Proving AI return on investment: measure the portfolio, not the pilot

Most organisations can point to AI activity. Far fewer can point to AI returns. Boards that approved early investment on the promise of productivity are now asking a sharper question: what has it delivered, and how do we know?

The difficulty is rarely that AI has failed. It is that value was never defined in a way that could be measured, so success cannot be shown even where it exists.

Why pilots cannot prove value on their own

A pilot is designed to answer "can this work?" — not "is this worth it at scale?". Pilot results also tend to flatter:

• the most motivated teams volunteer first;
• the best-suited processes are chosen;
• time saved is self-reported rather than observed;
• the cost of support, licences, data work and oversight is left out.

When a pilot is then scaled, the results disappoint and confidence in the whole programme drops. The answer is not more pilots. It is a measurement discipline that runs across all of them.

Manage AI as a portfolio

Treat every AI use case as a product in a single portfolio, scored on the same basis. For each use case, agree up front:

• the outcome it should move — a measure the business already tracks, such as cost per case, cycle time, first-time resolution or error rate;
• a baseline taken before the change, not reconstructed afterwards;
• a named benefit owner in the business, not in the technology team;
• the full cost — licences, integration, data preparation, human oversight and support;
• a decision point at which the use case is scaled, changed or stopped.

A shared scoring model lets leadership compare a contact-centre assistant with a finance reconciliation agent on equal terms. It also makes stopping weak use cases routine rather than political.

Separate capacity released from cash realised

The most common error in AI business cases is counting hours saved as money saved. Time released only becomes value when it is redeployed or removed. Report the two separately:

• Capacity released: hours freed, measured through observed process data where possible.
• Value realised: what that capacity was turned into — lower cost, higher throughput, better service or reduced risk.

Being honest about this distinction is what makes the numbers credible to a finance director. It also shows where management action, not technology, is needed to realise the benefit.

Build the evidence as you go

Benefits evidence assembled at year-end is weak evidence. Capture it as the work happens:

• record the baseline and its data source when the use case is approved;
• track the measure on a fixed cadence and log changes in method;
• note the assumptions behind each estimate and who challenged them;
• keep the decision trail for scale, change or stop calls.

This is the core of the Ekselens System of Proof: the difference between a figure someone asserts and a figure the organisation can evidence. In our assessments, the gap between the two is often the most useful finding for a board.

Include risk reduction, but price it honestly

Some AI value is defensive — fewer errors, faster detection, better compliance. That value is real but harder to show. Use planning estimates, label them clearly as estimates, and pair them with an observable indicator such as error rates, rework volumes or audit findings. A defensible estimate is worth more than an impressive one that falls apart under scrutiny.

A practical first step

Take the five largest AI use cases in flight. For each, check four things: is there a baseline, a named owner, a full cost and a decision date? Most organisations find at least half fail one test. Fixing those gaps takes weeks, not months, and turns a collection of experiments into a portfolio a board can manage.

AI will keep attracting investment. The leaders who keep that investment flowing will be the ones who can show, use case by use case, what it has returned.