Most AI adopters saw no change in labour costs
A New York Fed survey of 5,248 small employer firms puts AI use at 46 per cent while 77 per cent report no change in their labour costs, GitHub capped how many pull requests one contributor may open, and the Dutch regulator fined Uber EUR 825 million for deactivating drivers by automated decision.
How we Organise
The firms expect to hire, and their labour costs have not moved
The New York Fed asked 5,248 American small employer firms about AI last autumn. Forty-six per cent said the business or its staff were using AI tools, and a further 15 per cent planned to start within the year. Seventy-seven per cent reported no change in their labour costs from it. Thirty-one per cent reported higher sales. The same firms expect a great deal more than that. AI users ran a net 33 percentage points expecting to employ more people over the following twelve months, against 15 points among non-users. Ten points of the gap survives controls for the firm, the owner, the location and the previous year’s hiring.
We ask a finance team not to count a saving until the thing behind it has arrived. Our practice says it plainly:
No cost reduction is booked before the benefit it rests on has landed.
The usual objection is that the caution is theoretical, and this survey answers it from inside one population at one moment. It cannot tell you the forecast is wrong. It tells you the benefit has not arrived, which the authors say themselves.
How we Build
GitHub capped the queue instead of improving the filter
GitHub changed its pull request limits on 8 October so that drafts count towards them. Until this week a contributor could open any number of drafts even where a repository had capped the rest. The reason GitHub gives in its changelog is that maintainers are receiving more low-quality contributions than they can read. Counting drafts cuts the clutter, the notifications and the continuous-integration runs arriving with repository spam. What shipped is a ceiling on how many pull requests one person may hold open, applied before any review takes place.
We tell an engineering team to put the cheap checks first and let a human see only what an author has argued with:
Review is sized and routed so that a machine checks what it can before a person sees anything, including against the design principles the build must honour; findings reach a person only where the author disputes them, and the share of the machine’s findings the author accepts is measured, with findings from a check below the team’s threshold held back.
A month ago a study of 31,073 review-and-feedback pairs across 239 repositories found 56.3 per cent of a review agent’s comments rejected. That is why the closing clause is there at all. GitHub went the other way and capped admission instead of improving the filter. Its changelog gives no figure for how much spam the limit stops, and a public repository is not a firm’s own team, where you know who may contribute. If you have capped submissions rather than filtered them, we would like to hear what it cost: transform@dromologue.ai.
How we Assure
Eight hundred and twenty-five million for a decision no person took
The Dutch data protection authority fined Uber EUR 824,990,000 on 8 October. Between 2018 and 2022 Uber deactivated drivers’ accounts by fully automated decision: temporarily where it suspected fraud or where ratings were too low, permanently where low ratings persisted. Drivers lost their income from the platform while deactivated. The case began with complaints from 171 French drivers and reached the Dutch authority because Uber’s European headquarters are there. The authority found Uber had broken the prohibition on fully automated decision-making and had not adequately told drivers it was happening.
With a client we hold that an automated decision needs somebody able to stop it:
Every agent and every automated decision has an identified person who can stop it.
Nobody asks for that clause while the system is being designed, which is the difficulty with it, and there is now a figure to set beside the omission. Note what the ruling does not rest on. The authority decided the prohibition and the information duty, not whether any particular deactivation was wrong. Uber has appealed and no court has ruled. A firm outside the European Union faces no equivalent rule.