How to measure AI ROI in hours saved (without inventing numbers)

Most "hours saved" figures are guesses. The owner feels faster, the vendor quotes a percentage, and nobody counts the ten minutes spent checking each draft. That matters, because the decision to keep paying for an AI execution layer, add more tasks or switch tools should rest on numbers you trust.

The method below takes about an hour to set up and fifteen minutes a week to maintain. It works for any AI tool, not just ours.

Why gut feel is a poor measure of time saved

The best evidence we know of comes from a randomised trial run by the research group METR in early 2025. Experienced open-source developers worked on real tasks in codebases they knew, some with AI tools allowed and some without. With AI allowed, tasks took 19% longer. Before starting, the developers had expected a 24% speed-up, and afterwards they still believed they had been about 20% faster.

That study used early-2025 tools on complex software, and METR is careful to say it doesn't show AI rarely helps. Other work points the other way: an NBER study of customer-support agents found an AI assistant raised issues resolved per hour, with the largest gains for less experienced staff and little gain for the most experienced. The point for an operator isn't which number is right. It is that the effect varies by task and person, and that people are poor judges of their own speed. So measure.

Step 1: choose five tasks and sample a baseline

Pick five recurring tasks the agent will take on. Good candidates are frequent, well defined and easy to time: receipting a payment, drafting a follow-up, updating a CRM record after a call, preparing an invoice with notes, building a ten-row prospect list.

• Sample, don't estimate. Time at least five real instances of each task before handing it over. Start the timer when you start the task, stop it when the output is ready to send or file.
• Include the switching cost. If you have to open three systems to do the task, that time counts.
• Record volume. How many times a month does this task happen? Use last month's actual count, not a guess.
• Note quality. Did any of the baseline instances need fixing later? That is your baseline error rate.

Step 2: time the agent-assisted version honestly

After the agent has been doing a task for two weeks (so you are past the learning curve), sample again. This time, the clock measures your time, not the agent's:

• Time spent writing or speaking the instruction.
• Time spent reading and approving the draft.
• Time spent fixing anything wrong before it goes out.
• Time spent later fixing anything that went out wrong (the rework).

Agent processing time doesn't count against you unless you are waiting on it. If you have to sit and watch, it does.

Step 3: the spreadsheet template

One row per task. The formula for net minutes saved per month is: (baseline minutes − assisted minutes) × monthly volume − rework minutes. The rows below are illustrative, to show how the arithmetic works.

Check the arithmetic on the first row: (4 − 1) × 120 = 360 minutes, minus 20 minutes rework = 340 minutes, or about 5.7 hours. Keep the sheet next to the original timings so anyone can audit it.

Count recovered money in its own column

Some of the best outcomes aren't time at all. Agents that read every lease, timesheet and invoice find money that a busy person misses. Two of our published customer stories are about exactly that:

• Tony Lorenzo at Elite Partners Real Estate says his agent caught a commercial lease being billed about $1,049 a month below the signed lease.
• Irene Gyane at Fourth Woman Therapy says Cortex found completed work that hadn't been billed and turned it into invoices.

Individual results vary. Record found money as a one-off or recurring amount with the evidence attached, and never convert it into "hours". Mixing the two makes the whole report less credible.

What not to count

• Tasks you wouldn't have done anyway. If the agent built a 200-row list you would never have built by hand, that is new capacity, not hours saved. Report it separately.
• Drafts you rejected. A rejected draft saved nothing and cost review time.
• Agent minutes. The agent working for an hour is not an hour saved.
• The first two weeks. You are still writing rules and correcting style; it will look worse than reality.

Turning hours into a keep, expand or stop decision

Once you have a month of data, compare net hours saved against cost. Value an owner's hour at what it would cost to hire the work done, or what the owner bills clients, whichever is more honest for your business. Then compare against your subscription and any review help you pay for. For reference, Cortex plans and the optional human-in-the-loop add-on are listed in AUD on the pricing page.

• Keep and expand if net hours are positive, errors that reached customers are zero or falling, and review time per task is dropping.
• Fix the brief if rework is high on one task. That usually means the definition of done is vague.
• Stop a task if after a month it still costs more review time than it saves. Not everything should be delegated.

A good daily report makes this easier, because it records what was done and approved. We cover what to put in one in designing a daily exception report. And if you are deciding whether licence savings justify a switch, pair this sheet with our SaaS consolidation maths, or with the build-versus-buy comparison if you are weighing a DIY stack.

If you want a baseline to measure against, the outcomes page shows the kinds of work Cortex agents take on; choose five of yours and start timing this week.