A thousand staff hours a week from two agents, fifty more promised at the drive-thru
ASOS breaks down what its agents return in hours and pounds, Grab turns each production escape into a check that runs afterwards, and McDonald's own operators dispute the labour saving behind its ordering system.
How we Organise
ASOS can say what each of its agents gave back, in hours and in pounds
On this week’s earnings call, ASOS’s chief executive itemised what the retailer’s agentic rewiring has returned so far. Matching invoices to orders, work previously done by hand one line at a time, now saves about £1 million a year across the buying operation. A purchase-order agent and a catalogue agent together return close to 1,000 of those staff hours every week: 500 from the first, 300 from the second. An AI design studio produces more than 3,000 products for the site each week. An AI stylist logged more than 100,000 customer conversations in its first few weeks. The same call disclosed that the October breach, a push notification an unauthorised party sent through social engineering, cost between 2 and 4 per cent of gross merchandise value in the affected weeks.
Our finance practice asks that each release’s value be measured against its cost a month later, inside one record. ASOS does exactly that on each deployment, reading the savings out of the systems that the work runs in. It reports those figures to investors on the earnings call. It is not the first evidence against the statement. A measured account last month argued the wait itself may mislead, because the price of the same capability moves by close to a fifth inside the window. Both readings may stand. The figures remain the chief executive’s assertions on an earnings call. Hours saved cannot be checked against hours paid from outside the firm.
One month after a release, its value is read and set against its cost in the same record, and the cost carries the date and the price basis it was struck at.
How we Build
Grab treats every production crash as a case its checks must now catch
Grab’s mobile teams describe absorbing a surge of AI-generated code changes behind a layered crash defence. Static analysis runs before merge, with lint catching issues before the code lands. Early builds go only to internal users, who are monitored for crashes. Releases then roll out in stages, behind flags that trip when the crash rates pass set thresholds. Post-release triage maps each crash back to the change that introduced it. The loop that closes the account is the correction itself: post-mortem fixes become new lint rules and new test coverage. Each escape raises the floor for the next release.
Our engineering practice expects a reviewer’s correction to end as a case that keeps running in the test suite. Grab’s defence appears to be that sentence turned into machinery. The correction becomes a check that runs on every later build. A firm reading this and finding no such loop in its own pipeline has the gap measured for it. One caution applies: this is Grab’s own account. It publishes no crash rate before or after. The claim rests on the mechanism described, not on a verified improvement.
The case a reviewer’s correction became is a case the suite still runs.
How we Assure
McDonald’s scales ordering by voice while its own operators call the labour figure theoretical
McDonald’s will put its Archy voice-ordering system into most of its US restaurants by the end of 2027, with pilots abroad through the year. The company reports order accuracy above 90 per cent in English and Spanish, high customer satisfaction in early tests, and at least 50 labour hours a week freed per restaurant. In October, an analyst at BTIG reported that operators had been told little about specific productivity gains. He quoted one operator calling the target “highly theoretical”.
Our ethics position is that promises made to people when machines absorb their work are written down and reported on. The people whose hours this changes work for independent operators. A promise kept fully inside the role register, written down and dated exactly as the practice asks, can still leave the displaced workers uncounted. Nothing in the account says the company has made no commitments to its own employees. What the account shows is a mechanism reaching workers that the firm’s own record does not cover. A firm keeping this promise for workers outside its own payroll has a view we would like to hear: transform@dromologue.ai.
What was promised to the people whose work the machine changed is recorded against the role and reported against, and the people who never appear as claimants are looked for.