How to Measure AI ROI at Work (With a Calculator)

How do you measure the ROI of AI at work?

Measure AI ROI by comparing a task's results before and after AI, then putting a dollar value on the difference and dividing by the full cost of using AI. The formula is ROI = (annual benefit − annual cost) ÷ annual cost. The hard part is not the math. It is capturing a baseline before rollout, counting every cost (including review time), and counting only the time savings that turned into real work or real money.

TL;DR

• Capture the baseline first. If you did not measure how long the task took, how many errors it had, or what it produced before AI, you cannot prove anything afterward.
• Pick one workflow at a time. "AI at the company" is too vague to measure. "First-draft contract summaries" or "post-call follow-up emails" is measurable.
• Count the full cost. Licenses are the small part. Setup, training, human review of AI output, security and legal review, and ongoing admin all belong in the number.
• Discount saved time. An hour saved is not an hour of value unless it goes to other useful work. Apply a realization factor (40 to 60 percent is a defensible starting assumption) and say so openly.
• Use a control group for revenue claims. In sales especially, compare users with similar non-users over the same period before crediting AI with any revenue.
• Report a range, not a single number. Leaders trust a conservative, an expected and an optimistic case more than one confident figure.

What do you need before you start measuring?

Three things, and you need them before anyone turns the tool on.

One workflow. Choose a specific, repeated piece of work with a clear start and end. Good examples: drafting follow-up emails after sales calls, summarizing vendor contracts, answering internal HR questions, reconciling invoices, writing first drafts of weekly reports. Bad examples: "productivity" and "general use of the chatbot."

A baseline. Record how the work is done today. You want at least:

• Time per unit of work (minutes per email, hours per report)
• Volume (units per person per week)
• Quality or error rate (rework, corrections, complaints, audit findings)
• Output or outcome (meetings booked, tickets closed, deals won), if the workflow ties to one

Two weeks of real data beats a guess. If you cannot wait, ask five people who do the work to time themselves on ten tasks each.

An owner for the number. Someone in finance, operations or the business unit should agree to the assumptions up front, especially the hourly cost and the realization factor. Numbers that finance helped build survive budget reviews. Numbers the AI champion built alone usually do not.

How do you calculate AI ROI step by step?

1. Define the workflow and the unit of work. Example: "One post-call follow-up email."
2. Record the baseline. Minutes per unit, units per week, error rate, number of people.
3. Run a pilot with a comparison. Give AI to one group and keep a similar group on the old process for 4 to 8 weeks. If you cannot split groups, compare the same people before and after, and note anything else that changed (new quota, seasonality, a new manager).
4. Total the costs. Use the cost table below.
5. Value the benefits. Time, quality, revenue and avoided cost, each handled separately and conservatively.
6. Run the formula and the payback period. Then test it by changing your two or three shakiest assumptions.

What costs should you include?

Most failed ROI cases undercount cost. Include all of these:

Use a loaded hourly cost (salary plus benefits and overhead) rather than a raw wage. Ask finance for the standard figure for each role. Do not invent one.

What benefits should you count?

Keep four categories separate so no one can accuse you of double counting.

1. Time saved. (Minutes saved per unit × units per week × people × weeks worked) ÷ 60 × loaded hourly cost × realization factor.

The realization factor is the share of saved time that becomes productive work. If a rep saves 30 minutes on admin and uses it to make more calls, it counts. If it disappears into a longer lunch or more meetings, it does not. Start at 40 to 60 percent unless you can show where the time went.

2. Quality and risk. Fewer errors, less rework, fewer compliance misses. Value these by the cost of the old problem: hours spent on fixes, credits issued, penalties avoided. For Legal & Compliance teams, this is often the stronger case than speed. Only count it if you have the before-and-after error counts.

3. Revenue. More deals, faster cycles, higher conversion. Count it only when a control group supports it. Otherwise list it as an upside, not as part of the headline ROI.

4. Avoided cost. Work you would have hired or outsourced for, such as a contractor you did not engage. Count it only if the spend was actually planned and actually avoided.

Calculator: the AI ROI template you can copy

Copy this into a spreadsheet. Fill the blue-sky cells only with measured or finance-approved numbers.

