Britain’s next workforce layer will arrive before its management catches up

Adoption is earlier than the headlines suggest

Evidence for an agent-filled British economy remains thin. The Department for Science, Innovation and Technology’s AI Adoption Research, based on interviews with 3,500 businesses between February and May 2025, found that 16% used AI. Among those adopters, 7% reported using agentic AI, compared with 85% using natural language processing and text generation.

Those figures need their boundaries. The survey covered private-sector businesses with at least five employees in its selected sectors. It did not measure shadow AI use, and its fieldwork describes a particular period. It cannot establish how many British companies now have autonomous systems operating across their business.

It does puncture the assumption that widespread use of writing assistants proves widespread delegation to agents.

For this discussion, the useful distinction is whether software can carry a task through successive actions within an agreed scope. Consider a hypothetical overdue invoice. Software might explain the account history to an employee. An agent might also choose the next action, contact the customer and update the record. Those permissions create different management obligations even if both products appear behind the same chat window.

British companies therefore have time to prepare. The evidence does not justify telling every small business that it has already fallen behind. It does justify asking what happens when the tools it already buys begin offering to carry out work.

Agents have a distribution route into British businesses

Sage’s June 2025 statement of its agentic AI plans is instructive. It distinguished existing task automation and generative assistance from emerging agentic applications. More autonomous systems, capable of detecting business risks and initiating responses, were described as being in early testing, with an expected customer horizon of 12 to 36 months.

That was a statement of intent at the time, rather than proof that every described capability had reached customers. Its importance lies in the proposed route to market. An accounting supplier can introduce greater autonomy through an existing customer relationship. A small business does not have to commission its own AI platform to encounter the decision.

Across the Atlantic, Salesforce’s October 2024 Agentforce launch announcement described agents operating through its existing platform and automation tools. The announcement included Wiley’s account of taking a customer-service agent live following a pilot. Wiley reported an increase of more than 40% in case resolution compared with its previous bot.

That is a customer claim published by the supplier, with limited evaluation detail. It supports the existence of a deployment and a reported benefit. It does not establish what another company will save, or how much human checking its own workload will require.

Workday made the management implications explicit in its June 2025 Agent Partner Network announcement. Its proposed Agent System of Record would cover agent responsibilities, data access, permitted actions and performance. Workday was selling a way to administer agents alongside the work they performed.

Taken together, these announcements explain why I expect adoption to outrun management practice. Suppliers can package a capability and distribute it through established products. Each customer still has to decide who may authorise its use and who will deal with its mistakes.

A feature release can reach customers on a timetable. Changing how a finance team delegates authority requires agreement about the work itself.

Britain can sell expertise through agents

A&O Shearman’s April 2025 announcement with Harvey shows a different route. The firms announced agents for work including antitrust filing analysis and loan review. Their description covered breaking a problem into steps, working across documents and combining intermediate results under oversight.

The commercial arrangement matters. A&O Shearman said the agents would be used internally and sold to clients and other law firms, with the firm sharing software revenue.

Here, professional expertise becomes part of a product that can perform a defined sequence of work. That gives British service businesses a possible new source of revenue. It also forces a harder question about what customers are buying. Is the product supplying a draft for an expert to review, or accepting responsibility for completing a service?

Those are different propositions. Marketing language about a digital colleague cannot settle the distinction.

There are signs of the same approach beyond professional services. Microsoft’s December 2025 Agentic Launchpad cohort announcement described Raft AI using agents in logistics workflows, with people retained at designated control points. As a programme announcement, it provides supplier-side evidence of activity rather than an independent assessment of performance.

For British technology companies, this suggests a market in the detailed knowledge surrounding a task. A freight workflow needs a definition of an acceptable document and a route for resolving discrepancies. An accounts workflow needs someone to decide which exceptions justify intervention.

I would back a supplier that can demonstrate those decisions over one that merely demonstrates a fluent conversation. Customers need evidence that the system recognises the boundary of its assignment.

Supervision has to survive a busy Tuesday

Imagine a small British wholesaler piloting an agent to follow up unpaid invoices. This is a hypothetical example, not a customer case study.

The easy cases produce encouraging results. The agent finds an overdue balance and prepares a polite reminder. Then it encounters a disputed delivery. A credit note exists in one system but has not reached another. The account manager has agreed to wait until Friday, in a conversation that was never recorded.

A person familiar with the customer might know to stop. The agent can only work with the information and controls available to it. Giving it a broad instruction to reduce overdue debt leaves the business to discover which exceptions it failed to specify.

Management begins with deciding what the agent may do when its evidence is incomplete. Can it send a reminder without review? Can it offer revised terms? Which discrepancy suspends the workflow? Who receives that suspended case?

Putting a person at the end of the process does not answer these questions. Review requires time and evidence. If an employee receives a polished recommendation without the underlying account history, approval can become an act of trust in the presentation.

I would judge supervision by whether the reviewer can reject the action, understand why it was proposed and obtain help when the case exceeds their authority. A nominal approval button tells us little.

The economics also change once supervision is counted. Faster drafting produces a saving only to the extent that it reduces work elsewhere. Checking, correcting records and handling complaints belong in the calculation. So does keeping a manual route available when the agent stops.

