What will UK government workers do with an extra 26 minutes a day? The United Kingdom's Government Digital Service (GDS) has found that giving civil service employees access to Microsoft 365 Copilot saved them an average 26 minutes per day on office tasks. Microsoft 365 Copilot provides generative AI assistance with various Microsoft Office applications like Word, Excel, PowerPoint, Outlook, …
Not sure if I believe that.
Self-reported statistics are always going to be problematic as a data source.
Only a very small minority will bother reporting in the first place so don't think it can really be extrapolated against an entire organization.
I was using Copilot last year as part of a company pilot, perhaps I was holding it wrong, but didn't do much for me at all.
Let it do long mail trail summaries then verified a couple of them, had missed important points.
Handy for Teams summaries and what have I missed since I joined the meeting late type things but that's about it.
If you read the full report, you'll find support for your argument. Only about a third of the 20,000 staff completed the data collection. Like you I was part of a pilot (but to try Claude), and like you I found it didn't do much for me. But I'm a civil servant, and our prime minister has already nailed his colours enthusiastically to the AI mast, so the analysis has to support that.
Interestingly, the report was clear that by far the most frequent use of Copilot was for Teams summaries. Take that away, factor in the 1/3 response rate, and suddenly the maths looks far less compelling.
"if these 26 minutes a day are used for "strategic" activities I begin to understand the mess our government has been in for many years , and is still in."
The problem of government, and the reason it is in a mess in many countries is precisely because they don't have competently run strategy functions and staff. By career path I'm a business strategist, now a civil servant. The staff competencies of British civil service is organised into "professions", but there is no strategy profession. Statisticians, analysts, economists, occupational psychologists, risk profession, international trade negotiation, property, legal etc etc, but no strategy profession. And that shows. Whilst most departments have people with "strategy" job titles, the actual leadership of those departments mostly don't understand what strategy is (being from other professions), and are insufficiently curious about how the department might work better. It is not helped that politicians are for the most part even more clueless, and only think in terms of policies, and with a view to short term electoral calculus. Starmer and several of his cabinet are something of an exception, that he has defined the key missions of his government, and is not afraid of taking unpopular decisions in the longer term interest. But the British public with their goldfish attention span and hamster intellects won't care when it comes to election time, and there's still the problem that the administration of government lacks strategic steering. Nominally that sits with Cabinet Office, but (a) they've not grasped the key problem of a need for organisation strategy, and (b) they're first into the mangler of the government's decision to get rid of 10% or more of all civil servants. Seeing this pan out, it's rather like watching a manual worker who's been told to lose weight cutting off a few fingers to meet the goal.
The civil service is about 1.2% of total government spending. The real challenge of government efficiency should not have been saving 0.1% by the cuts to the branch of government that spends the other 98.8%. Look at the multi-billion cost overruns on every defence project, on HS2, on Hinkley Point and the rest of energy policy*, benefit fraud, on historic fuck ups from Horizon, BSE, infected blood, COVID fraud and waste, asylum processing delays and fraud, on planning delays and environmental red tape (Lower Thames crossing as but one example), just about every big government IT change (Defra payments, HMCTS, DWP), emergency services radio, the appalling deal to pay £3.5bn+ to hand over the Chagos Islands to Chinese aligned nation with no valid claim to them, etc etc etc
* Technically this isn't government spending, but every decision has been made by government and the costs added to your electricity bill
If we assume the time savings measurement is precise (and we probably shouldn't), it doesn't seem that much. But also, it doesn't necessarily mean an increase in productivity. Experiments on reducing work time, like 4-days week, often show counterintuitive results that the overall productivity does not change significantly; and people working twice more hours are not twice more productive either.
If you want to be pedantic, the precision was not given.
We only know 26 has two significant digits. Assuming it was rounded correctly[*], the last significant digit is the first digit of uncertainty (as a general rule, with some exceptions). So the value could be 26 ± 1 or 26 ± 8. Hard to tell. The best case scenario would be about 4% uncertainty; the worst case about 40%.
Comparing it to 8 hours does not say anything about the precision of the savings.
You are completely right about the accuracy though.
[*] Of course it was not – even people who should know better do it routinely wrong.
Like you I was assuming that the last digit was rounded to the nearest whole integer, so the true value could be anything between 25.5 and 26.5, ie a maximum error of 0.5 around a reported value of 26. That's how I got my 2% precision error (0.5/26 = 0.019).
