This is a PR puff piece dressed up as “research”.

First, 19 minutes a day is not productivity. It is noise. You could claw back more than that by letting people start at 10, avoiding peak commuting, or cancelling one standing meeting about another meeting. No AI required. Just basic adult management.

Second, the numbers are already wobbling. One government study says 26 minutes, another says 19, another says zero. That is not convergence. That is what happens when you measure vibes instead of outcomes. The only consistent result is “people felt nicer”, which is not the same thing as doing more useful work.

Third, the sample is cooked. Volunteers and nominees using shiny new tools always report gains. It is the Hawthorne effect with autocomplete. Statistical “adjustments” are not magic. They do not turn self-selection into causality.

Fourth, the tasks where Copilot “saves time” are exactly the ones civil servants should not be optimising blindly. Searching internal docs faster just means propagating outdated nonsense quicker. First-draft emails were never the bottleneck. Judgement, accountability, and decision-making were. Even the article admits the tool falls down the moment human judgement is required. That is the actual job.

Fifth, the cost-benefit is missing entirely. We are told minutes were saved, but not how much was paid per minute to a foreign, tax-optimised corporation. Billions leave the UK so staff can shave a few minutes off emails that probably should not exist in the first place. If the same department banned “reply all” and status updates, the gains would dwarf Copilot overnight.

Sixth, the risk is waved away. Every minute saved now comes with future minutes lost proving that sensitive government data was not inspected, retained, or subpoenaed under the US Cloud Act. That paperwork alone will eat the alleged gains for breakfast.

Finally, “comfort blanket” is the most honest phrase in the article, and not in the way it intends. Reducing stress by outsourcing thinking to a probabilistic text generator is not efficiency. It is sedation.

If Department for Work and Pensions wanted real productivity, it would fix workflows, kill pointless meetings, and stop confusing busyness with output. Instead, it bought licences from Microsoft, got a modest placebo effect, and called it transformation.

I'm sure it also "saved" West Midlands Police time in creating their evidence against Maccabi Tel Aviv ...right up to point it hallucinated about the West Ham match that had never happened.

Nothing in this "research" actually checked the integrity of any work that was produced, the users were selected based on volunteers, not on a random sample of workers, and the "results" were self-assessed by the users.

Rubbish in, rubbish out...

And in functioning country these things would have warranted an investigation, not a slide deck and pats on the back.

A lot of DWP processes, at least in my experience, seems to be about deterrence - making it so laborious to claim for various things that people who might otherwise be entitled give up the fight.

Finally, a purpose for which AI seems ideally suited.

I get baffled whenever "AI" is touted as being used to draft documents. Surely your staff should just be able to type documents themselves? Or has the general standard of written English degraded that far?

Yes. Yes it has.

This is exactly what happens in practice.

You write a few clear paragraphs.

AI pads them out with fluff to make them look “substantial”.

The recipient asks AI to strip the fluff back out again.

Then they reply by repeating the same process in reverse.

Net result: two humans doing the same thinking as before, plus a machine adding and removing noise in between.

Nothing is gained except big foreign tax-shy corporation gets paid.

Voicing a view from inside the Civil Service, I'm amazed how many emails I get that start off with that horrid, insincere AI giveaway "I hope this email finds you well....". Invariably from people who in person are intelligent, functioning adults well able to make requests or hold conversations without the broken crutch that is AI. These will be some of the people insisting that Copilot saved them n minutes per day.

Maybe I should conjure up an AI reply with the prompt "reply, explaining that the email didn't find me well, my sensitive intellectual constitution has been mortally wounded by the insincerity of a machine generated salutation that isn't even logically correct, because an email is not capable of finding or doing anything (other than perhaps causing offence)....etc etc."

...were spent working that out. Or did they employ AI to do it?

It's more likely to be that there was budget left over that needed to be used up before the end of the financial year... "we need a success story"

It's more likely to be that there was budget left over that needed to be used up before the end of the financial year... "we need a success story"

Civil servant here: Nothing of the sort. We're under a certain amount of gentle yet persistent pressure to cram AI into everything, because Starmer and his cabinet have picked AI as a magic bean to transform the lives of the plebs*. As civil servants, we're under an obligation** not to judge or obstruct government policy, but to enable and administer it, so whilst most colleagues agree AI is a piece of crap that adds nothing, our professional duty is to make full use of it, and then report back to justify what a boon it is.

* You might need to look up the exact words from his speeches or publications issued in his name; He probably only thought the word "plebs", and more likely said "citizens".

** The Civil Service Code. This is part of our contracts of employment.

I believe it took 19min/user/day to calculate.

I've spent more time in the Think Chamber than this saving - and I dare say it was a lot more productive in multiple ways

> Microsoft Copilot saved civil servants 19 minutes daily

But what is that "saved" time used for?

Is it needed to navigate / wrestle / wait for Copilot to produce something usable. Time spent checking what spurts forth. Chatting on social media. Attending more useless meetings or (just possibly) producing 19% of extra value - however that intangible metric can be measured?