Forgot your password?
typodupeerror
AI Microsoft

NHS Prescribes Half a Million Copilot Licenses For Its Paperwork Headache (theregister.com) 55

NHS England plans to roll out Microsoft Copilot to 505,000 clinicians and support staff after a 30,000-person pilot claimed the AI assistant saved users an average of 43 minutes a day on administrative work. The Register reports: The rollout won't happen overnight. NHS England said that each trust will receive a central allocation of licenses based on headcount, typically starting with around 2,000 Copilot seats, and that more than half a million staff are expected to have access by October 2026. The NHS has no shortage of administrative work to throw at the software. The rollout envisions Copilot helping with discharge paperwork, bed management, rota planning, meeting minutes, board papers, briefings, data analysis, and assorted HR, finance, and procurement tasks.

NHS organizations will also receive access to Copilot Studio, Microsoft's toolkit for building custom AI agents. NHS England said trusts will be able to develop agents for tasks such as handling Freedom of Information requests, processing complaints, reducing helpdesk workloads, and assisting with financial analysis. A governance framework called Agent 365 will oversee the deployment of those systems.

This discussion has been archived. No new comments can be posted.

NHS Prescribes Half a Million Copilot Licenses For Its Paperwork Headache

Comments Filter:
  • by RitchCraft ( 6454710 ) on Tuesday June 09, 2026 @12:04PM (#66182384)

    Prescribe me the best drugs!

    • Prescribe me the best drugs!

      You may actually get the best drugs. Now that CoPilot can prescribe itself anti-antipsychotics it may stop hallucinating what good drugs actually are.

      • by mjwx ( 966435 )

        Prescribe me the best drugs!

        You may actually get the best drugs. Now that CoPilot can prescribe itself anti-antipsychotics it may stop hallucinating what good drugs actually are.

        The UK is one of the few places where you can still buy small amounts of codeine in over the counter painkillers (less than 13mg from memory). This is one of the reasons we're not having the same levels of prescription pain killer prescriptions as the US or Australia.

        Australia banned it a bit over 10 years ago and everyone suddenly went doctor shopping for pain killers, now Australians are using strong painkillers as recreational drugs (A.K.A. Hillbilly Heroin) ... quelle surprise!

        It's the same story

        • by BranMan ( 29917 )

          Oh, I miss access to pseudoephedrine - that used to be my "nuclear option" for tackling my occasional acute but pretty severe allergy attacks. Take that, crash (couldn't do much else - it really knocked me out) and wake up the next morning more or less functional again.

          Now it's 2-3 days of total misery, stretched out to a week of not being able to sleep well. None of the current stable of antihistamines does a damn thing for me.

  • Uh huh. (Score:5, Interesting)

    by msauve ( 701917 ) on Tuesday June 09, 2026 @12:05PM (#66182388)
    >saved users an average of 43 minutes a day on administrative work.

    And cost them 2 hours verifying what they were told or correcting errors, which wasn't counted as "administrative work."
    • by vyvepe ( 809573 )
      The danger starts when people notice that the AI was already right about 30 times so they stop to check for errors.
      • Re:Uh huh. (Score:4, Informative)

        by cmdr_klarg ( 629569 ) on Tuesday June 09, 2026 @02:10PM (#66182634)

        The danger starts when people notice that the AI was already right about 30 times so they stop to check for errors.

        I'm assuming just a grammatical error: "they stop to check for errors" is a good thing, as they are stopping what they are doing to check for errors. If phrased as "they stop checking for errors" then that will reflect the meaning you want.

      • by AmiMoJo ( 196126 )

        The question is if Copilot makes fewer errors than humans do.

        • by vyvepe ( 809573 )

          I think Copilot does considerably more errors in the area of software development than trained humans. If Copilot would make less errors then why would developers need to review Copilot output? One could answer this question by: "To improve it even more." But if that would be the case then why just not use Copilot to do the review as well. Allegedly it does better than human. Since it is a review (different context) and due to sampling temperature the internal model "reasoning" would not be the same as duri

  • Copilot? Might as well just use Clippy, cheaper and better than Crapilot.
  • by vyvepe ( 809573 ) on Tuesday June 09, 2026 @12:16PM (#66182424)
    AI told us to cut off the left leg.
    Who was supposed to know it should have been the right leg?
    • Re:Not our mistake (Score:4, Insightful)

      by smooth wombat ( 796938 ) on Tuesday June 09, 2026 @12:43PM (#66182494) Journal

      AI told us to cut off the left leg.
      Who was supposed to know it should have been the right leg?

      Still better than removing a liver rather than a spleen [usatoday.com].

      • by mjwx ( 966435 )

        AI told us to cut off the left leg.
        Who was supposed to know it should have been the right leg?

        Still better than removing a liver rather than a spleen [usatoday.com].

        To be fair, that was an easy mistake for a Florida doctor to make as the average Floridian has no heart, no spine, no brain and their heads and arses are interchangeable.