AI ROI CALCULATOR

WORKFLOW: ______________________
PILOT PERIOD: ___ weeks   PEOPLE IN PILOT: ___   PEOPLE IN COMPARISON GROUP: ___

INPUTS
A. Loaded hourly cost of the role ........................ $____
B. Units of work per person per week (baseline) .......... ____
C. Minutes per unit BEFORE AI ............................ ____
D. Minutes per unit WITH AI (including review) ........... ____
E. People who will use it ................................ ____
F. Working weeks per year ................................ ____
G. Realization factor (share of saved time used well) .... ___%

COSTS (YEAR ONE)
H. Licenses / usage ....................................... $____
I. Setup (hours × hourly cost) ........................... $____
J. Training (hours × people × hourly cost) ............... $____
K. Governance and security/legal review .................. $____
L. Ongoing management (hours/week × weeks × hourly cost) . $____
M. TOTAL COST = H + I + J + K + L ........................ $____

BENEFITS (YEAR ONE)
N. Hours saved = (C − D) × B × E × F ÷ 60 ................ ____
O. Time value = N × A × G ................................ $____
P. Quality / risk value (documented only) ................ $____
Q. Revenue value (control group only) .................... $____
R. Avoided cost (planned spend only) ..................... $____
S. TOTAL BENEFIT = O + P + Q + R ......................... $____

RESULTS
ROI % = (S − M) ÷ M × 100
Net benefit = S − M
Payback (months) = M ÷ (S ÷ 12)

Build three columns: conservative (G at 40 percent, no revenue), expected, and optimistic. Present all three.

What does a worked example look like?

These numbers are illustrative, not benchmarks. Use your own.

Workflow: a 10-person sales team using AI to draft post-call follow-up emails and update CRM notes.

Baseline: each rep spends 25 minutes per call on follow-up and notes, and does 8 calls a week. With AI, including review, it takes 15 minutes. That saves 10 minutes per call, or 80 minutes (about 1.3 hours) per rep per week. For the example, call it 3 hours a week once you include other drafting AI also speeds up. Loaded cost is $60 an hour. Reps work 46 weeks a year.

Costs (year one):

• Licenses: $30 per user per month × 10 users × 12 = $3,600
• Setup and training: $3,800 (rep time plus an admin's time)
• Manager review and prompt upkeep: 1 hour a week × 48 weeks × $80 = $3,840
• Total: $11,240

Benefits:

• Hours saved: 3 × 10 reps × 46 weeks = 1,380 hours
• Realization factor 40 percent: 552 hours × $60 = $33,120
• Revenue claimed: $0 (the conservative case)

Result:

• ROI = (33,120 − 11,240) ÷ 11,240 ≈ 195 percent
• Payback = 11,240 ÷ (33,120 ÷ 12) ≈ 4 months

If the realization factor drops to 20 percent, benefit is $16,560 and ROI is about 47 percent. That still passes, and it is worth showing, because it tells a skeptical executive the case does not depend on one optimistic assumption.

How do you measure AI ROI in sales?

Sales is where people most want to claim revenue and where claims are easiest to attack. Measure in two layers.

Leading indicators (you can measure within weeks):

• Time spent on admin per rep per week (CRM updates, follow-ups, call prep, proposal drafting)
• Speed to first response on new leads
• Number of meaningful customer touches per rep
• Follow-up completion rate (what share of calls get a next-step email within 24 hours)
• Rep time spent selling versus not selling

Lagging indicators (take a full sales cycle or more):

• Pipeline created per rep
• Stage-to-stage conversion and win rate
• Sales cycle length
• Average deal size
• Quota attainment

How to keep the claim honest:

1. Use a comparison group. Pair reps by territory, tenure and segment. One group uses AI, one does not, for at least one sales cycle.
2. Fix the definitions before you start. Decide what counts as a qualified meeting or a touch, so no one adjusts the definition later.
3. Credit revenue only for the measured difference. If AI reps closed 8 percent more pipeline than matched non-AI reps, value that gap, then haircut it, because small groups are noisy.
4. Separate the tool from the training. Reps who volunteer for a pilot tend to be your strongest. Choose pilot members randomly if you can, or say plainly that you did not.
5. Watch for quality drops. Faster, more generic emails can hurt reply rates. Track reply rate and unsubscribes next to volume.

For most sales teams, the defensible headline is time returned to selling, with revenue shown as a separate upside line until a full-cycle comparison supports it.

What prompt can help you build the ROI case?

AI is useful for structuring the analysis, finding costs you forgot and stress-testing your assumptions. Paste this into any capable assistant, replacing the brackets. Do not paste confidential customer data or salary details your policy does not allow.

You are a finance analyst helping me build a conservative ROI case for an AI tool at work.

