Data Divers Dialogs · Analysis · approx. 8 min read
A CFO spends August explaining an awkward number. Three weeks earlier the ECB measured the same thing — with nothing to sell. What both of them describe, nobody measures.
On 12 August, Beth Gaspich had to explain a number that did not fit the rest of the quarter. The CFO of NICE had just reported record cloud bookings, $362 million of AI revenue base, an AI backlog growing at 72 percent. And in the middle of it, cloud net revenue retention sat at 106 percent. Existing customers were paying barely more than the year before.
Her explanation was unremarkable, and therefore interesting. The company was winning larger enterprise AI deals, she said, and many of those customers were still at the very beginning of deployment.
Translated: they are buying faster than they can install it.
That is, of course, the most comfortable explanation available to a finance chief in this position. Considerably nicer than "our product doesn't do what we said it would." You would not have to believe her. Except that three weeks earlier, somebody with nothing to sell had measured the same thing.
For its Occasional Paper 395, the European Central Bank surveyed around 6,000 companies across twelve euro-area countries. Roughly 70 percent use AI in some form, 30 percent not at all. It only gets interesting when you sort by size.
On "some use", large firms come in at 87 percent, micro firms at 60. A wide gap, expected, boring. On the share using it intensively, every single size band lands at seven to eight percent.
Bars: any use. Line: intensive use. Values for small and medium are interpolated between the reported endpoints; the flat line at seven to eight percent is documented across every size band.
Seven percent at twenty employees. Seven percent at forty thousand. The number is so unmoved by company size that on first reading you take it for a typesetting error.
Large firms are broader. They are not deeper.
That flat line is uncomfortable, because it rules out, one by one, everything you would normally blame. Capital? Large firms have more of it. Access to the technology? Same API for everyone. Data assets, specialist staff, consulting budgets — all things where size helps, and evidently none of them decisive, or the line would not look like a ruler.
Whatever separates the intensive seven percent from everyone else, it is not scale, capital or access. Those are precisely the three things European policy is currently subsidising.
Process quality drops out for the same reason. Corporates have the more considered, better documented, more audited procedures. If good process were the lever, they would be ahead. They are not ahead.
What remains is a single variable that has nothing to do with company size: how fast an organisation can change a process. Not how good it is. How quickly it becomes a different one.
And that does not depend on money. It depends on how many people have to agree, on their calendars, and on their willingness to sign for a risk. Large organisations are structurally slower at this. It is roughly the only thing at which size is a handicap.
HubSpot reported on 5 August net revenue retention of 102 percent and roughly 7,000 new customers — internally it had expected 9,000 to 10,000. In the same report: its own support agent resolves 72 percent of tickets without a human taking over.
monday.com on 10 August: 109 percent, the lowest in the company's history, with AI revenue doubling. Freshworks on 4 August: 104 percent. And NICE, where we came in: 106.
Four vendors, four segments, one week. Everywhere the AI business accelerates, everywhere retention gives way. For a software company that is the least comfortable combination there is, because it rules out the one explanation you would most like to have: weak demand.
They are buying. They are not installing.
DX analyses engineering data from more than 500 organisations. In the second quarter its Developer Experience Index fell from 67 to 65 — for the first time ever. In the same quarter, reported time saved through AI rose from 3.3 to 6.1 hours per week.
Both at once. Individuals got nearly twice as fast, and the organisation felt worse than before.
Two more numbers from the same dataset explain why. The median pull request grew from 42 to 72 lines. And in a subset of organisations the change failure rate rose by three percentage points.
Because work is not finished when it has been produced. It is finished when it has been reviewed, tested, approved and documented — and those stations are staffed by people whose numbers have not changed. When the pull request doubles in size, what doubles is not output but the review load per item. What comes out the other end is queue.
It appears in no AI metric, because it does not form inside the AI.
A survey of more than 170 Fortune 1000 executives by Atlassian's Teamwork Lab fits: six percent could name a specific, organisation-wide return. Not because nothing happened. Because what happened accrued individually and drained away organisationally.
The constraint is not intelligence per token. It is the time a change takes to reach a signature.
At the end of June, Capgemini employed 417,600 people — 68,200 more than a year earlier, a fifth more, with the full-year outlook raised. India's five largest IT services firms cut a net 7,389 roles over the same period, after adding 12,718 the year before. On record revenues. TCS shed 12,200 managers while training 114,000 employees in AI.
