How to measure the return on AI you already bought

Most advice on measuring AI return assumes you can still do the clean thing: record a baseline, then turn the tool on. That is the right sequence when you have not bought anything yet. It is not the situation most portfolio companies are in. The copilots, the accounts-payable automation, and the customer-service agent are often already on the books. The board is no longer asking whether AI is interesting. It is asking what the spend returned.

This note is a practical way to reconstruct a defensible “before,” compare it with what you can observe, and label the result honestly. It is not marketing measurement, and it is not a method for crediting campaigns or channels. It is a finance question: what did the AI already paid for change in the P&L, and how sure are we?

Most of the spend is already on the books

Finance is being asked for this number more often than it can produce it. In a 2026 CloudZero survey of 260 senior finance leaders, 135 of them CFOs, only 22% could tie AI spend to business outcomes, while 87% said they needed to do so within the year, and 66% of boards were conditioning further AI funding on proof of return (CloudZero, 2026).

Inside private equity, the measurement that exists often stops short of the P&L. Accordion’s PE AI adoption benchmark, a Q2 2026 vendor survey of technology and AI operating partners, found that 67% measure FTE impact or time savings, and only 38% measure revenue or margin impact (Accordion, 2026). The sample size was not shown on the source page. In the same survey, 47% of operating partners described a tool-by-tool, use-case-by-use-case approach with no shared measurement (Accordion, 2026).

McKinsey, writing in April 2026, noted that horizontal tools such as chatbots, copilots, and summarizers can help people work faster and still rarely change a P&L, and reported, from its latest Global Survey on AI, that 60% of respondents had still not seen enterprise-wide EBIT impact from their AI programs (McKinsey, April 24, 2026). The survey’s sample size was not stated in that article.

If the tool is already live, you do not get a clean pre-period by wishing for one. You reconstruct it from the records you have, or you say that you cannot.

What “return” has to mean here

A usage dashboard is not a return. Tokens, seats, and tickets “touched by AI” are activity. They can support a claim. They are not the number a CFO takes to the board.

Return, here, means a change in a financial or operational line that finance will recognize, set against the full cost of the initiative, with the method written down. Hours saved count as a P&L result only if you can show where the capacity went: a cost that came out, output that went up without adding staff, or work that stopped being done. Otherwise the honest label is that time was freed, not that EBITDA moved.

Every figure should carry a confidence label: measured, estimated, projected, or not yet attributable.

A four-step way to reconstruct the count

You do not need a new data platform to start. You need a short list of initiatives, their cost, the records that already exist, and a rule for what you will and will not claim.

1. Inventory the spend, not just the licenses

List every AI cost line for the initiatives in scope: subscriptions and usage, implementation and integration, training, ongoing review time, and internal staff time. The point is total cost of ownership, not the vendor invoice alone. If you cannot name the cost of an initiative, you are not ready to name its return.

2. Rebuild the “before” from data you already have

A reconstructed baseline is a picture of the workflow as it ran before the tool was in use, built from system history rather than from memory. Useful sources are ordinary ones: general-ledger and accounts-payable detail, ticket or order systems, ERP cycle times and error or rework rates, timekeeping or headcount by function, and the official rollout dates. If you know who got the tool, and when, you may be able to compare earlier and later groups. If those dates were not kept, the comparison gets weaker, and the label should get weaker with it.

Use a period long enough to include a normal operating cycle. Where seasonality matters, compare like with like. Where volume, mix, headcount, or a parallel redesign also changed, say so. A baseline that ignores those shifts will not survive a controller’s questions — and should not.

Gaps are expected. Accordion’s 2026 benchmark reported that 75% of portfolio companies arrive at close unprepared or only partially ready for AI deployment, and that a poor data foundation ranked first among barriers to scaling (Accordion, 2026). A gap in the record is not a license to fill it with a vendor’s savings estimate. It is a reason to narrow the claim.

3. Choose a comparison, or say you cannot

McKinsey’s advice for new rollouts is to build measurement into the design, through A/B tests or staggered deployment, so results stand up to scrutiny (McKinsey, April 24, 2026). When the tool is already live, you look for the comparison that the history will still support. This is also how TribuTrace reconstructs a baseline.

Four options cover most cases:

A staggered rollout you can still see. Some teams, sites, or regions received the tool later than others. If the groups are comparable, the difference in the same period is the most defensible reconstructed comparison.
A before-and-after with the other changes named. You compare a like period before and after go-live, and you list the other initiatives, volume shifts, and staffing changes that also moved in that window. The result is usually estimated, not measured, because you cannot fully separate the causes.
A matched comparison. A similar workflow, product line, or sister company that did not get the tool. This is only as good as the match. Write down why you consider the two groups comparable, and where they are not.
Not yet attributable. The data cannot support a comparison that finance will stand behind. That is a valid finding. It tells you what to capture before the next renewal or the next purchase.