Workflow: [describe the task and unit of work]
Team: [roles and number of people]
Baseline: [minutes per unit, units per week, error rate, any outcome metric]
With AI (pilot results): [minutes per unit including review, error rate, outcome]
Costs I know: [licenses, setup hours, training hours, review time]
Loaded hourly cost: [figure approved by finance]
Working weeks per year: [number]

Do the following:
1. List any cost categories I have missed (review, governance, management, rework).
2. Calculate hours saved, time value at realization factors of 30%, 50% and 70%, total cost, ROI % and payback in months. Show the arithmetic.
3. Tell me which three assumptions most affect the result and how the ROI changes if each is 25% worse.
4. List the questions a skeptical CFO would ask and what evidence would answer each.
5. Flag any claim I am making (especially revenue) that my data does not support.

Do not invent benchmarks or industry statistics. If you need a number I have not given you, ask for it.

What to check in the answer:

• Redo the arithmetic yourself, or put the figures in your spreadsheet. Language models make calculation slips, especially in multi-step math.
• Confirm that every input in the output is one you supplied. If the answer includes an "industry average" you did not provide, delete it or find the real source.
• Make sure review time is in the "with AI" minutes and not forgotten.
• Check that revenue is not sitting inside the headline ROI without a control group.

Where not to trust AI: it cannot know your loaded cost, your real adoption rate or whether saved time was used well. It also tends to produce confident, round-number benefits. Treat its output as a draft and a checklist, not as evidence.

Why do AI ROI numbers fall apart under questioning?

These are the usual failure points:

• Counting self-reported time savings as fact. People overestimate. Time a sample of tasks, or compare system data such as tickets closed or emails sent.
• Ignoring adoption. A license that 30 percent of staff use is a 30 percent benefit at 100 percent of the cost. Include actual weekly active use in your calculation.
• Measuring only the pilot enthusiasts. Early adopters save more than the average employee. Apply a lower rate when you project to everyone.
• Counting hours that were never a cost. If no one will be hired, cut or reassigned, saved time is capacity, not cash. Say "capacity gained" and show what it was used for.
• Skipping the quality side. Faster work with more errors can be negative ROI. Always track one quality measure.
• Reporting once. Usage drops after the novelty fades. Re-measure at 30, 90 and 180 days.

How should you present the result to leadership?

One page is enough:

1. The workflow and the decision you want (expand, hold, stop).
2. The baseline and the pilot result in plain numbers.
3. Total cost and total benefit, in the three scenarios.
4. ROI and payback for each.
5. What you did not count (revenue upside, morale, risk reduction).
6. The risks and the controls: who reviews output, what data is allowed, who owns the tool.
7. The next measurement date.

Executives and managers tend to respond to payback period and the downside case. Legal & Compliance will ask about data handling, review steps and audit trail, so put those in the cost section rather than leaving them to be discovered.

Can you build your own AI ROI tracker instead of using spreadsheets?

If you are measuring several pilots across teams, the spreadsheet gets messy: baselines in one file, time logs in another, costs in a third. Remy is a product agent that builds full-stack business apps from a plain-English description, so you could describe a small ROI tracker and have it built and deployed. For example: a form where each pilot owner logs baseline and pilot minutes per task, a table for costs, role-based sign-in so finance can approve assumptions, and a dashboard that applies the formula above and shows conservative, expected and optimistic cases. The app runs on a live URL, and the code is yours in a git repo.

Frequently Asked Questions

5 questions

01What is a good ROI for an AI project?

There is no universal number, and any figure quoted without a source for your industry should be treated with suspicion. A better test is whether the conservative case still beats your company's usual hurdle for similar investments, and whether payback fits your budget cycle. Many teams look for payback inside the first year for tools that need little custom work.

02How long should I run a pilot before measuring ROI?

Run at least 4 to 8 weeks for time-based workflows, so novelty wears off and you collect enough volume. For sales revenue claims, plan for at least one full sales cycle. Measure again at 30, 90 and 180 days to see whether usage and savings hold.

03How do I measure ROI when the benefit is quality or risk, not time?

Count the old problem's cost: rework hours, credits, penalties, audit findings or escalations, before and after. If you lack before data, you cannot claim the improvement, so start logging now and report it as a future measure. Compliance teams often benefit most from a documented error rate and a review log.

04Should I include the cost of employee time spent learning AI tools?

Yes. Multiply training hours by the number of people and their loaded hourly cost. It is real spending even though no invoice arrives, and leaving it out is one of the first things a finance reviewer will catch.

05Can I use AI to calculate its own ROI?

You can use it to structure the analysis, find missing cost categories and challenge your assumptions, using the prompt above. Do not rely on it for the final arithmetic or for benchmarks. Verify calculations in a spreadsheet and use only inputs you can source from your own data or from finance.