The "ROI" trap

When a CFO asks "what is the ROI of our AI project", the honest answer is rarely a single percentage. It is three or four numbers, two of which are not monetary. Trying to compress them into one ratio either inflates the result (most common) or hides genuine value (also common). The teams we see succeed at scaling AI past the first project all stop pretending there is one number, and instead agree on a small portfolio of metrics. Below are the three frameworks we use, in the order we typically introduce them.

Framework 1: Time saved per task, validated

The most defensible AI metric is time saved on a clearly bounded task. The trick is in the validation. A vendor case study saying "saves 4 hours per week per employee" usually means: someone estimated it once. The validated version is: pick 8–12 employees, time the task before deployment for two weeks, deploy the AI, time the same task for two weeks, calculate the difference and the variance. If the variance swallows the gain, the gain is not real. We have seen confidently reported "30% time savings" disappear under proper baselining; we have also seen modest 8% gains turn out to be rock-solid and aggregate to seven-figure annual savings at scale. The discipline is what produces a defensible number, not the size of the number.

Framework 2: Quality and consistency

Time savings without a quality measure is dangerous. If your AI assistant produces sales proposals in half the time but they convert at 60% of the previous rate, you have lost money. Build at least one quality metric per AI workflow. For sales proposals: win rate. For support: first-response resolution rate and CSAT. For internal Q&A: accuracy of answer (sampled and rated by an SME). For document drafts: edit distance from human-finalised version. Quality metrics typically need 90 days of data to be meaningful — they move slower than time savings — but they are what tells you whether the time saved is real value or hidden cost.

Framework 3: Capability shift

The third metric is the one CFOs are least comfortable with and the one that turns out, in retrospect, to be the most strategically important. It is not "did we save money", it is "what can we now do that we could not before". Concrete examples: support team handling 40% more ticket volume without adding headcount; sales team able to localise into a third language without a translator on staff; analyst team running 3x as many ad-hoc data queries because they no longer need to write SQL. None of these show up cleanly as ROI in year one. All of them are why teams that adopted AI seriously in 2024 are pulling away from competitors in 2026.

Pre-launch: baseline everything

You cannot measure impact without a before-state. The single most common mistake we see is launching the AI tool first and only then trying to assemble a baseline. By that point, memories are biased, processes have already adapted slightly, and the comparison is suspect. Spend the two weeks before launch baselining the three or four numbers you will report on. Even a rough baseline beats a perfect post-hoc estimate.

Reporting cadence that survives contact with reality

Monthly is too noisy for AI metrics in the first six months. Quarterly is too slow to catch problems. The cadence we recommend: a weekly internal review of operational metrics (latency, error rates, usage) and a monthly business review of impact metrics (time, quality, capability). Both go to the same dashboard, both have an owner, both are reviewed by the executive sponsor every quarter. The dashboards that disappear after launch are the ones nobody is responsible for. Name an owner.

A note on costs

AI costs are easy to undercount. Visible: API/model costs, vendor fees, infrastructure. Less visible: SME time during deployment (often 60–120 hours), ongoing prompt and retrieval maintenance (typically 0.2–0.5 FTE for non-trivial deployments), training and onboarding hours across the user team. We recommend building a 24-month total cost of ownership before signing anything; it usually doubles the visible cost number, and it sets honest expectations with finance.

What to put in the board pack

Three numbers, in plain language. Time saved per week per affected employee, with the methodology footnoted. One quality metric, with direction (better, same, worse) and confidence (high/medium/low). One capability statement (something you can now do that you could not before). Total 24-month cost of ownership against total 24-month projected value. Five lines, not fifty. We are happy to share our reporting template — get in touch if it would be useful.