I don't follow you when you said the true value could be 26+/- 8, how did you get +/-8?
This is a discussion I've been having as well - how do you quantify the saving?
I'm quite prepared to believe that users saved in the region of half an hour a day. Some of the documents which have to be produced by civil servants in all organisations are prime for automation of some kind in my opinion. However from an organisational point of view, how are you quantifying that saving? More importantly how do you bank that saving? The organisation I work for has 500 ish people in it, so the current yearly cost of copilot for all would be north of £100k. Taking average salaries, that's between 3 and 4 FTE employees. I can tell you right now I don't have an extra £100k floating around in my budget just waiting to be used, and it's a big enough number that getting sign off for it as an IT project will be next to impossible (we don't like spending money on IT here apparently), without the ability to demonstrate a monetary saving elsewhere.
So simply put, which 4 (or more) people are being let go from the organisation to pay for this? Or how else are you documenting any return on investment for this project?
"So simply put, which 4 (or more) people are being let go from the organisation to pay for this? Or how else are you documenting any return on investment for this project?"
It's been well publicised that government are arbitrarily getting rid of about 50,000 civil service jobs over the next four years, around 10% of the total. Invitations to take VR have already been handed out where I work. Lucky that the people who will be leaving either know nothing and do nothing, or that AI will take up the slack, isn't it?
If I were cynical, I might wonder whether this is simply a case of AI being used to generate greater volumes of the kind of corporate BS that blights the lives of the people who do the real work. I think readers of El Reg will know what I mean: emails, reports and presentations from HR, senior management and consultants, full of buzzwords but oddly devoid of any actual meaning. On the plus side, if I can toss these into Copilot and ask it for a one-line summary in language that a 7-year-old can understand, I'll save myself actually having to read all of the tedious crap and save myself a lot more than 26 minutes each day.
For #deity's sake, don't let Farage know. He'll only use that as cast-iron proof that public departments are over staffed!
Personally, I think they need to measure productivity not time-per-task. As pointed out above, there's actually money wasted if a task has to be redone if the 'help' turns out to have provided woefully inaccurate/wrong data.
And like that, with one twenty-one-word statement, Mr Anonymous Coward accurately summarised the output of slightly over half a million UK government employees, delicately capturing the nuances of their working lives as they grapple with directives from an ever changing host of ministers beholden to an even more fickle electorate. Thank you for letting us bask in the warm golden shower of your wisdom.
slightly over half a million UK government employees
And yet they still can't even answer their phones. HMRC callers spent 798 years on hold in FY 2023. Even if they spent every one of those 26 minutes saved helping a customer, it wouldn't make any difference at all.
Or, to use the number next to it, an average wait time of over 23 minutes. Which isn't good, certainly, and we should try to do something about that. However, it's not surprising that a department where people call with complicated questions and thus probably take a lot of the time of the people answering calls has high wait times. The way to speed that up is to have more customer service people or fewer callers, and since the set of potential callers are everyone who pays tax, that's not likely to be an easy option. Still, I've had a lot longer hold times with many places. I have a feeling that, if the average hold time was brought down to a number everyone would be happy with (three minutes work for everyone?), there'd be complaints about the people who were employed to answer calls but didn't have any and were just sitting around.
Don't worry when they do answer them you have to go through security.
I'm retired now but I was self employed. Once only one of the tax forms arrived so I phoned to ask for copies of the missing ones. Blank forms you understand. I had to go through security before I was allowed to ask my question - "can I have my forms please".
I can understand the need for security for some queries but this was silly, and a waste of two peoples time.
I have access to CoPilot and GPT. CoPilot is really awful in comparison and GPT isn't great.
Not sure about adding stardust but it can be used cobble together the bulk of something that you could do yourself but probably wouldn't bother about otherwise.
So a bit of value added but little that is actually insightful.
And you do need to go through everything the little bastard spits out as it likes to throw in the odd subtle bit of disinformation or complete bollocks when you're not looking closely at it.
I've used copilot to write the odd excel macro which is handy I suppose but not worth handing over a dossier on your entierly life to get.
Using Copilot takes a small amount of time, but they then have to do virtually the same amount of work that would be necessary to do it without Copilot, to be able to properly verify the output (which absolutely needs to be done because, as we are well aware, there are frequently major errors in the output from these things).
I would have thought it would actually increase the amount of time needed, not reduce it?. Or are they not doing the verification?
