How to measure AI ROI

Most businesses can't say whether AI is paying for itself. Not because the benefit isn't there, but because nobody wrote down what things took before. Once that baseline is gone, all you're left with is impressions — and impressions lose to a spreadsheet at budget time.

Measure tasks, not people

"Are we more productive?" is unanswerable. "How long does it take to produce a monthly client report?" is answerable, and that's the level to work at.

Pick three or four tasks that are:

Resist measuring everything. Four tasks measured properly beats twenty measured vaguely.

The calculation

Per task:

(minutes before − minutes after) × times per month × hourly cost ÷ 60

Then, against it:

licences + training hours × hourly cost + setup time

Keep the hourly cost honest — salary plus on-costs, not just the wage. Overstating it inflates the result and the number stops being believable.

An Australian small-business example

A 12-person Adelaide engineering consultancy picks three tasks:

TaskBeforeAfterPer monthSaved
Site report first draft90 min35 min1211 hrs
Tender summary120 min50 min44.7 hrs
Client update emails15 min6 min406 hrs

That's about 21.7 hours a month. At $75/hour loaded, roughly $1,630. Against around $400 in licences and an initial 20 hours of training and setup.

The training cost is recovered in the first month, and the ongoing figure is defensible because it's built from timed tasks rather than a feeling. Note what's absent: no claim about "transformation", no company-wide percentage. Three tasks, measured.

What to watch out for

The part most businesses skip

Decide what the recovered hours are actually for before you start. More client work? Faster turnaround as a selling point? A vacancy you don't need to fill?

Without that decision, twenty hours a month diffuses into slightly longer breaks and slightly less pressure — real for the people involved, invisible on any measure. That's how a project with genuine returns ends up looking like it delivered nothing.

Frequently asked questions

How long before we can measure anything?
Give it six to eight weeks. The first fortnight is learning and the numbers look worse than reality; after that they stabilise. Measuring in week one tells you about the learning curve, not the tool.
What if the time saved just becomes slack?
Then the ROI is real but unrealised, and that's a management question rather than a tool question. Decide in advance what the freed hours are for — more client work, faster turnaround, or a role not backfilled. Otherwise the saving quietly evaporates.
Should we count quality improvements?
Count them where you can tie them to something — fewer revisions, fewer complaints, faster approvals. Note the rest qualitatively but don't inflate the number with unmeasurables; it undermines the credible parts.
What's a realistic saving?
It varies enormously by task. Drafting, summarising and first-pass research tend to show the largest gains. Work needing judgement, relationships or physical presence shows little. Beware anyone quoting a single company-wide percentage.
Do we count the training time as a cost?
Yes — including it is what makes the figure defensible. Licences plus training time plus the productivity dip while people learn. A number that ignores those is the kind that gets picked apart.

Put this to work

Ad On Group runs AI training and enablement for Australian teams through Ad On AI — a three-month, self-paced program that takes non-technical staff from their first prompts to working AI agents.

Talk to us →

Keep reading

← All resources