return of roughly 177%, and it improves in year two when the build cost does not recur.

The honest caveat: recovered capacity is only money if you do something with it. If those 20 hours belong to someone who now does higher-value work, or if it lets you take on more volume without hiring, it is real. If the person simply has a quieter week, you have bought comfort rather than return. Say which one it is.

Count three things separately, not one blended number

The most common error in AI ROI maths is collapsing everything into a single payback figure. Keep three inputs apart, because they have very different reliability.

• Labour reduction — the most defensible, because it reconciles against payroll and timesheets.
• Revenue lift — real but harder to attribute; more leads contacted quickly should show up in conversion rate, not just in pipeline.
• Infrastructure cost — a negative input that recurs, and the one most often forgotten in year-two forecasts.

The measures that are not hours

Time saved is the easiest number and rarely the largest one. Several others are worth baselining even though they take more effort to convert into currency.

Error rates matter because errors have a tail — a mis-keyed invoice costs the correction plus the relationship damage plus the audit exposure. Cycle time compression often shows up as revenue rather than cost: leads contacted in five minutes convert at multiples of leads contacted the next day. Customer satisfaction is slow but shows up in retention, and retention is where the compounding is.

Make the maths reconcile

The test worth applying to your own numbers: would this survive someone from finance checking it against payroll and revenue records? If the claimed saving implies a headcount reduction that did not happen, or capacity that nobody can point at being used, the number will not hold.

This is not pedantry. A defensible modest number is more useful than an impressive one that collapses under the first serious question, because the defensible one gets you the budget for the next phase.