Think I know the answer to that one :/
Arguably the people who report the highest time-savings from AI are the first an organisation should be looking to get rid of - either they're not checking the outputs of the AI and consequently doing a bad job or they're getting perfect results but their job can be replaced entirely by the AI. Given how LLMs work, I would bet on the former in almost every case.
Acting as a filter for incompetence is one of the few ways AI might consistently improve efficiency.
“ Microsoft 365 Copilot provides generative AI assistance with various Microsoft Office applications like Word, Excel, PowerPoint, Outlook, and Teams. It allows workers to accomplish some tasks through a natural language chat interface instead of mouse movements and menu clicks.‘
I remember using natural language interfaces in the late 1990s for Windows, okay these used voice and so did not require using a keyboard and mouse/touch-screen to interact with a chatbot.
I suspect some one using Office 2000 on Windows 2000 with a “natural chat interface” of the time, would be more productive(*) with a lower energy footprint than someone doing the same task with W11, 365 and CoPilot.
(*) I appreciate the newer products have greater functionality etc., however, the real issue is UX/UI and bloat such as telemetry.
LOL anyone who has fought voice recognitian in a car will know for sure that moving the mouse and clicking on a link is far more efficient than repeating yourself 16 times to the machine even after spending a month learning the right sequence of words to make it do what you are actually asking.
AI is inherantly untrustworthy at the moment, "Do X" doesn't necessarily mean that the AI will "do X", it does some probablistic nonsense that means it picks the most likely option of what "do X" actually means. On a bad day you think "do X" means "turn on the kitchen lights", and the AI goes for "launch missiles". But it's ok because the supplier has put in a disclaimer of "sometimes AI makes mistakes" so shareholder value is maintained.
Exactly this. I have asked it to write some Home Assistant automations. Usually the yaml it spits out is obsolete, presumably because there are more examples of older HA syntax than the current stuff. If I point out that it has given me obsolete code then it will tell me that I'm right and produce something more current. My small number of experiences with it so far make me think of it as a glorified search engine. But as that, it sort of delivers if you know what you're looking for.
See this a lot with anything to do with resolving MS Windows and Office problems. There is so much stuff out there of varying vintages, that both Google’s and MS’s AI search will generate a solution that concatenates stuff from several versions, which naturally is of little help with a problem involving current versions.
I suspect it is similar with Linux and other tech products due to the number of distro’s and longevity of development. So for example, whilst Linux may be dropping support for older x86 CPU’s, I fully expect LLM AI’s to generate code that erroneously includes considerations only applicable to these earlier processors.
-- fought voice recognitian in a car --
me: SELECT TRACK
car: sorry steering command are not available
third attempt:
car: say name of track
me: landslide (just for those who don't know its Fleetwood Mac - and no I didn't say that to the car)
car: gives me a choice of several tracks none of which are landslide
If I want real fun I try "select artist" and "ygdrassil"
The study defines active Copilot users and adoption rate in the second paragraph of chapter "4. Results and findings":
>Active users of M365 Copilot are defined as users with at least one interaction with M365 Copilot in the previous 30 days.
>The adoption rate is defined as the ratio between active users and number of users with a M365 Copilot licence,
I.e. if Copilot manages to suggest a useful response to some chat message and you click "I'll look into it" instead of typing it once in a month you are a active Copilot user. Using these definitions the adoption rate is probably showing how many users with a license used Teams at all and not how many really use CoPilot in any meaningful way.
One interaction a month for the full Copilot is an appalling minimum to work from. That single interaction would have to save the whole $30 per month license cost which is unlikely.
Even once a week is low.
A lot of the interactions are stupid anyway. Copilot please populate this form for me, which will probably be most of the same entries every time. Here's a thought, do some better templates and just choose between them. Small upfront investment in time and more likely to be accurate.
What I'm seeing from teh company pilot are people very keen to evangelise the wonder that is AI and no one ever seems to publicly state that it did nothing for them. I assume no one wants to be tarred and feathered for pissing on the new corporate golden calf.
therefore 100% of comments about AI are positive, 100% of people save time using it and it is a wonderful lover and a kind grand parent and brought world peace to them
"thereby freeing public sector staff from laborious tasks through automation"
So a script could free them from the same 23 minutes per day work. I see no reason for anything with AI to be included. If people could save 23 minutes, it means that the work they were doing before was unoptimised. Fix that, and you get the 23 minutes AND save the AI costs.
Profit!
