The earnings call is the one room where every public company is eventually asked what AI returned. In the second quarter of 2026, 65% of S&P 500 firms mentioned AI on those calls. Eleven percent quantified a productivity benefit for a specific use case. Two percent quantified an impact on earnings (Goldman Sachs US equity strategy, via Investing.com, Aug 2026). The subject came up almost everywhere. The number came up almost nowhere.
It is tempting to read that as caution — lawyers, guidance, the usual reticence about anything that might become a promise. It is simpler than that. Most of those companies cannot answer, because answering needs an artifact they do not keep: AI spend attached to specific work, an outcome someone accepted, and a cash effect documented and dated. Disclosure is not measurement. The funnel measures who can put a number on it in public, and the narrow end is what having a record looks like from the outside.
The drop is not modesty. It is a missing record.
Putting a number on AI’s effect on earnings takes three things most organizations never assembled. The spend has to be attached to named work, not sitting in one ledger line called AI. The work has to have an observed before and after, with someone recording that the output was accepted rather than merely produced. And the money has to reach a documented movement — an invoice that fell, a contract not renewed, a unit cost that moved on the same accepted volume. Skip any one of the three and the honest options are a story or silence. Silence is the more respectable answer, which is presumably why it wins.
The eleven percent shows where the wall stands. Those firms could quantify a benefit for one use case — a support queue, a coding workflow, a claims step — and then stopped. Getting from a use-case number to an earnings number means crossing from the operation into the accounts: showing that released time was actually reassigned or removed, netting the cost of running the AI itself, and defending the attribution against everything else that moved that quarter. That crossing is the whole job, and most of the firms holding a use-case number never attempted it in public.
Read the funnel carefully, because it does not say what a vendor would like it to say. Goldman’s own note records a small and statistically insignificant difference in earnings growth between the companies quantifying AI productivity gains this quarter and other S&P 500 companies. Being in the two percent is not evidence that AI paid off. It is evidence that the company can say what happened. Those are different claims, and only the second one is inside management’s control.
The shape is not new either. On Q4 2025 calls, 70% discussed AI, 10% quantified an impact on a specific use case, and 1% quantified an impact on earnings (Fortune, Mar 2026). The verbs differ between the two notes — discussed in one, mentioned in the other — so these are two readings of the same shape rather than two points on a line. Read either way, the top of the funnel is crowded and the bottom is nearly empty.
The earnings story so far belongs to the sellers
There is real AI money in the Q2 2026 results. Hyperscalers and AI infrastructure firms grew earnings 54% year over year and contributed roughly half of the index’s profit growth (Goldman Sachs US equity strategy, via Investing.com, Aug 2026). That is the selling side of the trade, and it is measured to the decimal, because selling produces revenue and revenue has an accounting standard behind it.
The buying side is measured too, just on the wrong axis. The share of US businesses on Ramp’s platform paying for at least one AI tool rose from 43.8% in September 2025 to 50.6% in April 2026. Adoption has a number. Spend has a number. What the spend returned is the one quantity with no meter attached, and it is the only one a board is actually asking about.
So the aggregate reads like a boom while the individual case reads like a guess. That is not a contradiction. It is what happens when the sellers’ revenue is audited and the buyers’ return is anecdotal. The two percent is the small set of companies where the buyer’s side of that trade has a number too.
Build the record before the question arrives
The decision is not what to say on the next call. It is what to keep, starting now, so that there is something to say on a later one. Take the use cases that are material and, for each, attach the AI spend to the work it is actually doing. Observe that work before and after against an acceptance standard stated in advance. Record what was accepted, what came back for rework and what was abandoned — an agent that fails at a task consumes the same budget as one that ships.
Then be strict about categories, because the strictness is what makes the number survive a sceptical reader. Cash is a documented, attributable movement in money paid or received, net of what the AI cost to run. Capacity is released time, and it stays in hours until a manager decides what happened to it. Structural value — traceability, control, repeatability — is real and is not cash. Modeled upside is anything about future revenue, and it stays labelled as modeled. These four sit beside one another; the moment they are added into one total, the record stops being usable in front of anyone who did not write it.
And date everything. Model prices move, capability moves, and the work itself moves. A unit cost observed in one quarter is an assumption by the next, so every entry carries a review date, and a material price or capability change pulls that review forward.
What would count as proof?
- A named use case with its AI spend attached to the work it is doing — not one line called AI in the general ledger.
- An observed before and after on that work, with the acceptance standard stated in advance and rework and abandonment counted, not only the successes.
- The cash effect documented and attributed — the invoice, contract or unit cost that moved — net of what running the AI cost.
- Released time kept in hours, with a named decision recording what happened to that capacity: reassigned, removed, or neither.
- Modeled upside labelled as modeled, reported beside the observed figures and never summed with them.
- A date and a review date on every figure, with evidence that at least one price or capability change pulled a review forward.
Every one of those is checkable by someone who does not trust you. That is the standard an earnings call applies, and it is the reason the number is hard rather than a reason to skip it.
What remains unclaimed?
Quantifying an earnings impact is not the same as having one. Goldman found no statistically meaningful difference in earnings growth between the companies that quantified AI productivity gains and everyone else, so a number said out loud is a statement about the record, not a result. Silence proves nothing either: most of the distance between mentioning AI and quantifying it is companies that never built the artifact, not companies that tried and failed.
Adoption is not return — a business paying for AI tools has bought capacity, not established value. A use-case gain is not an earnings effect until the money moves and the attribution survives everything else that moved with it. Released time is not cash until someone decides what happens to it. And a supplier’s revenue is not a customer’s return: the fastest-growing earnings line in the index belongs to the firms selling the tools, which says nothing at all about what the buyers got.
Two percent is a low bar, and it is the whole bar
Two percent of the index could put a number on it. The bar is not high — a figure, and the record standing behind it. It only has to exist before someone asks.