What did the second-quarter earnings season show about AI results?
Adoption widened and proof did not follow it. Apollo's chief economist reports that 69 percent of S&P 500 companies described a live AI deployment, up from 64 percent the quarter before, while only 29 percent quantified a result.
The thinning continues above that. Apollo puts the share reporting a metric tracked over time at 2 percent, and finds no company at all breaking AI value out as its own KPI or as a line in the profit and loss account. So the same disclosure season that produced a broad majority of live deployments produced almost no measured ones, and no audited ones.
Where companies did quantify something, the results lean hard one way. Cost accounts for 70 percent of the disclosed proof points, against 22 percent for revenue. Apollo reads part of that as timing, since an efficiency gain lands inside an operation that already exists well before a new revenue line takes shape.
Set the two headline figures against each other and the conversion rate falls out. Of the companies describing a live deployment, fewer than half have attached a number to it: 29 of 69, which is about 42 percent. That subtraction is ours rather than Apollo's, and it is the number an owner should hold onto, because it is the rate at which a deployment has so far turned into evidence among the largest and most heavily scrutinized companies in the market.
| Rung on the disclosure ladder | Share of S&P 500 companies |
|---|---|
| Describes a live deployment, no quantified result required | 69% |
| Same measure, prior quarter | 64% |
| Quantifies a result | 29% |
| Reports a metric tracked over time | 2% |
| Reports AI value as its own KPI or profit and loss line | None |
| Share of quantified proof points attributed to cost | 70% |
| Share of quantified proof points attributed to revenue | 22% |
Why does a disclosure ladder matter to a private company?
Because a buyer grades the same way, without publishing the scale. The index behind Apollo's charts sorts disclosure into five levels, from a stated plan with no result yet, through a live deployment with no number, to a quantified result, then a result tracked over time, and finally a result reported as its own line.
Read as a ladder, the second-quarter numbers are a distribution and not four separate facts. Sixty-nine percent reached the second rung. Twenty-nine percent reached the third. Two percent reached the fourth. Nobody reached the fifth. Each rung is roughly an order of magnitude harder than the one below it, and the whole market stalls between the second and the third.
A private company sits on the same ladder whether or not it knows the rungs have names. An owner who says the operation has adopted new tools is on the second rung. An owner who says the change removed a measurable amount of cost from a named process last year is on the third. An owner whose management accounts have carried that measure monthly since is on the fourth, and is now in a very small minority of any company of any size.
The difference between those positions is not presentational. It is the difference between a claim a buyer discounts and a number a buyer can underwrite, and this data says the step between them is where nearly everyone stops.
“Every seller in this market now has an AI paragraph in the equity story, and the market has just shown how little of that paragraph survives measurement. A buyer does not pay for adoption. It pays for a number that was tracked before anyone thought about selling, because that is the only version it cannot be talked into and cannot be talked out of.”
How will a buyer test an AI claim in your numbers?
By asking where the saving went. A cost reduction that is real has a destination: a smaller line in the accounts, a headcount that did not get replaced, a contract that was not renewed. A claim that cannot be followed to a destination is a claim about behaviour, not about earnings.
That test bites hardest on the add-back schedule. An efficiency gain presented as an adjustment to earnings asks a buyer to capitalize a saving at the deal multiple, which is the most expensive thing any line in a quality of earnings exercise can ask for. When the largest companies in the market can produce a tracked metric only 2 percent of the time, an adviser on the other side has every reason to treat an untracked one as unproven and to strike it.
The cost-versus-revenue split is worth reading as a hierarchy of what gets believed. Cost carries 70 percent of the disclosed proof points and revenue 22 percent, which tells you which kind of claim companies have actually been able to evidence. A revenue claim asks a buyer to accept that customers behaved differently for a reason inside your operation, and it will be tested against retention, pricing and win rates that existed before the change and after it.
Expect the timing to be tested as well. If the measurement begins in the year of the sale, its baseline is whatever the seller says the old process cost, and a baseline chosen after the fact is not a baseline. A measure that predates the decision to sell is worth more in a process than a better measure that does not.
What should an owner or a sponsor do before the next process?
Instrument one thing properly rather than claiming several. Pick the single process where the change is clearest, define the measure, fix the baseline, and carry it in the monthly management accounts. One metric with a year of history behind it is worth more than a page of deployments, and this data suggests it will put you ahead of the overwhelming majority of disclosed claims anywhere.
Then decide what the claim is for. If the saving is durable and measured, it belongs in the run rate and should be visible in the trailing numbers rather than argued for in an appendix. If it is real but recent, present it as a trend with its baseline attached and let the buyer extrapolate. If it is neither, leave it out of the numbers entirely and let it sit in the operating narrative where it can do no harm.
Sponsors should apply the same grading at entry. A target whose thesis rests on an efficiency gain that nobody has tracked is being underwritten on the second rung of the ladder, and the exit will be graded on the third or fourth by whoever buys it next. The cost of instrumenting that measure during the hold is trivial against the discount applied to an unproven one at exit.
The broader point is about where the burden of proof has moved. While deployment was the scarce thing, saying you had deployed was informative. With a broad majority of the largest listed companies saying it, the statement has stopped separating anybody from anybody, and the scarce thing is the number underneath. Being able to produce that number is now the differentiator, and by this reading it is one that roughly seven in ten companies still cannot claim.
As of September 2026
Sources: Apollo Global Management, The Daily Spark, "From Who Is Deploying AI to Who Can Prove the ROI", by Torsten Slok, Partner and Chief Economist, published 11 September 2026 and read first-hand at apollo.com, for the S&P 500 second-quarter earnings season figures of 69 percent of companies pointing to a live deployment against 64 percent the previous quarter, 29 percent quantifying a result, 2 percent reporting a metric tracked over time, no company breaking AI value out as its own KPI or profit and loss line, and cost accounting for 70 percent of disclosed proof points against 22 percent for revenue. The five-level disclosure ladder paraphrased in the second section and in the table's first column follows the L1 to L5 methodology published at thevaluegap.io, which Apollo's charts credit as the underlying index; that source is used for the definitions of the levels only, and no figure in this article is taken from it or attributed to it. The conversion rate of about 42 percent, being 29 of 69, is our arithmetic on Apollo's two figures. The reading of the ladder as the grading a buyer applies in diligence, the treatment of an untracked saving as an add-back that will be struck, the reading of the cost and revenue shares as a hierarchy of what a buyer will believe, and all guidance to an owner or a sponsor are ours; Apollo addresses public-market investors and says nothing about private middle market processes. The companion article cited on AI disruption already priced into listed multiples carries its own sources.
Corrections: factual errors are corrected on the page and the correction dated. Write to admin@kadenwoodgroup.com.
This position sits within our business valuation practice.