    • So funny story about this. My friend currently has a broken leg. We often send silly joke AI generated images to each other, so to cheer her up I fired up Nano Banana Pro with the prompt: "Draw a picture of this person laying on the hospital bed with her left leg in a cast. She is wearing a hospital gown. The left leg is elevated. She is surrounded by racoons [her favourite animal], and one is giving her a cuddle to make her feel better. The scene is in a recovery room, well lit with light coming through th

      • I have had 4 surgeries in the last 12 years. During preparation for each, I was pleased that staff, while looking at paper records, asked me to recite what was to be done, and at what location on my body, before marking either me or a tag.
        • It seems silly but this is all because mistakes happen. I can't go an donate blood without being asked for my date of birth 3 times, once by the receptionist giving me the forms to fill out, one by the doctor prep pepping the conditions of the donation, and one by the nurse reading the chart.

          The friend cited above apparently also had the doctor ask her to confirm which leg needed surgery which is fun given one is normal and the other was blue with bruising, massively swollen, and resting on the remains of t

  • Oh dear (Score:4, Interesting)

    by fleeped ( 1945926 ) on Tuesday June 09, 2026 @12:17PM (#66182430)

    That's where the taxpayer money will go? To even less interaction with humans, and even more money to Microsoft and its unreliable shitty software?

    • Less administration means more time to take care of the patients?
      • I can assure you that any cost savings will not be used to hire more personnel.

        • One would certainly hope not. Too large a bureaucracy is the reason they're wasting so much time on administrative tasks.
      • Re: Oh dear (Score:4, Insightful)

        by fleeped ( 1945926 ) on Tuesday June 09, 2026 @02:05PM (#66182628)

        There are precise automation systems that can be developed to do the admin work. Not non-deterministic agents. If the underlying systems are an archaic clusterfuck already (and afaik it is), unleashing the unreliable agents on it is going to be ... double plus shit?

        • by jezwel ( 2451108 )
          This is it exactly - the higher ups are too lazy/unqualified to identify and resolve business process automation so are counting on pointing the finger at CoPilot when things don't improve enough. I went from a team of 5+ to 3 (minimum allowed) after working through this in our group, and that started out as interpreting what people had free-text written on a printed-out form, and while at it also saved hundreds of FTE hours monthly further down the line in automated service allocation and provision.
        • The precise automation systems depend on humans doing their job correctly.

          This is a signal that they will be underpaying these employees even worse than they are now. They are expecting churn and turnover so they don't have to pay people at salaries earned through years of service.

          The AI solution won't need the data to be perfect going in. Of course, the data coming out also won't be deterministic, as you say. It'll be GIGO as per usual.

    • Britain's population increased from 56 million to about 60 million between 1970 and 2005, at a rate of about 114,000 per year. It increased from 60 million to 68 million between 2005 and 2020. at a rate of 500,000 per year. That might have something to do with NHS staffing and budgetary issues.
      • And you think AI agents are the budget-conscious answer? I'd love to see the budget allocated to (or, required by) MS AI, and more importantly, how it evolves in the following year, with the IPO races doing the rounds now. It would be a shame if AI becomes more expensive and then it's too late to get rid of it due to sunken costs, right?

        • by JBMcB ( 73720 )
          I didn't talk about AI at all. I'm talking about the underlying issues potentially causing the NHS's capacity issues. However, I see an issue in your reasoning:

          It would be a shame if AI becomes more expensive and then it's too late to get rid of it due to sunken costs, right?

          What is the sunk cost? You rent AI services. There is a nominal cost in integrating it into some workflow, but the main cost are purchasing the service as needed.

          • "Purchase the service as needed" - do you expect needs to go up or down with time? Also, do you expect AI costs to go up or down with time? And it's not exactly AWS compute that they're booking, in terms of elastic scaling. And I'm sure you're well aware that when institutions or organisations build/plan infrastructure around MS, it's really, really hard and costly to get off it. Or you think they can just replace the agents with something else that's magically cheaper?

      • by Xarius ( 691264 )

        45+ years of neoliberalism, privatisation, mostly-Tory and right-wing labour governments deliberately starving the NHS of resources and selling it off for parts for a quick buck have caused the problems. Not an increase in the population. This is a common strategy, deliberately starve public services of resources then claim that the "state is failing" and private companies with their profit-greedy owners are asked to step in to "do it better"

        An increase in the population naturally gives you more bloody work

      • by AmiMoJo ( 196126 )

        Not really. The people who came tended to be young and healthy, and many of them worked in the NHS.

        The problem is that out population is ageing, and they expect expensive, long term treatments and care that were seemingly not anticipated when they were paying National Insurance (the income tax that nominally pays for pensions and free healthcare in the UK). Not nearly enough was saved for this eventuality, not nearly enough was invested. Opportunities to take care of it were squandered.

    • by evanh ( 627108 )

      And so much for the shift away from dependence on US corporations. Doesn't seem that stated intent is having an effect.

