Funding AI in a PE-backed company means applying the same commercial test to an AI spend request that every other investment in the business has to pass: what number it claims to move, the baseline it moves that number from, what it costs fully loaded, and who owns the result once it ships. The decision is rarely hard because the technology is hard. It is hard because AI proposals arrive described as capability rather than as a return, and a management team with no way to convert a capability claim into a number already in the plan ends up approving or declining on how convincing the presenter was.
This guide is written for the executives and sponsors who have a proposal in front of them and have to approve it, decline it, or send it back. It covers what such a request has to evidence, the three things most of them turn out to be, the ninety-day signal worth asking for, and whether to put an AI premium in the exit model. It names no tools and explains no technology. The underwriting discipline that applies to every spend line sits in investment marketing, the payback calculation in marketing payback period, and the company-level revenue build in the annual commercial plan. What follows is only what an AI request has to show that an ordinary spend request does not.
Start here, because most of the confusion in these decisions comes from treating the category as exempt. AI changes the cost and the speed of certain inputs. It does not change the acquisition economics of the business, the payback test applied to money spent, the data quality required to see a result, or who is accountable for the revenue number. A request that needs one of those four suspended in order to look attractive is a decline.
The measurement point is a precondition rather than a consequence. A company that cannot measure revenue by acquisition source cannot measure what AI did to it, in either direction. This is not a theoretical gap: in one Claymore engagement we found 38 percent of new business arriving through word of mouth that no system was recording. A company in that position can neither attribute an improvement to a tool nor rule out that it came from somewhere else. The sequence for fixing it sits in fixing marketing attribution and building a single source of revenue truth, and it is worth reading before any proposal is scored rather than after.
Two cautions on evidence. The published numbers on AI adoption, returns and valuation premiums are largely self-reported survey opinion, much of it from firms that sell AI services, and a recognizable name on the cover does not make an opinion a measurement. Use the company's own data or none. And no proposal should be underwritten on capability that has to arrive later for the case to work. The honest version of an uncertain decision is to size it as uncertain, not to resolve the uncertainty with a prediction.
The pressure to get this right is not abstract. S&P Global's 2026 research found 71 percent of GPs and 53 percent of LPs now prioritize operational value creation, so a sponsor is being asked what the operating companies actually did, and an AI line with no evidence behind it reads badly in that conversation.
This is the structural reason the usual template does not fit. A channel request claims customers: expected volume, cost to acquire, payback period, named owner. An AI request almost always claims something one step back from that. Quotes drafted per day. Minutes to respond to an inbound call. Hours of analyst time returned. Those are real quantities and they are not revenue, and the step between them is where the proposal usually goes quiet.
The move that closes the gap is to name the constraint the input sits inside. If the input being improved is not the thing limiting output, then making it faster produces a saving and no revenue at all. A tool that halves the time to produce a quote is worth very little in a business already issuing quotes the same day and losing them on price. The same tool is worth a great deal where quotes sit for four days and the loss reason is a competitor who answered first. Identical claim, identical cost, opposite answers, and the only thing separating them is a constraint the proposal rarely mentions.
So the first question is not what the tool does. It is which number in the operating review is currently held down by the input it improves.
Nearly every AI proposal that reaches a management team is one of three things. Each is legitimate, each carries a different test, and each should be signed by a different person.
Classification is also the fastest way to find the proposals that are not investments at all. A request that cannot be placed in any of the three categories, because it names no input, no constraint and no number, is a procurement decision dressed as a growth initiative, and it should be approved or declined on procurement grounds: price, contract term, exit cost, and whether anyone will use it.
Once the category is agreed, four questions settle the decision. They are deliberately ordinary, which is the point.
The last question is the most useful and the one most often skipped: what does this proposal predict you will see in ninety days that you would not have seen anyway? Not whether it has paid back. Judging an investment on ninety-day payback is its own mistake, and the right horizon sits in the payback guide. The ninety-day question is narrower. It asks whether the mechanism is doing the specific thing the case said it would.
The counterfactual half matters more than the forecast half. Over any quarter the business will also have seasonality, a price change, a new location, a sales hire and whatever the market did. A prediction that cannot be separated from those is how a tool gets credit for a good quarter it had no part in. Write the prediction, and what it has to be distinguished from, on the same page before the money is committed.
