AI agents could make staying small a serious ambition for UK founders

Britain’s starting point is already small

At the start of 2025, the UK had an estimated 4.3 million private-sector businesses employing nobody beyond their owners, representing 75% of the business population. These were not all one-person limited companies. The category includes sole proprietorships and partnerships, alongside companies classified as having no employees beyond a working proprietor. That distinction matters when estimating the potential market for agent-run businesses. The Department for Business and Trade’s business population estimates describe an established pattern of working without a wider payroll, rather than evidence of an AI-driven business boom.

Readiness is a separate question. In the Longitudinal Small Business Survey 2024, 53% of businesses without employees used technologies or web-based software to sell or manage their business. That measure covers ordinary business software, not specifically AI. The same survey found that 59% wanted to grow sales over the next three years, while just 27% reported sales growth over the preceding year.

Those figures suggest a more useful starting point than the hunt for a lone founder with a billion-pound valuation. Britain already has owners who want more business and have limited capacity to absorb it. The question is which parts of their working day can be delegated without making the service worse.

For a plumber, automating enquiries cannot repair another boiler. It might reduce the time between a missed call and a confirmed appointment. For a specialist producing digital reports, delegation could reach further into delivery. The economic effect depends on what constrains the business, so a national count of AI subscriptions would tell us little about whether owners were becoming more productive.

The examples prove different things

The international success stories deserve attention, provided we distinguish an AI product from a company operated by autonomous agents.

PPC Land reported that Dutch entrepreneur Pieter Levels’ PhotoAI generated US$132,000 in monthly revenue, based on his own public disclosures. Its account of the PhotoAI business describes a founder building and selling AI-generated photography through a compact operation. The revenue is founder-reported, rather than independently audited in that article. It demonstrates a commercial opportunity for a small operator selling AI capabilities. It does not establish that agents independently manage every business function.

More direct evidence of delegated work comes from Salesforce’s own production data. Its Agentic Enterprise Index for the first half of 2025 reported a 119% increase in agents created and deployed within participating organisations. Reported actions included querying records and creating customer-service cases. The dataset selected businesses with agents active in production throughout the period, so it describes committed adopters rather than the whole market.

Within the same report, escalations to people rose from 22% in the first quarter to 32% in the second. Salesforce presented that as agents becoming better at recognising when human help was needed. Commercially useful automation can therefore increase the importance of a clearly defined handover.

For a solo owner, every such handover returns to one person.

A British example brings that constraint closer to the customer. In an April 2026 account of building Nyata AI, Ibrahim Maina described drawing on his experience in a Bristol homeless hostel to build a conversational service for finding local support. He reported having mobile apps available and a business development partner onboard, but no revenue at that point.

Maina identified maintaining accurate information about local organisations as a harder problem than making the model call. Opening hours change; services move. His description was precise. “The data layer is the real product.”

That is an early founder’s account, not proof of a profitable autonomous company. It does suggest where a British founder’s advantage might come from. Knowledge of a particular community or trade gives the software a purpose. Generating more software cannot supply that knowledge by itself.

A company built around delegated decisions

Consider a hypothetical British specialist who helps independent retailers maintain supplier catalogues. Today, the work might involve reading inconsistent spreadsheets, resolving product descriptions and preparing updates for an online shop.

An agent-based service could attempt the repeatable preparation, then route uncertain matches to the owner. Another automated check could compare proposed changes with the original supplier file. Publication could require approval when a price changes beyond an agreed limit.

The commercial advance would come from reducing the owner’s work per accepted catalogue update. Merely processing more files would be a poor measure if the founder spent the evening correcting them.

This distinction changes how we need to think about the company itself. Its working design becomes a set of decisions about authority. Which records can software read? Which changes can it make? When must it stop? Who can restore the previous state?

A founder could retain control while delegating substantial execution. Customers might receive a faster service, and the owner might serve more shops without adding an employee. Equally, a poorly designed version could multiply incorrect prices before anyone noticed.

I would expect the strongest one-person businesses to choose their boundaries deliberately. They would offer a narrow service with recognisable acceptance criteria, rather than promise to handle anything a customer asked. Their owners would know enough about the work to recognise an implausible result.

Staying small then becomes a design choice. The founder can decline work that introduces too much judgement or support, even when the software appears capable of attempting it.

Human attention remains an expensive input

Research on agent systems gives a concrete reason to resist the idea that adding more agents automatically removes the need for supervision.

The 2025 paper *Why Do Multi-Agent LLM Systems Fail?* examined seven multi-agent systems across more than 200 tasks and identified 14 distinct kinds of failure. These included agents repeating steps, losing context and checking results incorrectly. The study concerns particular systems and tasks, rather than a universal failure rate for present-day products. Its practical warning is that coordination and verification introduce problems of their own.

Giving one agent the job of checking another does not, by itself, establish an independent check.

British public-sector experience also shows how much care can sit behind a bounded assistant. In March 2026, the Government Digital Service reported 90% answer accuracy in its testing of GOV.UK Chat. Assessment combined subject specialists with automated tools, and answers drew exclusively on GOV.UK guidance. An answer could fail the accuracy standard because it addressed only part of a question. This was a question-answering service, not an autonomous company, but its testing illustrates the work required even within a restricted information domain.

For a solo founder, the relevant calculation must include time spent investigating exceptions. A cheap subscription can become expensive if every customer interaction creates another review task. Conversely, a more expensive service could earn its place if it reduces those interventions without concealing mistakes.