This is why a small business may sensibly buy a managed service around an agent. Its provider must then explain who handles an exception and how quickly. A subscription to software alone cannot answer the telephone when a customer disputes what that software has done.

The strongest objection is that dependable autonomy may arrive slowly

There is a serious case against the workforce-layer forecast. Agents may remain too unreliable or expensive to receive broad authority. Buyers may adopt useful assistance while retaining human control of consequential actions. Established automation may handle a bounded process more predictably.

Gartner’s June 2025 forecast anticipated that more than 40% of agentic AI projects would be cancelled by the end of 2027, citing rising costs, unclear business value or inadequate risk controls. It also warned about existing assistants and automation being relabelled as agents.

That remains a forecast, with a deadline that has not yet passed. But its explanation deserves attention. A business can be right to cancel a project after discovering that conventional software solves the problem more cheaply.

Britain’s experience with other AI systems supplies a further caution. A UCL-led evaluation of NHS chest-diagnostic AI deployment, reported in September 2025, found that contracting had taken four to ten months longer than anticipated. By June 2025, 23 of the 66 participating trusts were not yet using the tools in clinical practice. Challenges included connecting to varied existing IT systems and finding time for clinical staff to participate.

These were diagnostic support tools, not autonomous workforce agents. Their relevance is organisational. Even a funded programme with a defined purpose had to accommodate the systems and people already doing the work.

That evidence limits any confident prediction of a sudden national takeover. Slow deployment could give managers time to learn. Better tools could reduce the checking required. Businesses might become more selective after disappointing pilots.

My expectation is still that delegated software work will spread, but unevenly and within boundaries. An agent does not need to run an entire department to change who performs its routine work. A collection of narrow permissions can become a workforce layer without producing anything resembling an autonomous employee.

The strongest version of the counter-argument changes the likely shape of adoption. It leaves the need to manage those permissions intact.

Removing junior work creates another management debt

Britain’s employment debate needs equal care. Redundancies do not, by themselves, prove that agents have replaced people. Research on exposure to language models cannot identify the effect of autonomous agents alone.

King’s College London’s October 2025 account of research into the UK labour market reported that firms whose workforces were highly exposed to AI capabilities had reduced employment by 4.5% on average. Junior positions fell by 5.8%. The study analysed job postings and LinkedIn profiles from 2021 to 2025.

Those findings warrant attention, while remaining distinct from a count of jobs lost to agents. They raise a management question that the promise of automated work does not resolve.

Where will the next generation learn to recognise a bad answer?

If a firm removes routine document review from a junior role, it needs another way for that employee to learn what a sound review looks like. Assigning a new starter to supervise an agent does not automatically supply the judgement needed to challenge it.

There is a constructive alternative. Trainees can compare their own conclusions with the agent’s output and explain disagreements to an experienced colleague. Saved time can support closer feedback or exposure to harder cases. But that requires a deliberate training design and a budget for supervision.

A company that treats every automated hour as immediately removable labour risks consuming its own training capacity. The cost will appear later, when experienced staff leave and their replacements have reviewed fewer real cases.

Employees also need a reason to report where agents fail. If discovering a weakness is treated as resistance to progress, the pilot loses useful evidence. Managers who want honest results must make it possible for staff to say that a manual step still earns its place.

Britain’s technology sector can compete on the work around the model

For British software firms and service providers, the opportunity extends beyond building an agent. Customers will need help specifying its authority and measuring its results.

A credible service could include connecting the required records, testing against difficult historical cases and monitoring actions after launch. It could also cover handing the work back to people when the system becomes unavailable. Those activities require knowledge of the customer’s process that a model demonstration cannot supply.

This changes what a useful sales demonstration looks like. Show the buyer a disputed transaction. Remove a required document. Present conflicting instructions. Demonstrate what the agent records when it refuses to proceed, then show how a person takes over.

Support and exit deserve the same attention. A customer needs to know whether it can export the action history and retain the rules used to govern the workflow. Changing supplier becomes harder if the business can no longer explain how its routine decisions are being made.

There is a choice here for the UK channel. It can sell installation followed by loosely defined support, or it can sell responsibility for keeping a specified service working within agreed limits. The second proposition is harder to price, because it includes the awkward cases. It is also more useful to a business without its own AI engineering team.

Britain’s advantage will have to be demonstrated in completed work that customers can inspect.

Watch the authority being delegated

Counting agents will tell us less than examining what they are permitted to do.

I would watch for published evidence of sustained use, with the scope of authority made clear. Useful reports will include human correction rates and the cost per accepted outcome. They will explain whether a resolved case stayed resolved, and what happened when the system encountered information it could not reconcile.

Inside a business, the next step can be small. Choose one recurring workflow and write down the actions that require approval. Name the person who owns the result. Test the awkward cases before widening the agent’s access. Keep the previous process available until the new one has demonstrated that it can carry the work.

For boards, the revealing question is whether an agent can acquire a new permission without the accountable manager understanding what has changed. For employees, it is whether supervision comes with enough time and authority to challenge an output.

I expect Britain’s next workforce layer to emerge through thousands of these ordinary decisions. Its arrival may look less like hiring a digital employee than accepting another feature in an application renewal.

Before that feature acts on a customer’s account, someone needs to be able to explain who authorised it, what it may do and how to stop it. That person is already managing an agent, whatever their job title says.