Headcount math: how to estimate AI ROI without lying to your CFO

The most common AI ROI calculation in any pilot proposal is fake.

It usually goes something like this: “This tool saves your team 10 hours per week, which at a $150/hour blended rate is $78,000 per year, so the $75,000 pilot pays back in under 12 months.” Clean math. Easy to put in a slide. Persuasive in a 30-minute pitch meeting.

It’s also wrong almost every time.

The wrongness compounds in three places. First, the hours-saved estimate is usually whoever-talked-to-the-vendor’s casual guess, not a measured number. Second, the saved hours don’t actually convert to dollars at the blended rate — most companies don’t fire anyone, so the saved hours show up somewhere else in the org without showing up on a P&L line. Third, the calculation ignores all the costs that aren’t the pilot price: integration overhead, training time, error rates, ongoing maintenance, the inevitable second engagement to handle edge cases the first one didn’t.

A good CFO knows all of this. When a vendor opens with the $78k-back-on-$75k math, the CFO smiles politely and walks the deal back to procurement to die. So we use a different framework.

The honest version

There are three numbers that matter. Get all three accurate and the ROI conversation becomes real.

Number 1: FTE-equivalent absorbed

This is NOT total hours saved across the company. It’s the equivalent in headcount of the work the tool actually absorbs.

The calculation is:

FTE-equivalent = (hours saved per week × weeks worked per year) / 2,080

2,080 is the standard US working-year hours (52 weeks × 40 hours). Adjust if you measure productivity differently.

Worked example. Six credit analysts each spend 30% of their time pulling data from internal systems and writing standardized memos. The tool absorbs 80% of that work.

Hours saved per week = 6 analysts × 40 hours × 30% × 80% = 57.6 hours
Hours per year = 57.6 × 52 = 2,995 hours
FTE-equivalent = 2,995 / 2,080 = 1.44 FTE

So the tool absorbs about 1.4 FTE worth of work across your analyst team.

Note that this number is honest about scope. It doesn’t say “we saved your firm a million dollars.” It says we absorbed the equivalent of slightly less than 1.5 full-time analyst headcounts of repetitive work. Whether that’s economically valuable depends on what those FTEs cost and what they do with their freed time.

Number 2: The realized dollar value

This is where most ROI math goes wrong. Saved hours are not dollar savings unless one of three things happens:

(a) You hire fewer people next year than you would have. Replace the next planned headcount add with the tool. This is the cleanest dollar conversion. If you’d have hired a $90k-fully-loaded analyst, that $90k stays in the P&L.

(b) Your existing people produce more output that you can monetize. If freeing 1.4 FTE worth of time means your analysts review 30% more loan applications and you actually convert more deals as a result, the dollar value is real and measurable.

(c) You cut headcount. Rare in practice. Most firms re-deploy freed time. But if a workflow really was redundant work and you’re shrinking that team anyway, the math is one-to-one.

If none of (a), (b), or (c) applies, the dollar value of the saved hours is zero. The freed time gets absorbed into whatever your team was doing in the margins. The FTE-equivalent is real but doesn’t show up on a P&L.

A CFO will respect a vendor who admits this. A CFO will dismiss a vendor who pretends every saved hour is fungible cash.

For our worked example, if the bank’s next planned hire was a credit analyst at $95k fully loaded and we absorb 1.4 FTE, the realized first-year dollar value is roughly $95k × 1.0 = $95k (you decide not to hire ONE planned analyst; the additional 0.4 FTE of freed time goes to expanded coverage).

Number 3: Total cost of ownership

The pilot price is not the total cost. There are four more cost buckets that need to be in the math.

Integration overhead. Your in-house engineers will spend some time supporting the integration. Estimate 0.1 FTE for the first quarter post-launch, then 0.02 FTE ongoing. At $130k fully loaded, that’s about $13,000 in year one.

Adoption time. Your operators spend time learning the tool, reviewing its outputs, retraining when it changes. Estimate 5 hours per operator in the first month, 1 hour per month thereafter. For our 6 analysts at $95k fully loaded ($46/hour), that’s 6 × 5 + 6 × 1 × 11 = 96 hours = roughly $4,400 in year one.

Ongoing maintenance. Either you pay the original vendor a retainer or you pay your in-house team to maintain the tool. Budget 5-10% of the build cost per year for either path. On a $75k pilot, that’s $4,000-7,500 per year.

Error and edge case handling. The tool will produce wrong outputs sometimes. Operator review catches most. The cost of bad outputs that slip through depends on the workflow. For a credit memo tool with operator review on every output, this is essentially zero. For a fully autonomous tool with no review, it can be the largest cost in the system. Scope your tool accordingly.

For our worked example, total year-one cost is about $75,000 + $13,000 + $4,400 + $6,000 = $98,400.

Putting it together

Year-one payback for the worked example:

Realized value:     $95,000
Total cost:         $98,400
Year-one net:       -$3,400

This pilot doesn’t pay back in year one. It pays back in year two, when the recurring cost drops to about $16,000 and the recurring value (one fewer analyst on payroll) stays at $95k+. Cumulative year-two net is positive $79,000-ish.

That’s a fine outcome for a real pilot, and a CFO who has seen this math respects it. The vendor’s pitch should ALSO show this math. Anyone selling you “12-month payback” with optimistic numbers is either confused or selling.

What to do with this

Three things, in order:

Before the pilot starts, get the FTE-equivalent number right. Watch operators do the work for a few hours. Time the steps. Use those measurements, not a vendor’s estimate.

Before the pilot starts, identify which of (a), (b), or (c) applies for dollar conversion. If none of them does, the pilot might still be worth running for non-financial reasons (employee retention, faster cycle times, lower error rates) but don’t pretend it’s a financial decision.

Build the year-one and year-two total-cost picture before the engagement, not after. The CFO will ask. Having the math ready before the question gets asked is the difference between a pilot that gets approved and one that dies in procurement.

This is honest math. It’s not flattering to the vendor. It’s also the only kind that lets pilots actually get signed and survive the CFO review six months in. The vendors who show you this math are the ones to work with.