(Sorry, but that would mean Manglement allowing IT to actually talk to the people doing the work and to make the improvements they need, so we know that's never going to happen...)
This ^^^^^^
So many times I have seen complex process reporting or analysis done on bastardised spreadsheets instead of something written with a clear UI. One wonders where we’d be if organisations had invested in providing tools that fit the job instead of the other way around.
When everything looks like a nail, companies use hamm …. er spreadsheets!
>” I have seen complex process reporting or analysis done on bastardised spreadsheets instead of something written with a clear UI”
There are many times I have used Excel for accessing a DB, analysing data and creating a report; I have helped my users by giving it a clear UI and protecting cells and sheets.
I therefore conclude a “clear UI” does not mean there isn’t a “bastardised spreadsheet” behind it.
Sometimes you have to use the tools to hand, however, that doesn’t mean you forget to exercise good software design principles.
What was clawed back was the time wasted looking for where the options had moved that I needed to use. Or how the new formula format was breaking my workflow.
This is unpaid overtime that MS spent with frivolous u/I changes... But they are going to sell it as more extra work for the masses. Ugh.
Being cynical, this study has the fingerprints of a certain major system integrator’s Microsoft practise all over it.
I expect the licenses were provided gratuitous, to facilitate the signing off of a much enlarged 365 subscription, at a time when the UK government should be getting off a service ultimately under the control of the Weirdhouse…
These tools only increase productivity (assuming you want to keep accuracy the same) if you formulate it so that the cost of (accurate) verification is accounted for.
They're weak oracles, generating candidate solutions, at a cost c.
You then have to compute, over a day, for how many tasks was the verification cost _much_ smaller than the generation cost, and how much postprocessing was needed to make it acceptable.
There are plenty of tasks where that equation can work, but only if you're aware of that tradeoff.
This is especially the case because it's an online problem, you don't know per se if the next use will pay off, so you need to use your own intuition (maligned term for optimized heuristics based on experience) to figure out if and to what extent past was predictive for future gain.
That isn't helped by most of these tools being as volatile as smoke, where models (versions) are swapped without much control.
Some of these models have restricted use cases where, for the right discerning user, aware of accuracy and constraints, they can save significant productivity, especially on tasks that aren't in the class of essential but accidental complexity (e.g. figuring out what and how of solving, rather than the draining steps of navigating broken software or poorly sourced tools).
Doesn't matter if AI does save time or not: All it means is the worker will be expected to do more work elsewhere.
You don't 'get' the time back: You just free it up to take on other duties and do more work meaning you produce more for the company (public or private - it's the same principle) all for the same pay.
6,578 divided by 60 is 109.6. To convert that to "4.59 days" you have to be reckoning on 24 hours in a day. What kind of slave drivers are you?
I would assume the report is claiming a saving in *working* days. That's (109.6 / 8 = ) 13.7 working days per year, which matches Microsoft's figure, assuming an eight-hour workday.
Meeting summaries can be useful, although we're still in a world where people are checking the output - presumably once you rely on it the manual checking goes and we just have to pray the accuracy is better! Although how accurate are people taking minutes and notes anyway?
What is genuinely interesting (and frightening, depending on your perspective!) is that O365 and CoPilot are rolling out capabilities where usage data is being used to produce recommendations where AI could automate roles, aka O365 is going to be telling HR which jobs it thinks it could eliminate! It's going to be a supercharged version of the productivity monitoring of the past, and I'm sure there will be some trip ups along the way. "According to CoPilot you've been doing nothing all day" ... "But I've been slaving away all hours, using this non-Microsoft application!". There will no doubt be some interesting test cases of contract law if or rather when AI accuses you of being a lazy bugger!
I like how this comment directly appears above an ElReg survey:
1. How familiar are you with Cohere’s enterprise AI solutions?
It’s the next question that is so funny:
2. Which of the following brands do you most associate with Secure AI?
There is no option for “None of the above”.
I was one of the lucky few who got access to Copilot during the government trial. Well, "lucky" might not be the right word—because the day I was notified that my license would become active was the same day I went on two weeks of leave. So, right off the bat, half my trial period was gone. That left me with just over two weeks to explore Copilot, and in that time, I pushed it to its absolute limits.
Despite the short timeframe, I managed to develop multiple full syllabuses, complete with structured study guides, textbooks, and practice exams. I generated multiple versions of course materials for accessibility needs, ensuring inclusivity. I created content that would have taken a team of six people two years to complete manually. I also drafted complex queries for large-scale databases, ensuring long-term usability even after the trial ended. I enhanced meeting transcripts in Teams, which were far superior to standard auto-generated transcripts.
