Measuring AI ROICounting what actually changed

Why AI ROI is so often overstated#

AI ROI claims tend to follow a familiar pattern: a task that took an hour now takes ten minutes, multiplied across the team, multiplied by salary, equals an impressive annual figure. The arithmetic is fine. The assumptions are not.

The saved time may be partly eaten by review and correction. It may not go anywhere productive. The faster output may be weaker. And the costs usually include only the subscription, not the set-up, training, governance and maintenance. An honest framework accounts for all of this, which makes its conclusions less exciting and far more useful.

Time saved is a promise. Value created is a receipt.

Four levels of value#

Activity

How many people use the tool and how often. Useful for adoption, useless as proof of value. A tool can be used constantly and add nothing.

Efficiency

Net time or cost saved per task, after accounting for review, correction and rework. Measure before and after on the same task type, timed honestly.

Quality

Whether the output is better, the same or worse. For ads, conversion rates; for content, engagement and editorial scores; for research, accuracy against source. Quality often decides whether efficiency gains are real.

Business outcomes

Whether the business moved: more pipeline, lower acquisition cost, better retention. This is where time saved must eventually show up, through reinvestment in work that matters, or through lower costs.

Count the full cost#

Subscriptions are the visible part of AI cost. The rest is easy to miss, and leaving it out is the most common way ROI gets inflated.

• Licences and usage. Seats, credits, usage-based charges, which can grow unpredictably.
• Set-up. Integration, configuration, building brand context and prompt libraries.
• Training. Time for people to learn the tool and the workflow.
• Review time. The ongoing human effort of checking and correcting output.
• Governance. Policy, approvals, audits, legal review.
• Rework and incidents. Fixing errors that got through, including any reputational cost.
• Switching. The cost of changing tools later, if output and workflows are locked in.

A useful habit is to estimate these costs per workflow, not per tool. One assistant might serve five workflows, each with its own set-up, review and governance burden. Allocating cost this way shows which workflows carry their weight and which ride on the others.

Establish a baseline before you start#

Without a baseline, every ROI figure is a guess. Before introducing AI to a workflow, record how long the task takes, how much it costs and what quality and outcome levels look like today. Two to four weeks of honest measurement is usually enough.

For outcome-level measurement, the gold standard is a comparison group: some campaigns, regions, teams or customers use the AI-assisted approach, others do not, and you compare results. Where that is impractical, compare before and after while watching for other changes that could explain the difference. See incrementality testing.

A worked calculation#

Try setting the reinvestment share to zero. If the result goes negative, the workflow only pays back if the saved time is genuinely put to work, or if it leads to fewer hours being paid for. That is a management decision, not a software feature.

Measuring quality and outcomes#

For quality, define the measure before you start. For AI-assisted ads, compare conversion rates against human-only variants in the same test. For content, use an editorial score plus downstream engagement. For research, check a sample of AI summaries against the source material and count errors.

For outcomes, be patient. A faster content workflow will not change pipeline in a month. Agree the time horizon in advance and resist declaring victory or failure too early.

Five traps that inflate the numbers#

Most inflated ROI figures are not dishonest. They come from reasonable-looking shortcuts that compound. Knowing them makes your own figures more defensible and helps you question vendor case studies.

• Estimating instead of timing. People guess how long tasks took before AI, and guesses run high. Time real tasks.
• Ignoring review. Counting drafting time saved while leaving out the editing and fact-checking it created.
• Best-case extrapolation. Taking the most impressive task and assuming every task saves the same.
• Counting capacity as cash. Treating freed hours as money saved when nobody's hours or costs actually changed.
• Attribution by coincidence. Crediting AI for an outcome that improved for another reason, such as seasonality, a price change or a new channel.

Vendor case studies are worth reading for ideas, not for numbers. Their conditions, data and definitions are rarely the same as yours. Your own baseline and comparison are the only figures that should drive budget decisions.

Reporting AI ROI to leadership#

Leadership needs a short, honest summary per workflow: what was done, what it cost in full, what it saved net, what happened to quality, what happened to outcomes and what you will do next. State the confidence level openly. A modest, well-evidenced result builds more credibility than a large, assumption-heavy one. See board reporting.

When to stop#

An honest ROI framework will sometimes say no. If a workflow saves little time after review, lowers quality or shows no outcome effect after a fair period, stop it. Stopping is not failure; it frees budget and attention for workflows that work. See AI marketing strategy for how pilots fit into a wider plan, and CAC and LTV for the outcome measures that matter most.

Equally, be willing to scale quickly when the evidence is strong. A workflow that saves real time, holds quality and moves an outcome deserves training, documentation and wider rollout, not another quarter of cautious piloting. Measurement exists to make both decisions easier: stop the weak, back the strong.

Published by Fabulous.Media, a network of specialist marketing agencies. Updated 9 October 2026. Platform features change often; check current official documentation before acting on platform-specific detail.