The owner’s absence provides a useful test. Can routine work continue safely during a day off? Can a trusted external specialist recover the service if the founder is ill? Can customers obtain help without waiting for the only person who understands the system?

A company that requires constant personal vigilance has not escaped the limits of self-employment. It has changed the work its owner must stay awake to do.

The strongest objection is commercial

The strongest argument against a new generation of successful solo companies is that their competitors can buy the same tools.

If agents make a service cheaper to produce, new entrants can copy the offer. Existing suppliers can use agents while retaining their customer relationships. Buyers may demand lower prices. Under those conditions, productivity gains could flow to customers or software providers while founders struggle to retain a margin.

I do not think the available examples settle that objection. A profitable founder’s revenue screenshot tells us little about the people who built comparable products and never found a buyer. Nor does a successful launch establish that customers will renew once alternatives arrive.

My case for solo companies therefore rests on something narrower than widespread access to capable software. An owner needs an advantage that survives when competitors acquire similar tools. That might be permission to use valuable data, knowledge of an awkward workflow, or customers who already trust their judgement.

Agents could make those advantages easier to turn into a repeatable service. They cannot guarantee that the advantage exists.

Nor should a one-person payroll be confused with an economy containing only one person’s labour. Hosting and payment services involve other organisations. Specialist contractors may handle security or accountancy. A founder buying those services has a business supported by other businesses.

There is nothing illegitimate about that arrangement. It does mean that the costs and dependencies belong in the assessment. Comparing an AI subscription with the salaries of an imagined team omits the work still being purchased elsewhere, as well as the founder’s own time.

I would judge the model by durable profit after those costs, acceptable service and manageable working hours. Remaining the only employee is optional.

Britain can sell the dependable parts

The UK technology sector has an opportunity to supply the components that make such businesses manageable.

London-based Convergence offers evidence of British work on the underlying agent technology. In May 2025, Tech.eu reported Salesforce’s acquisition of Convergence, founded by Marvin Purtorab and Andy Toulis the previous year. The report described a team specialising in agent design and autonomous task execution. Financial terms were undisclosed.

That is evidence of demand for UK agent expertise, rather than evidence that a one-person company has replaced a team. Both kinds of business could develop together. Small operators would buy capabilities built and maintained by larger specialist organisations.

For British software suppliers and managed service providers, I would place a bet on products that reduce the work of checking and recovering delegated actions. A founder needs to see what an agent changed, understand why it stopped and revoke its access without dismantling the whole business. Clear usage charges and usable exports matter when one owner has to assess the commercial terms.

Support has a place in this market. An owner might happily buy an on-call service that provides cover during illness or investigates a broken connection. Selling dependable assistance could be a better proposition than encouraging every customer to become an expert in agent orchestration.

Existing British companies also need to reconsider how they assess small suppliers. Employee count is a weak substitute for evidence about delivery. A one-person supplier with a tested recovery process could be a sensible choice for a bounded service. A supplier that cannot explain how customers retrieve their data is harder to trust, whatever its headcount.

Accountability and the missing apprenticeship

Autonomy does not remove the business from its obligations to customers. The Competition and Markets Authority’s March 2026 guidance on consumer law and AI agents says businesses remain responsible for agents’ actions, including when a third party provides the technology. It calls for testing and monitoring, with problems addressed promptly. This is not legal advice; consult your legal counsel.

That principle belongs in the product design. If an owner cannot identify misleading information or correct an inappropriate decision, giving the agent more freedom increases exposure faster than it increases useful capacity.

There is also a less immediate question for the UK. Where will future founders acquire the expertise that lets them judge the software’s work?

In October 2025, King’s College London reported research into AI exposure and UK employment. Drawing on job postings and LinkedIn profiles, the study found employment reductions concentrated in junior positions at highly exposed firms. The university reported a 5.8% fall in those roles. That finding is not proof that each missing position was directly automated, but it raises a serious concern about access to early career experience.

The solo-founder argument works particularly well for someone who already knows a profession. It becomes less comfortable when we ask how a newcomer learns that profession if the preparatory work no longer creates an entry-level job.

A country can gain productive independent businesses while weakening the route through which people become capable of running them. British employers and training providers need to treat supervised work with AI as a skill-building opportunity, with learners required to inspect evidence and defend decisions. Familiarity with the interface is not enough.

Watch what happens after the launch

The evidence to watch now is what happens when an agent-supported company acquires repeat customers.

Does revenue grow without a matching increase in the founder’s working hours? How much time goes into corrections? What proportion of delivery depends on contractors? Do customers stay when a rival offers something cheaper? Those answers would establish far more than the number of agents a founder has configured.

For an owner considering this model, a paid, bounded service is a better starting point than a plan for an autonomous organisation. Establish what a customer will accept, record the current delivery cost and delegate a task whose results can be checked. Expand the agent’s authority only when the evidence supports doing so.

Hiring remains a legitimate outcome. A second person may provide expertise or cover that produces more value than another layer of automation. Refusing to hire merely to preserve the solo-founder label would turn a useful business option into an expensive personal rule.

I expect Britain to produce more companies whose owners can choose to stay small because agents handle repeatable work. Their contribution will be visible in services that become affordable to offer, and in owners who can accept another customer without sacrificing another evening.

The convincing one-person company will be one whose customers receive what they paid for, whose owner earns a sustainable living, and whose systems still behave sensibly when that owner closes the laptop.