88% of companies run AI. 6% can show it on their earnings.

The survey behind the headlines is The State of AI in 2026, McKinsey's annual read on enterprise AI, released in late August 2026 with 1,719 professionals and business leaders answering across industries and regions. The Register covered it on August 25 under the headline that AI is finally 'on the road to ROI', then spent half the article pointing out that the earnings needle has not moved.

Both things are true, and the definitions matter. McKinsey counts an organization as an AI high performer when it attributes at least 5% of EBIT to AI use and calls the impact significant. Six percent of respondents clear that bar. 37% report any EBIT impact at all, a number The Register describes as a plateau from the previous survey. So the honest summary is this: adoption went up again, spending went up again, and the share of companies that can find AI in their earnings stayed flat.

Meanwhile 80% of people using AI report individual productivity gains. Hold those two numbers next to each other. Four out of five users say they personally got faster. Slightly more than a third of organizations can find any of that speed in the P&L. The ladder between those figures is where this whole story lives.

The mechanism is boring, which is why it survives. An AI tool lands on a team. People use it, and they genuinely get faster. Then the saved minutes scatter across the same handoffs and the same approval chain that existed before the tool. Nobody assigned the freed capacity anywhere, nobody set a number it was supposed to move, so it shows up as slack and the slack gets absorbed. The company paid for speed and bought comfort.

McKinsey's own data puts the fix in the org chart. In the previous edition of the survey, high performers were roughly three times as likely as everyone else to have fundamentally redesigned workflows around AI, three times as likely to report strong senior leadership ownership, and far likelier to aim AI at growth rather than shaving costs at the margin. Same models. Same vendors. Different org chart around the tool.

This is consistent with the harshest number in the field, MIT's finding that 95% of generative AI pilots show no measurable P&L impact. Different unit of measurement, same shape: pilots that were never wired to an earnings line cannot show up on one. We traced 62 figures on this question in our AI ROI benchmark, and the pattern held across every serious source. The technology clears the bar. The deployment trips over it.

The most decision-changing figure in the release got reported and then skipped past, and we found no coverage that worked out what it means. Nearly a third of respondents told McKinsey their organization decided against buying one or more software products and built the functionality in-house with coding agents instead.

Sit with that. The build-versus-buy line, the one that has defaulted to buy for two decades because engineering time was the scarcest thing in the building, just moved. When an agent can produce the internal tool in days, the calculation that justified a five-figure annual SaaS contract for a feature your team half-uses stops holding. A third of companies have already acted on that. Every software renewal conversation this year happens in the shadow of it, whether the vendor acknowledges it or not.

Two cautions before you swing the axe. First, the survey counts decisions, not outcomes: an agent-built tool still needs an owner and upkeep, and a rebuilt tool that quietly breaks in month four costs more than the subscription did. We wrote up what actually breaks in production automations from our own build logs, and none of it was the model. Second, running agents is not free: 20% of respondents say AI operating costs already constrain how much they use the technology. Building beats buying only when someone did the math on both.

The agent numbers moved faster than anything else in the survey. Among organizations with more than $1B in revenue, 40% are now scaling AI agents, up from 27% a year earlier. That is a 13-point jump in the segment with the most process debt, the most compliance review, and the slowest procurement on earth. Big companies do not move like that for a demo.

The workforce expectations shifted with it. 39% of respondents now expect AI-driven job cuts in the year ahead, up from 32%, while 43% still expect little or no change. Read those together with the scaling number and it looks like reallocation: the companies furthest along are redrawing roles around agents rather than around headcount.

If you run a mid-market company, the uncomfortable part is the clock. The enterprises are through the experimentation phase that most agent-piloting companies are still stuck in, and the readiness work they paid consultants millions to figure out is now written down in public surveys. The playbook is free. The window where following it is a competitive edge rather than table stakes is not open indefinitely.

Play it forward for your own company, in both directions.

If you copy the 6%: you pick one workflow, not eleven. You redesign it around the agent instead of bolting the agent onto the old shape, you put a senior name on the number it is supposed to move, and you define that number before anything ships. The productivity that used to evaporate lands somewhere you can point to: invoices posted per bookkeeper, tickets closed without escalation, proposals out the door per week. That is the version where a team of eight produces like a team of twenty and the delta shows up at EBIT, which is the only place anyone senior looks.

If you do what most of the 88% did: you buy seats, your people get faster, and the gains scatter. A year in, finance asks what the AI line item bought and the honest answer is vibes. Costs compound while you wait, since a fifth of companies already report operating cost as a constraint, and risk compounds too: in the prior survey more than half of organizations reported at least one negative consequence from AI use, with inaccuracy the most common. The high performers run defined human validation on agent output at roughly three times the rate of everyone else, which is exactly the discipline agent evals exist to make routine. Skipping it does not make you fast. It makes you the case study in next year's edition.

The spread between those two futures is not model quality. Both futures run on the same models. The spread is whether anyone did the unglamorous scoping work up front.

The survey's practical instruction is one sentence long: find where AI touches your earnings before you scale it, not after. That is the whole job of an AI opportunity assessment, and the 2026 numbers are the strongest argument yet for doing one on paper before doing it in production.

The first pass does not require a vendor. List the workflows where people repeat judgment on schedule: coding invoices, screening applicants, triaging tickets, assembling the Monday report. For each one write down four things: how many units a month, how rule-bound the middle of it is, how often exceptions land on a human today, and which P&L line moves if the volume doubles without new hires. Rank by that last answer. A workflow with no earnings line attached is a hobby, and McKinsey just published what a hobby portfolio looks like at enterprise scale.

Where we come in is the step after: pressure-testing the shortlist against what agents can reliably do this year and shipping the first system. That is what our 10x audit is, the scoring work with a build attached. A starter build runs $1,500 to $2,500 fixed, which is less than most companies spend per quarter on seats nobody measures. And if your processes are undocumented to the point where nobody can list the exception rate, do not hire us yet. Fix the documentation first. An assessment run on folklore produces a roadmap to nowhere, whoever sells it to you.