Rebuilding a baseline from records the company already has is a narrower job than building measurement from scratch. It works with the ticket logs, invoices, time records and system data that exist today, rather than new instrumentation. And when those records will not support a claim, “not yet attributable” is a valid answer, not a failure of the method.

4. Put cost and result in one ledger

For each initiative, write one line that a finance owner can defend: the total cost for the period; the operational change you observed, with its data source; the financial translation, if you can make one; the method you used; the confidence labels; and a keep, fix, or stop recommendation.

Keep the supporting detail close: the period, the volume, the other changes in the business, and the limits of the comparison. If hours were freed and you cannot show a cost out or a volume up, say that. If a vendor promised a saving you cannot find in the data, say that too. The ledger is the artifact. The board or LP one-pager is a one-page cut of the same lines, not a different story.

A hypothetical example

The following is a worked illustration only. It is not a client result, and the figures are invented to show the method.

Suppose a portfolio company turned on an invoice-coding tool in accounts payable nine months ago. No baseline was written down. Finance can see the subscription and implementation cost. It cannot yet say what the tool returned.

The inventory shows a license, a one-time integration, and some exception-review time. The reconstructed “before” comes from the ERP: invoices processed, time to code a clean invoice, exception rate, and rework cost, taken from the same season a year earlier and from the months just before go-live. Rollout was staggered: two sites first, two later.

At the early sites, cycle time and exception rates moved after go-live in a way that did not appear at the later sites in the same window. Volume was stable, and no other AP redesign ran in parallel. That comparison can be translated into a cost-per-invoice figure and labeled measured, with the limits stated: two sites, nine months, this workflow only. The later sites are not yet attributable. Hours no longer spent on coding have not shown up as a headcount reduction; some went to exception review. Until finance can show where the rest of the capacity went, any EBITDA claim on “hours saved” stays off the ledger or is labeled estimated and boxed as incomplete. A keep/fix/stop call might be: keep the tool at the early sites, fix the exception load before expanding, and do not claim a company-wide return yet.

None of those numbers should be copied out of this illustration. Only the shape of the work should be.

When “not yet attributable” is the right answer

It is the right answer when rollout dates were not kept, when the only “before” is a recollection, when too many other changes hit the same workflow, when the system cannot produce a consistent unit of work, or when all you have is a vendor’s projected saving. That is a decision input, not a failure of the finance team. It tells you what to instrument before the next renewal, and what not to put in an LP letter.

What to take to the board

A short pack is enough: the inventory of AI spend, by initiative; the ledger, with cost, result, label, and keep/fix/stop for each item in scope; the method, including what you could not compare; and what will be measured the same way next quarter, and who in finance owns it. Lead with outcomes, not model names. Include the initiatives you would stop. A pack with no stops and no labels will be read as a narrative, not as a count.

If you want a structured pass at this for one portfolio company, the AI Value Baseline is the fixed-scope version of the same work: a handful of initiatives already in use, a reconstructed baseline, a labeled ledger, and a routine your team can run afterward.

Sources

Figures were checked against the live pages on November 10, 2026.

CloudZero, “Finding the ROI of AI: The Finance Perspective,” 2026 survey of 260 senior finance leaders, 135 of them CFOs. The report states that 22% can tie AI spend to outcomes, 87% say they must within the year, and 66% of boards condition further AI funding on proof of return. https://www.cloudzero.com/finance-needs-ai-roi-2026-survey-report/
Accordion, “The PE AI adoption benchmark,” responses collected in Q2 2026, a vendor survey of technology and AI operating partners. The page reports that 67% measure FTE impact or time savings, 38% measure revenue or margin impact, 47% describe a tool-by-tool approach with no shared measurement, and 75% of portfolio companies arrive at close unprepared or only partially ready for AI. Sample size was not shown on the page. https://www.accordion.com/the-pe-ai-adoption-benchmark/
McKinsey, “From promise to impact: How companies can measure—and realize—the full value of AI,” April 24, 2026. The article reports, from McKinsey’s latest Global Survey on AI, that 60% of respondents had not seen enterprise-wide EBIT impact from their AI programs. It also recommends building measurement into rollout through A/B testing or staggered deployment, and notes that horizontal tools such as copilots rarely change a P&L. The survey’s sample size was not stated in the article. https://www.mckinsey.com/capabilities/quantumblack/our-insights/from-promise-to-impact-how-companies-can-measure-and-realize-the-full-value-of-ai