Personal AI assistants will become normal in UK workplaces

Britain has already moved beyond the demonstration

Lloyds Banking Group offers a concrete indication of what routine use can look like. In October 2025, it reported that nearly 30,000 Microsoft 365 Copilot licences had been issued, with 93% of licensed colleagues actively using the tools. A staff survey reported an average saving of 46 minutes a day, through work including document summaries and meeting preparation. These are the bank’s reported results, with time savings drawn from employee responses, rather than an independent measurement of additional output. Even with that qualification, the Lloyds deployment announcement describes established use at scale.

Across the wider economy, the picture is less settled. The Office for National Statistics reported in July 2026 that around 35% of UK businesses with at least ten employees used an AI technology, up from around 12% in late 2023. Its figures covered AI broadly, so they cannot tell us how close Britain is to giving every worker a personal assistant. They also showed a pronounced sector difference, with adoption at 58% in information and communication and 13% in construction. That is evidence of uneven uptake, not a permanent ceiling on what either sector might do. See the ONS analysis of AI in UK businesses.

Abroad, the distinction between trying AI and relying on it remains visible. Gallup’s US findings for the second quarter of 2026 put workplace AI use at 52% of workers, but daily use at 15%. Writing and research were leading applications among users. The Gallup workplace survey supports a future built around assistance with everyday work, while showing how far occasional use remains from a daily habit.

Those measures deserve to remain separate. A business using one AI application, an employee trying a chatbot and a licensed user returning every working day are different observations. Combining them into a single adoption story makes the future look more certain than the evidence allows.

For employers, the useful question is whether staff return to an assistant because it helps them finish a recurring task.

The email comparison needs a harder test

Email carries a message that a person has chosen to send. A generative assistant can produce the message itself, including statements the person did not supply. That introduces a different responsibility. Checking the recipient and attachment is one thing; establishing whether a convincing paragraph is true is another.

Email also has a reason to spread that a personal assistant does not automatically share. To exchange email, both parties need access to email. A worker can benefit from an assistant even if the customer or colleague on the other side never touches one. Widespread availability alone will therefore struggle to make an unhelpful assistant indispensable.

These differences strengthen the case for testing each task. They do not rule out normality.

Allen & Overy’s February 2023 announcement of Harvey’s deployment across its global practice made human review explicit. The firm said around 3,500 lawyers had submitted roughly 40,000 queries during the trial, while stating that output needed careful review by a lawyer. Its original Harvey announcement put adoption and supervision in the same account.

That is a more credible model of workplace normality than software whose answers nobody checks. An assistant can become familiar while its authority remains limited. Preparing a first draft, finding passages in an approved document collection or proposing questions for a meeting may justify regular use. Deciding what the business promises a client requires a separate judgement.

We should expect employees to learn those boundaries as part of learning the job. Familiarity must not become a reason to remove them.

Personal will mean knowing the work

The compelling version of a personal assistant is one that requires less repeated explanation. It knows which project the employee means, can locate the approved documents and understands what the employee is allowed to do.

Consider a hypothetical maintenance business preparing a quotation. An assistant could assemble the customer’s request and relevant service history, then draft a response using an approved price list. The engineer would still decide what work is necessary. The business owner would still approve the commitment. Success would mean a correct quotation reaching the customer sooner, with less administrative effort across the whole job.

That scenario asks more of the company’s records and permissions than of the assistant’s conversational charm. If the price list is obsolete, fluent prose makes the resulting quotation more dangerous. If service records belong to another customer, retrieving them is a breach of the intended boundary.

A glimpse of this more specialised direction appears in A&O Shearman’s April 2025 announcement with Harvey. The proposed agents addressed work such as antitrust filing analysis and loan review, using matter-specific documents and curated data. The announcement of legal workflow agents described systems built around defined professional tasks.

My expectation is that a worker’s apparent “personal assistant” will eventually draw on separate services with different permissions. One might help with writing; another might retrieve customer records. A third might prepare an action that requires approval. The employee could experience a single point of help while the employer retains boundaries behind it.

That arrangement has a better chance of becoming dependable than giving one conversational system unrestricted access to the business.

It also leaves an awkward question. When an assistant learns how an employee works, who controls that accumulated context? A useful future product will need a clear answer about what belongs to the employer and what the individual can retain when they leave.

The strongest objection is the checking bill

The serious argument against assistants becoming normal is economic. If checking the output consumes the time saved producing it, routine use makes little sense. A polished mistake that reaches a customer can cost more than the original task.

British research gives that objection substance. The Department for Science, Innovation and Technology’s survey of 3,500 private-sector businesses with at least five employees was conducted between February and May 2025. Among AI adopters, 75% selected improved workforce productivity as an impact, yet 77% reported no revenue change since adoption. The AI Adoption Research report also recorded difficulties in attributing business results precisely to AI.

Neither percentage settles the argument. Revenue can remain flat while a company reduces costs or improves service. Conversely, employees feeling quicker does not establish that the company completes more useful work.

