Measuring AI ROI the way a buyer will measure it

Why most AI ROI numbers don't survive diligence

The most common AI ROI claim in the lower middle market is "we saved X hours". It is also the claim a quality-of-earnings provider will most readily strike out. Bain's CEO guide to generative AI is unusually direct on this: hours saved is not the same as cost removed, and time saved that does not translate into either headcount avoidance or measurable throughput is invisible in the financial statements[1].

The bar a QofE provider will apply is the bar AICPA's quality-of-earnings guidance sets out: adjustments must be evidenced, repeatable, and traceable to the underlying books and records [3]. Productivity claims that don't meet that bar simply don't make it through to the adjusted EBITDA the deal is priced from.

The four metrics buyers will credit

1. Headcount avoidance (or redeployment)

The cleanest metric. "We did not hire the additional CSR our volume forecast required because the triage agent handles 60% of inbound autonomously." Verifiable against the volume curve and the org chart.

2. Error-rate reduction

Error rates that map to either revenue (mis-quotes, missed renewals) or cost (rework, refunds, write-offs). A documented before/after with the underlying ticket or ledger evidence.

3. Throughput per FTE

Volume of work (orders processed, tickets resolved, claims handled) divided by direct FTEs in the function. A 12-month trend showing the ratio rising while quality holds is one of the cleanest signals of operational leverage in lower-middle-market diligence — and it flows straight through to multiple via the operational maturity premium GF Data tracks [2].

4. Gross-margin lift

Where AI compresses cost-to-serve (automated reconciliation, automated documentation, automated tier-1 support), the impact lands in gross margin. A trailing-twelve-months margin curve with an inflection at the deployment date — and the deployment date documented — is the strongest form of AI ROI evidence.

How to log it so it actually survives

• Capture the baseline before deployment. Volume, error rate, FTE count, gross margin — all snapshotted with date.
• Log every AI run. Counts, costs, and outcomes. The "did it work" question must be answerable from data, not memory.
• Tag the deployment in your accounting system. A clear date and a brief note in the chart of accounts that lets a QofE analyst align operational changes with financial inflection points.
• Re-measure quarterly. A rolling four quarters of evidence is the credible minimum. Less than that and a buyer will discount the claim.
• Tie it back to the multiple narrative. "Throughput per FTE up 38% over 18 months at constant quality, gross margin up 240bps, no incremental headcount" is the form of sentence that argues for multiple expansion. "We use AI more" is not.

The point

AI ROI is real. AI ROI is also the easiest area for a target company to oversell and a buyer to discount. The owners who get paid for their AI work are the ones whose claims a QofE provider can verify in an afternoon — not the ones whose claims live in a management deck.