Agentic AI

How to measure AI ROI, before and after the project

How to measure AI ROI: set a baseline, count every cost, run a measured pilot and decide on thresholds agreed in advance. With a fill-in business case template.

Someone in your business wants to fund an AI project, and you are the one who has to say yes. The demo looked convincing, the supplier talks about hours saved, and nobody can quite tell you how much the current process costs today. That is the real problem with AI ROI (return on investment): it is rarely the technology that is hard to judge, it is the missing “before”.

In this article we set out a method for finance directors and chief executives: measure the baseline, count every cost, run a pilot that can answer the question, and decide on thresholds you agreed before the results came in. You will also find a fill-in template for your AI business case.

Why AI ROI is hard to measure, and why that is no excuse

Many AI projects are judged on impressions: the tool “feels faster”, the team “likes it”. Good signs, but they do not pay for usage fees or upkeep.

Three things make AI ROI slippery:

  • No baseline. If you never measured how long a task took before, any number after is a guess.
  • Hidden costs. The build is visible; the time your staff spend checking the AI’s output often is not.
  • Hours saved are not money saved. An hour freed only has value if it goes to something useful: more volume, faster service, less overtime, or work that was waiting.

None of this rules AI out. It means building the case like any other investment, with numbers you can check.

How to measure AI ROI: start with a baseline

Measure the current process for two to four weeks before anything changes. Choose one task, not a whole department: reading documents, sorting emails, preparing quotes. If you are starting from invoices, our guide to AI invoice processing shows where the time usually goes.

Four measures are enough.

Time

Minutes per item, from start to finish, including the checking and the back-and-forth. Ask two or three people to log a sample: estimates are usually too low.

Volume

Items per week or month, and how much that varies. A process that handles 50 items in a quiet month and 400 in a busy one does not have the same case as a steady one.

Errors

How many items are corrected later, and how long each correction takes.

Delays

How long an item waits between arriving and being finished. Delay is often the cost customers notice most, even when it is hard to put in money.

Write the baseline down, date it, and have the process owner sign it off: it is your reference for every decision that follows.

Count every cost, not just the build

A supplier’s quote covers the build. Your business case needs the full cost over at least two or three years.

Build and integration

Design, development, connection to your existing tools (email, document storage, CRM or ERP) and testing. This is usually a one-off cost, paid as a fixed price or by the day.

Usage fees

AI models are usually billed by use. Prices change often, sometimes several times a year, so price your volume at today’s rates, note the date, and allow a margin.

Human review

In most serious projects, a person checks some or all of the output, at least at first. That time is a real cost. Leave it out and your pilot will look better than the live system.

Upkeep and change

Monitoring, adjustments when your documents or rules change, model updates, and the hours spent training staff. Budget for them every year, not just at launch.

A fill-in template for your AI business case

Here is a template you can fill in. Copy it, replace the figures with yours, and keep the money side and the service side separate.

Illustrative example (fictitious figures). A team processes incoming documents. Amounts are in any currency, and an hour of staff time costs 50 once salary and overheads are included.

Line Baseline After pilot
Documents per month 2,000 2,000
Minutes per document, checking included 6 4
Staff hours per month 200 133
Documents corrected later 4% 2%
Hours spent on corrections per month 20 10
Average wait before processing 3 days 1 day

Gains. 220 hours a month before, 143 after: 77 hours saved, or 924 hours a year. At 50 an hour, that is 46,200 a year, if those hours go to useful work.

Costs. Build and integration 30,000 and training 3,000, both one-off. Usage fees 4,000 and upkeep 6,000 a year.

Result. Net yearly gain of 36,200 (46,200 minus 10,000). The one-off 33,000 is paid back in about 11 months. The shorter wait (three days to one) stays on its own line: it matters, but we do not turn it into money unless you can show what it earns.

Fill in the “after” column only with measured figures, never with the supplier’s promise.

Run an AI pilot project that can answer the question

A pilot is not a demo. It is a short, real run on your own data, with your own people, designed to fill in the “after” column. Our article on AI agents for SMEs explains how to choose the task; here is how to measure it.

Set the thresholds before you start

Agree in writing what result means “go”, “adjust” or “stop”. In the example above, it could be:

  • Go if minutes per document drop to 4.5 or less, corrections do not rise above the baseline, and payback is under 18 months.
  • Adjust if time drops but corrections rise: fix the weak spots, then measure again.
  • Stop if time barely moves after the adjustments. Stopping a pilot is a good result: you spent a little to avoid spending a lot.

Keep the pilot real

Use real documents, including the messy ones. Keep a person checking every output at first, then reduce checks only where the error rate allows it. Run it long enough to cover a normal cycle, often four to eight weeks, so a quiet week does not flatter the numbers.

Report a range, not a single number

An honest pilot report gives a range rather than a single figure, and says what that range depends on in your own process. We quote our own real reference the same way: in the United Arab Emirates in 2025, an AI agent for document processing, built with Zazco Tech and led by our founder, Anass Chabbab, in his previous roles, saved 25% to 50% of the time spent on that work (see the UAE document AI case). We quote the whole range, not its best end.

After go-live: keep measuring

The baseline does not retire when the project goes live. Track the same four measures every month, alongside usage fees and review time, on the same page as your other indicators. Our guide to an executive dashboard with the right KPIs (key performance indicators) shows how to keep that page short.

Watch for three signs:

  • Review time creeping back up, often because documents or rules changed.
  • Usage fees rising faster than volume, which calls for a look at how the model is used.
  • Saved hours that never reappear as extra volume, faster service or less overtime.

And be ready to say when AI is not the answer. If a task is rare, short or already well handled by a simple rule in your existing software, a better form or a small automation may beat any AI project. We would rather tell you that in a first call than build something that never pays back.

For document work, an AI document agent is among our products in preparation.

Frequently asked questions

How long should an AI pilot project last?

Long enough to cover a normal cycle of your activity, which is often four to eight weeks. A pilot that runs past three months risks becoming a half-finished live system with no clear decision. Fix the end date and the thresholds before you start, and hold the go, adjust or stop meeting on that date.

What is a good ROI for an AI project?

There is no universal figure, and we would be wary of anyone who quotes one. A good return is one that clears the thresholds you set before the pilot, for example a payback period your finance team accepts, with no rise in errors. Compare it with your other investments: an AI project should meet the same bar as a new machine or a new hire.

Can you measure AI ROI without a baseline?

Only roughly. You can rebuild a partial baseline from old records, such as timestamps in email, ticket or document systems, but it will miss checking time and corrections. If the project has already started, measure the current state now and treat it as your reference for every later change. A late baseline is far better than none.

In short

AI ROI is measured like any investment: a dated baseline, the full cost over several years, and measured results. Time, volume, errors and delays are enough to start. A pilot with thresholds agreed in advance turns a debate about impressions into a clear go, adjust or stop decision. After go-live, keep the same measures on your dashboard so the gains stay real.

Weighing an AI project and unsure what it should deliver? Book a free 30-minute discovery call: we look together at your process, what to measure first, and whether our agentic AI services are the right fit, starting with our AI and automation audit.

Related servicesAgentic AI

Have a project in mind?

Let’s talk about it on a 30-minute discovery call, free and with no obligation.

Book a discovery call

Further reading

All articles