Every CEO I talk to asks the same question in a slightly different voice: how will I know if this worked? It is the right question, and it is the one most vendors answer with a slide about productivity. This post gives you a way to answer it yourself, with the numbers you already have, before you commit a dollar.
Why is AI ROI so hard to measure in a mid-market company?
Three reasons, and none of them is the technology. First, most companies deploy AI as a general tool rather than as a change to a specific process, so there is no single number that should move. Second, the baseline is never recorded, because the project started with enthusiasm rather than a measurement. Third, the cost side is undercounted, because the subscription fee is visible and the two hundred hours your operations manager spent configuring and checking it are not.
Large enterprises get around this with analytics teams and long pilots. A mid-market company does not have that luxury and does not need it. What it needs is discipline about one number, applied at the start. When a five-person team I worked with generated $32 million in revenue, the difference was not more tools. It was that every system we built was pointed at one measurable moment in the sales process, and we knew what that moment was worth before we touched it.
Which four numbers are worth tracking?
Pick one of these as the primary measure for a first project. Track the others as supporting evidence, but let one number decide.
1. Revenue captured that was previously lost
Calls that went to voicemail after hours and were never returned. Quotes that were sent once and never followed up. Web inquiries answered the next day instead of in the next minute. Each of these has a close rate and an average value in your own records. An AI assistant that answers, books or follows up turns a known leak into revenue you can count, and this is usually the strongest first measure because it shows up on the top line.
2. Cost per completed unit of work
What does it cost you today, fully loaded, to process one intake form, one claim, one invoice or one client onboarding? Divide the monthly cost of the people and systems involved by the number completed. Measure it again at day 90. This number is honest in a way that "hours saved" is not, because it only moves if the work actually got done cheaper.
3. Cycle time on a money-bearing process
Days from inquiry to proposal. Hours from first call to booked appointment. Days from service delivered to invoice sent. Shortening these pulls revenue forward and often raises conversion, since the prospect who hears back in ten minutes is still interested. Record the median before and after, not the average, so one slow case does not hide the change.
4. Headcount avoided
If the project meant you did not make a hire you had budgeted, that is a clean, defensible return, and it is often the single largest line in a mid-market ROI calculation. It only counts if the hire was real: approved, budgeted and now withdrawn. A vague sense that you will need fewer people someday does not belong on the scorecard.
Which three numbers mislead?
Hours saved. This is the number every vendor leads with, and on its own it is worth nothing. Time saved becomes a return only when it is redeployed to something you can see in the numbers: an avoided hire, a cleared backlog that was costing you revenue, or staff moved onto selling or billable work. If you cannot say where the hours went, do not count them.
Adoption and usage. Logins, prompts per day and the percentage of staff using the tool tell you people are curious. They do not tell you the business made money. Usage is a leading indicator at best, and a distraction at worst, because it lets a project look successful while the process it was meant to change sits exactly where it was.
Customer satisfaction scores taken too early. Satisfaction is worth tracking, but a 90-day pilot is too short for it to mean much, and a survey sent during a change often measures the disruption rather than the result. Record it, keep it off the decision, and look at it again at six months.
What does the full cost of an AI pilot include?
Put every one of these on the cost side, or the calculation is fiction.
- Software subscriptions and any usage-based fees, at the volume you actually ran, not the volume on the proposal.
- Implementation, integration or consulting fees, whether paid to a vendor or to a firm like Torque AI.
- Your own staff's hours: configuration, testing, reviewing outputs, handling exceptions and attending the meetings. Estimate honestly and multiply by a loaded hourly rate.
- The manual fallback you kept running in parallel, which during a pilot is a real double cost.
- Any one-time cleanup of data or process that the AI could not work without.
I covered what these line items typically add up to in the post on what a revenue engine is and why most small firms lack one, and the short version is that internal time is usually the largest cost and the one most often left out.
How do you set up the measurement before the project starts?
Week 0: Write the number down
Name the one process, the one number and its current value from the last 30 days. Write the target and the date you will check it. Name the person who owns the number. If this takes more than a page, the project is too broad.
Week 0: Run a premortem on the measurement itself
Ask the team to write the memo explaining, 90 days from now, why the number did not move or why nobody believed it when it did. The usual answers are a baseline that was never recorded, a definition that changed mid-pilot, and a cost that was never counted. Fix those three now. This is the same premortem-first discipline that keeps marketing budgets honest.
Days 1 to 90: Report every two weeks
The owner reports the number, the running cost and nothing else. Two weeks is frequent enough to catch a definition drifting and infrequent enough to avoid reacting to noise.
Day 90: Decide on the scorecard
One page: the baseline, the day-90 value, the full cost, the modeled payback period and a recommendation to scale, adjust or stop. The CEO signs it either way. A stopped project with a clean scorecard is a success, because the method now works for the next one.
What does a good result look like?
A first pilot that pays back its full cost inside six to nine months is solid. The best first projects pay back inside the 90 days, usually because they were aimed at a leak that was larger than anyone had measured, such as overflow calls or unfollowed quotes. If the modeled payback is longer than a year, the problem is almost always scope. Narrow the process rather than extending the pilot.
The scorecard also tells you something the number alone does not: whether your company can run an AI project with a measurement from start to finish. That capability is the actual return on the first project. Once you have it, the second and third projects are cheaper, faster and far easier to approve, because the board has seen a page it trusts.
If you would rather have the first number chosen, the baseline recorded and the ROI model built by someone who has done it across law firms, wealth managers, insurance agencies and family-owned companies, that is what the AI Clarity Sprint does in three weeks at a fixed fee. Details are on the engagements and FAQ page. The prompts and procedures behind this kind of work are in ChatGPT for Profit, Volumes 1 and 2.
Frequently asked questions
- What is a good ROI for an AI project in a mid-market company?
A first pilot should pay back its full cost, including staff time, within six to nine months of going live, and the best first projects do it inside the 90-day pilot itself. If the modeled payback is longer than a year, pick a narrower process. The point of the first project is to prove the measurement method as much as the result.
- How long does it take to see ROI from AI?
For a narrow process that touches revenue, such as inbound call handling, quote follow-up or intake, a measurable change usually shows up within 60 to 90 days of going live, provided a baseline was recorded first. Company-wide tool rollouts take much longer to show anything, because nobody agreed on what to measure.
- Should I count time saved as ROI?
Only if the time was redeployed to something that makes or saves money and you can say what that was. Hours saved that turn into longer coffee breaks are not a return. The safer approach is to count time savings only when they avoided a hire, cleared a backlog that was costing revenue, or moved people onto billable or selling work you can see in the numbers.
- What costs should be included in an AI ROI calculation?
Software subscriptions and usage fees, any implementation or consulting fees, the hours of your own staff spent configuring, testing and reviewing the system, and the cost of the manual fallback you keep running during the pilot. Leaving out internal time is the most common way companies convince themselves a project worked when it did not.
- Do I need a data team to measure AI ROI?
No. One number, a 30-day baseline, a 90-day comparison and a one-page scorecard are enough for a first project, and the owner of the pilot can keep them in a spreadsheet. A data team becomes useful when you are running several AI projects at once and want to compare them, not before.
Want the number, the baseline and the ROI model built before you spend?
The AI Clarity Sprint gives mid-market companies, law firms, wealth managers and insurance agencies an AI audit, deployment recommendations with ROI models, and a 90-day plan in three weeks.