Money
How to measure ROI on an AI employee
Tobiloba Odejinmi · 27 Mar 2026 · 6 min · 951 words
Direct answer
Measure an AI employee in three tiers. Capability ROI is whether the process can run without you hovering. Realized ROI is whether hours and errors actually moved off a named desk. Strategic ROI is whether the company redesigned how work gets done. A seven-day build can give you the first. It does not make you a Frontier Firm. IBM and Microsoft have been clear: only about 9% of leaders say they are running one of those.
- Do not put week-one output in the strategic ROI column.
- Capability means it runs. Realized means a person got time back and quality held.
- Track the process, the exceptions, and the reviewer load, not 'AI usage'.
- The 9% figure is a warning, not a target you hit with one workflow.
What are the three ROI tiers?
I use three words so finance and the team do not argue past each other. Capability. Realized. Strategic. Most vendors mash them into one chart and call it transformation. That is how you get a week-one build dressed up as a company strategy.
Capability is binary enough to be honest. Did we ship a process that runs on real cases? Realized takes a month of volume and a person who will say whether their Tuesday changed. Strategic is rare. IBM puts the Frontier Firm number at 9% of leaders. Treat that as the background, not as a promise attached to a $7,500 week.
What counts as capability ROI?
Capability is the thing I can stand behind at handover. The process is mapped. The tools are connected. A reviewer can open a log. Someone on your side can explain the path. If a case comes in on Monday, the system does the first pass without a Slack thread to me.
That is worth money. It is also the ceiling for week one. I have built platforms that later served thousands of providers. None of them became 'strategy' on the day the first path went live. They became useful when the path kept working.
When does realized ROI start?
It starts when you can say, without a flourish, that a person did less of the pile and the quality did not fall over. At SmartComply, review work fell by about half. That is realized. We did not call it a new operating system for the company.
You need a baseline. Hours on the process before. Exception rate before, even if it is ugly. Then the same numbers after, plus inference and reviewer time. If you skip the baseline, you will invent the win. I will not put a fake dollar amount on a case I have not measured with you.
- Hours on the named process, not 'team productivity'
- Exception rate and time-to-correct a miss
- Reviewer hours as a real line, not a rounding error
- Inference spend on production volume, not the sample
Why is a week-one build not strategic ROI?
Strategic ROI means you changed how the firm works. Agents and people share the work on purpose. Data and ownership are good enough to scale past a pilot. That is the 9%. Most companies are still running experiments and calling them programs.
A seven-day build can be the first honest brick. It cannot be the building. If a vendor tells you otherwise, they are selling you a category, not a process. I would rather you leave the first week with one live workflow and a clean scorecard than a story about becoming a Frontier Firm.
What should you track in the first thirty days?
Track the cases that completed without a person. Track the cases that needed a person, and why. Track how long the person spent. Track failures that were silent. Silent failures are how you lose trust, and lost trust wipes realized ROI even if the hour count looks pretty.
I still sit in reviews. The useful ones are boring. Someone can explain a miss without opening six tabs. If you cannot do that at day thirty, you do not have realized ROI yet. You have a system that is still onboarding you.
How do you report this without padding?
Put capability in one sentence: it is live, here is the owner, here is the log. Put realized in a small table: hours, exceptions, reviewer time, run cost. Leave strategic off the page until you have more than one process and a year of not lying to yourselves.
If the table is modest, say so. Modest and true beats a payback date you cannot defend when someone in finance asks where the cash moved.
Questions people ask
What is capability ROI?
The workflow exists in production. It can take a real case from trigger to a result or a clean handoff. That is a useful milestone. It is not proof the business is cheaper to run.
What is realized ROI?
A named person spends less time on the pile, the miss rate is something you can explain, and the run costs are on a sheet. If you cannot point at those three, you still have a demo.
What is strategic ROI?
The operating model changed. People orchestrate agents. Work is redesigned, not just sped up. That is the Frontier Firm story. One week of build is not that story, and I will not sell it as if it is.
How soon can I claim payback?
Claim capability at handover if it is true. Look at realized ROI after you have enough real volume to trust the exception rate. Do not date a strategic claim to day seven.
Which numbers belong in a board pack?
Volume handled, hours off the named process, exception rate, reviewer hours, inference spend, and incidents. Leave 'productivity uplift' out unless you can show the desk it came from.
Written by
Tobiloba Odejinmi
Head of Engineering at 10mg Health. I have run engineering at Zeeh Africa and sold Insurpass and Shopl. I still write the code. If you have one process that still runs on people copying things, we can look at it in thirty minutes.