    • So... everyone talks about vibe coding these days... yet nobody within the administration comes along and says "oh, instead of paying MS for 500k licenses... how about we just vibe code something similar with an open local model"...

      (e.g. do they *really* need the latest frontier model for routine administrative tasks... or would a 7gig local model do just as well?).

  • So now NHS has nearly half a million headaches. I'm not sure this is an improvement.

  • by spacepimp ( 664856 ) on Tuesday June 09, 2026 @12:21PM (#66182454)

    The alarming rate at which 505k employees burn through the annual token allotment with little to nothing to show for it except some super wasteful reports that are likely cheaper to produce in meat space will be the topic of a future article posted here in Early 2027.

    • I'm just spitballing here, but couldn't a more local LM be setup and trained on a more narrow basis? We're not talking audio/video generative AI but more along the lines of text generation with some sorting and organizing of files.

      The doctor will still have to of course go over the output, but that might be faster then some of these people's input skills. They are doctors, not computer nerds. I've seen many hunt and peck on the keyboard. It would be a whole lot easier if a local computer station recorded th

  • by Junta ( 36770 ) on Tuesday June 09, 2026 @12:47PM (#66182498)

    Not specific to AI, and I frankly can't speak to NHS specifically, but it sounds awfully familiar...

    So many things where bureaucratic junk demands awkward forms and processes, and efforts to automate all that stuff instead of streamlining the underlying mess...

    To the extent this works (and I can believe it based on other bureaucracies I've been involved with), it's because there's all sorts of dumb boilerplate crap in the process, lots of material generated that is never read, lots of fields to populate that don't matter to anyone. To the extent it ever matters that goes away as the people just stuff meaningless crap in those fields...

    The human is still having to provide the crux of the important bit, but there's just so much fluff that is blatantly obvious that LLM can do whatever with that could have been omitted or dealt with better.

    • by jezwel ( 2451108 )

      So many things where bureaucratic junk demands awkward forms and processes, and efforts to automate all that stuff instead of streamlining the underlying mess...

      ... it's because there's all sorts of dumb boilerplate crap in the process, lots of material generated that is never read, lots of fields to populate that don't matter to anyone. To the extent it ever matters that goes away as the people just stuff meaningless crap in those fields...

      The human is still having to provide the crux of the important bit, but there's just so much fluff that is blatantly obvious that LLM can do whatever with that could have been omitted or dealt with better.

      Concur. Some examples
      - you need approval from finance to spend $$$. Solution - build that approval into the requesting process - that way you only get requests that have financial approval already. Bonus - cost centre is included and can be charged automatically.
      - people request things for other people in the Justification field (which almost never had anything useful in it anyway). Solution - add field for who gets the service, remove the Justification field (they can sort that out with their manager of

  • by pesho ( 843750 ) on Tuesday June 09, 2026 @12:49PM (#66182502)
    I am sure saving 43 minutes a day as determined by a Linked-in post will be a great benefit to everyone's health. I hope they checked the accuracy of the AI output. Because it is not like UK will ever prosecute and convict innocent employees based on the inaccurate output of a computer system [wikipedia.org]. That, would never happen, right?
  • saves 43 minutes until the lawsuits start happening. I am sure lawyers going to love this non human suggestions. Garbage in and Garbage out.
  • Using the AI to help with analysis or data/literature review is really very good. But I'm surprised how inept they can be in their own use case

    Siri - "I don't want xxx (or I want yyy) on my phone, how do I do that" - I've found Siri to be useless for this
    Co-pilot - "how do I fix xxx or not have it do yyy on my doc/excel/ppt" - disappointing, at best you get a complicated list of things to try.
    Gemma for Mac (locally) - "how do I make backups of chats" says that I can't do that, need to use Gemma in a
  • A lot of paper-work is poorly factored, requesting the same info in multiple places rather than having ways to cross-reference. Instead of factoring the system, they are automating the repetition. This may just encourage yet more redundancy.

  • We. regret to inform you Mr Smith - you have ovarian cancer.
  • "We're wasting too much time on paperwork, so let's find a way to do paperwork more quickly" sounds sensible on the surface, but when you recognize that time is just one resource, that translates "We're wasting too many resources on paperwork on resources, so let's spend more resources into spending resources on paperwork more quickly."
  • An audit here in Ontario, CA, found that AI Scribe produced Hallucinations (9 out of 20 vendors), incorrect information (12 out of 20 vendors), and incomplete information (6 out of 20 vendors). "For example, the submitted notes included statements that there were "no masses found" or that there was presence of anxiety in the patient, although this information was not discussed in the recordings." The whole report is interesting: https://www.auditor.on.ca/en/c... [auditor.on.ca]
  • and in the usa cutting health insurance paper work BS will save an lot more time

  • I foresee a lot of real cases of this mock Onion obituary. [theonion.com]

    Diana Yanko, 61, died on Tuesday after an AI incorrectly filed her charts, another AI denied her claim, and a third AI turned off her life support.

The opulence of the front office door varies inversely with the fundamental solvency of the firm.

Working...