Imagine an assistant allowing account managers to produce twice as many proposals. If every proposal then needs extensive correction by the same sales director, the business has accelerated the arrival of work at its existing bottleneck. The account managers’ task times improve while customers continue waiting.

That is a reason to measure the complete process, including review and rework. It is also a reason to reject particular uses without declaring the whole category a failure.

Economists Erik Brynjolfsson, Daniel Rock and Chad Syverson argued in their 2018 research brief on AI and the productivity paradox that new technologies require complementary investment and changes in how organisations work before their full benefits appear. Their argument helps explain why buying software and obtaining business gains can be separated by a long stretch of practical work. It does not guarantee that every current assistant will eventually justify its cost.

I expect assistants to become normal because they can earn their place task by task. They do not need to perform every part of a job. But there is no honest basis for promising that every employee will benefit, or that better models will remove the need to judge where assistance is worthwhile.

A company that stops an unsuccessful pilot has learned something. Keeping every licence to preserve an adoption target teaches it less.

An assistant needs the employee’s trust too

Demand is already coming from employees as well as management. Research commissioned by Microsoft and conducted by Censuswide in October 2025 surveyed 2,003 UK employees. It found that 71% reported having used unapproved consumer AI at work, with 51% doing so weekly. The Microsoft report on unapproved AI use is supplier-funded evidence from a defined sample, rather than a census of British working life. It still gives employers a practical reason to investigate what staff are already doing.

An approved assistant has to be useful enough to displace the unofficial alternative. Blocking a familiar tool while supplying an unusable replacement leaves the underlying task untouched.

Trust also runs in the other direction. An employee may welcome help with a difficult draft while objecting to managers treating every abandoned prompt as evidence of poor performance. If the assistant doubles as an undisclosed assessment mechanism, calling it “personal” becomes hard to defend.

The TUC’s May 2024 account of its proposed AI employment bill argued for workers to understand and influence important AI-driven decisions about them. That document presented a proposal, not an enacted entitlement. Its relevance here is the underlying distinction between technology employees use and technology used to manage employees.

Employers can address that distinction directly. Explain who can inspect conversations and what usage records will be used for. Involve staff in choosing the work to test. Agree how to handle mistakes before asking people to disclose them.

There is also a bargain to make about saved time. A promise of relief from repetitive administration will ring hollow if every recovered minute produces an immediate increase in workload. Companies are entitled to seek better output from an investment. Employees are entitled to ask what changes in their working day. A rollout that avoids that conversation leaves acceptance to chance.

Britain’s technology opportunity is in making the assistant useful

For British technology companies, the commercial opportunity extends beyond supplying another chat window. Customers will need help connecting assistants to dependable records and deciding which actions require approval. They will need support when a supplier changes a model or a document source stops updating.

Linklaters offers one indication of the expertise involved. In November 2025, it announced a global team of 20 specialist AI lawyers, combining legal knowledge with training in AI tools and workflow creation. The Linklaters announcement described people working within practice groups to help other lawyers use the technology. Buying access to a platform had not eliminated the need for professional judgement about its use.

A small British business cannot simply copy that staffing model. A credible local service provider could supply a narrower version, concentrating on a defined process and supporting it after launch. For the hypothetical maintenance company, that might mean keeping the assistant’s reference material current and checking whether quotation corrections are falling.

This is where I would look for durable UK technology businesses. A supplier that understands the customer’s work can sell a result the customer recognises. A generic demonstration becomes less persuasive when the buyer asks who will fix the connection to its booking system on a busy Monday.

Contracts need to reflect that continuing responsibility. Buyers need to know what support covers, how changes are tested and what they can take with them when they leave. An assistant that accumulates useful business context could become expensive to replace even when its monthly fee looks modest.

British providers have an opportunity to make that dependence manageable. They should be able to explain how the customer exits as clearly as they explain how it starts.

Watch what happens at renewal

The next convincing evidence will come from continued use after the launch campaign has ended. Watch whether staff choose the assistant for recurring work, whether completed work needs fewer corrections and whether the employer renews on the strength of those results.

For a British company deciding what to do now, a bounded trial is enough to begin. Choose a task whose current process can be observed. Record the time spent producing and checking the work, along with errors that reach the next person. Give staff approved information and a clear way to report problems. Decide in advance what improvement would justify continuing.

Keep the authority question separate. Permission to prepare an order is different from permission to place it. Permission to draft a reply is different from permission to send it. Each expansion deserves evidence from the work already entrusted to the system.

Personal AI assistants will become as normal as email when employers can provide them without treating every deployment as an experiment, and employees can depend on them for defined parts of their day. That will happen unevenly, with tasks abandoned along the way.

The revealing moment will be a new starter asking for access because the job is harder without it. By then, the employer needs a better answer than a licence. It needs to know what the assistant is allowed to do in that person’s name.