Caveat: Capgemini's increase includes the integration of Intelligent Business Operations; the organic share has not been broken out. The Indian reduction is shaped by pyramid restructuring, attrition and visa policy.
Both camps have access to the same models. The dividing line runs between firms that rent out capacity and firms that sell change — and AI collapsed the price of the first good. Capgemini's chief executive Aiman Ezzat summed up the client side in early August: it is not as simple as it was made out to be at the start.
In regulated environments this tips into the absurd. That is where the density of approval points is highest: every validation-relevant change needs a documented, traceable, signed artefact. This is not a flaw in the system, it is the system. It exists so that changes do not happen quickly. Anyone serious about patient safety builds friction in on purpose.
The observation stands nonetheless: exactly where the most compute per head would be available, the least changes. The bottleneck is not a system. It is a person with the qualification, the authority and the willingness to carry a risk.
In practice that is usually one individual — a quality lead or a process owner, with a calendar booked three weeks out and the prospect of personally answering for that signature in an audit. Until it is there, the change does not exist. Even if it has been technically finished for a month.
A side observation, offered as observation rather than thesis: the timelines are stacked with alarm posts. Jobs gone, industry gone, consulting gone, tomorrow the next tipping point. It gets tiresome — you could build something in the time. And anyone building something rarely runs into the limits of a model and almost always into an approval date.
Plenty. Net revenue retention is a lagging metric; it reflects renewals negotiated one to two years ago. Part of the compression is probably plain small-business churn, dressed after the fact in an AI narrative. And if an agent resolves 72 percent of tickets on its own, the customer needs fewer licences — falling retention would then be a sign of successful deployment, with a pricing problem as the consequence.
Then a counter-example from the same week. Atlassian reported 28 percent revenue growth, retention above 120 percent and a backlog up 44 percent. No compression anywhere. Anyone wanting to explain why has several plausible candidates and no evidence — and should say so.
The most serious objection concerns the core. The ECB broke its data down by company size, not by change velocity. No survey does the latter. So we are inferring from an absence: because every size-dependent explanation fails, the cause must be size-independent — and change velocity is the most plausible size-independent variable anyone can name. Plausible is not proven. It could equally be the business model, the industry, or particular individuals.
Two weeks are also a snapshot, not a trend. With ERP and with cloud, the interim numbers looked like failure for years, right up until they didn't. Salesforce reports on 26 August; the next ECB round follows in the autumn. Until then nobody knows.
For workforce planning this produces an uncomfortable calculation. The scarce resource is not the person who operates the AI but the one allowed to take responsibility for its output. TCS trained 114,000 employees in AI. That creates not a single additional approver.
And a detail that disturbs the planning. An analysis of ChatGPT Enterprise across more than 1,500 organisations puts the heaviest use among early-career staff — whose output is reviewed by precisely the experienced ones. So cutting the entry rung does not relieve the bottleneck. It moves it up a floor and drains the pipeline to exactly the role that is scarce. Approval capability is built from domain judgement, institutional knowledge and trust. It is the one thing in an organisation you cannot order in.
Nowhere.
There is a benchmark for almost everything. For customer satisfaction, for data-centre energy efficiency, for how long a ticket spends in first-line support. What gets measured is revenue per employee, utilisation, licences deployed, pilots launched, hours saved per head.
None of those numbers says how long a company takes to do something differently than before.
That is not negligence but inheritance. Every one of those metrics comes from a time when capacity was scarce and change was rare. Both have reversed; the measurement system has not. And what is not measured is not managed — it is bought.
None of this would need a new system:
There is good news in that, if you care to read it that way. Without domain judgement, authority and a signature, none of what the models produce all day makes it into production.
AI needs us — and it knows it.
Which leaves the last of the four numbers. No new system, no licence, no project: you would only have to count what was actually done differently last quarter than the quarter before. An afternoon's work, in any company.
In most of them, nobody knows it.
On the numbers. Survey, financial and headcount figures come from the primary sources linked below. That change velocity is what sits behind them is an explanation by elimination, not a demonstrated cause — the ECB surveyed by size, not by speed. DX time savings are self-reported; pull-request sizes and failure rates are measured.
Sources: ECB Occasional Paper 395 · HubSpot Q2 2026 · monday.com Q2 2026 · Freshworks Q2 2026 · Capgemini H1 2026 · diginomica on DX/Atlassian · Atlassian Q4 FY26