All of this work is still being used today. The syllabuses will likely be in place for the next decade, and the queries I generated continue to power critical databases long after my Copilot license was revoked.
One of the biggest unspoken challenges of the trial was that many departments had just transitioned from Windows 8.1 to Windows 10, alongside an upgrade to a new Office suite. This meant that before users could even begin experimenting with Copilot, they were already struggling to adapt to a new operating system and unfamiliar Office tools—without formal training. For many civil servants, this created a double adaptation burden: learning a new OS and Office suite while still managing their regular workload, and figuring out how to use Copilot effectively in an environment where many were still getting used to basic software changes.
What’s also missing from the report is the fact that licenses were only awarded for **30 days** before being taken away and reassigned to someone else. This meant that just as users were starting to get familiar with Copilot, they lost access, making it impossible to build long-term efficiencies.
Another overlooked factor is that many civil servants have only **basic ICT skills**. Even if they had the imagination to see how Copilot could be useful, they often lacked the technical ability to apply it effectively. Without structured training, many users defaulted to simple tasks rather than leveraging Copilot for deeper workflow improvements.
Perhaps the most significant missed opportunity was that **Copilot was not trained on any government data**. At the very least, it should have been trained on **publicly available documents and non-confidential data**—the kind of information that could be obtained through the **Freedom of Information Act** or simply scraped from government websites. Instead, it was deployed without any pre-training on even the most **mundane, widely accessible** government data. This severely limited its usefulness, as it could have been leveraged far more effectively if it had been optimized for the specific needs of civil servants.
Many critics assume that the saved 36 minutes per week would simply be used for tea breaks or idle time. In reality, for overworked civil servants, those minutes often meant completing tasks that otherwise wouldn’t get done due to workload constraints, reducing time spent on mundane admin work, freeing up capacity for higher-value strategic tasks, and improving accessibility, allowing users with disabilities to work more efficiently.
The real takeaway from the Copilot trial isn’t just the average time saved—it’s the potential for exponential efficiency gains when used correctly. The problem wasn’t Copilot itself—it was the lack of training, rushed implementation, and simultaneous system upgrades that prevented many users from fully leveraging its capabilities. For those who understood how to use it, the results were game-changing. Imagine what could have been achieved if all users had proper training, full access for the entire trial period, and a version of Copilot that had been **pre-trained on relevant government data**.
>” Despite the short timeframe, I managed to develop… I generated… I created…I also drafted…”
Sounds wonderful, just that being able to develop ie. Draft, something doesn’t mean the content is complete, coherent and relevant. That all needs review time.
Also, I doubt you could stand up and defend any of these products, eg. Why the syllabus was ordered/structured the way it is and the order of topics etc.
The reason some things take time is because, being human, you do much of the sense checking whilst you structure and then write the report; Copliot will do none of that, it meagrely is stringing symbols together based on the roll of dice. If the products of CoPilot are so immediately usable, we now have to question just what your expertise is and its value; in a world that believes the civil service is overmanned, it does seem your job is redundant, as it can be fulfilled by a CoPilot question creation expert.
Your last point “ I enhanced meeting transcripts in Teams…”
Is interesting, as to do this to any effect you will have had to actually attended the Teams meeting, made copious notes, so as to be able to compare old and new and then meaningfully enhance the Teams output. Now we are talking about transcripts ie. The aural version of OCR, a meeting transcript is only really needed after the event to cover your arse, what is actually needed are actual minutes: agreements, actions etc. So the question has to be what is the real value of transcripts over a simple recording. I suggest if you are having to revisit a meeting’s transcripts you probably weren’t paying attention (note we are talking about meetings not training session and workshops).
>” One of the biggest unspoken challenges…”
You following comments are really informative and good feedback.
Nice reply Boomer.
However, a team of experts confirmed everything content is complete, coherent and relevant.
Already being used in place.
i.e. is better than ie.
Coherency and correctness.
Another point, most meetings are not worth being involved in -- you simply do a two times or four times speed listen on the meeting -- look at the notes and everything else that came with it and then redirect everybody who is acting without purpose and efficiency.
No my job is not redundant it just freed me up of all the administrative tasks that can be done by AI and has made my job much more about what my job actually is and that is interacting with real people leaving computers and AI to act interact with themselves.
Thanks again
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