A usable prediction has three properties. It is specific to the mechanism rather than to the outcome, so a faster quote turnaround is readable where a better quarter is not. It is visible in a report that already exists, because a prediction requiring new reporting will not be checked. And it is capable of coming back negative, which is the property most pilot designs quietly remove. A pilot with no success criterion and no stop date is not a pilot. It is a subscription with a trial period attached.
Most proposals should be sent back, and the reason is nearly always the baseline. That is not a stall. A request returned with one specific instruction, produce this number for the last four quarters and come back, is cheaper than a funded initiative nobody can assess a year later.
Approve when the category is agreed, the baseline exists, the cost is fully loaded, the owner already holds the number, and the ninety-day prediction can fail. Decline when the claim cannot be tied to a constraint, when the predicted signal is indistinguishable from what would have happened anyway, or when the case only works if the payback test is suspended for this one line. Treating those as close calls is how a company ends up with a portfolio of tools nobody owns.
One note on scale. The right first commitment is usually the smallest one that can still produce the ninety-day signal, which is often far smaller than the proposal in front of you.
Buyers ask about AI, so sponsors reasonably ask whether there is a premium to underwrite. The honest answer is that the evidence for one is thin and self-reported, and a survey of expectations is not a valuation input. Putting a premium in the model on that basis creates a number that has to be defended in diligence with material that will not support it.
What does travel is ordinary. A measurable improvement in a number a buyer already underwrites, evidenced from the company's own records, holds its value in diligence whatever produced it. A better conversion rate, a shorter cycle, a lower cost to serve, each reproducible from source data, is worth more than any narrative about tooling. In one Claymore engagement, commercial work of that kind produced a 36 percent revenue lift over 24 months on flat marketing spend, and what made it credible to a buyer was that every component could be rebuilt from the company's own transaction records. The evidence set a buyer will ask for is laid out in the commercial evidence pack and the run-up to a sale in commercial exit readiness.
The practical discipline, then, is the same one that governs the rest of the spend: fund the mechanism, evidence the number, and let the buyer decide what the mechanism was worth. An AI proposal that cannot name the number it moves, the baseline it moves it from, and the person who owns that number afterwards has not been evaluated yet, whatever the meeting concluded.
Fund the requests that can name a number already in the operating plan, show that number's current value from the company's own data, state a fully loaded cost rather than a license fee, and name an executive who is already accountable for that number afterwards. In practice that tends to mean work sitting inside a real constraint, or the measurement foundation that other requests depend on. Decline or defer anything whose return rests on capability the business cannot demonstrate today, and never fund a three-year plan whose returns depend on capability that does not exist yet.
Treat it as an investment case for an input rather than for a channel. Name the input the proposal makes cheaper or faster, name the constraint that input sits inside, and show what happens to output when the constraint moves. If the input is not the binding constraint, the saving is real and the revenue effect is zero. Then state the baseline, the fully loaded cost including integration, data work, internal hours and ongoing ownership, the accountable owner, and what you expect to see in ninety days that you would not have seen anyway.
A cost substitution wearing revenue framing, where software does work people currently do; a real capacity gain inside an existing constraint, where the input being improved is the thing actually limiting output; or a measurement project that has to happen first, because the targeting, forecasting or pricing promise rests on data the company does not yet hold in usable form. Each one is legitimate. Each one has a different test, a different cost profile and a different person who should sign it, which is why classifying the request before testing it saves the argument.
Not as a line in the model. Buyers do ask about AI, and the published evidence for a price premium is mostly self-reported survey opinion about expectations, much of it from firms selling AI services, which is not a basis for a valuation assumption. What travels at exit is a measurable improvement in a number a buyer already underwrites, evidenced from the company's own records. If the work produced a better conversion rate, a shorter cycle or a lower cost to serve, underwrite that. The mechanism behind it is a detail of the story, not the value.
The executive who already owns the number the proposal claims to move, not the person who proposed the tool and not a technology function by default. Ownership of the outcome and ownership of the implementation are separate jobs, and the second does not substitute for the first. If the only named owner is the sponsor of the request, and that person is not accountable for the number in the operating review, the initiative has no owner and will not be measured after the first quarter.
Have a revenue problem the board is asking about? Start a conversation.