Your Next Hire Might Not Be a Person: The UK Founders Scaling With Agent Specs, Not Job Specs
Photo: Federal Communications Commission, Public domain, via Wikimedia Commons
For years, the flex at UK founder meet-ups was headcount. Twelve people, fifteen, twenty-two — proof that something real was being built. In 2026 the flex has flipped. The founder with nine staff and a seven-figure P&L is the one everybody wants to buy a coffee for.
That didn't happen because the models got clever overnight. It happened because the cost of hiring in Britain climbed steadily, the cost of software fell, and a generation of founders got comfortable handing whole job functions to something that never asks about the pension contribution.
The shape of a small company has changed
Draw the org chart of a typical UK SME five years ago and you'd see layers of people whose jobs were mostly moving information around: a sales admin chasing quotes, a support agent copying answers from a knowledge base, an operations coordinator reconciling spreadsheets.
That middle layer is precisely where AI agents have landed. Not as a chatbot bolted onto a website, but as named, budgeted, supervised functions inside the business. Founders describe them the way they'd describe a junior team member: here's what they own, here's what they cost, here's how we check their work on a Friday.
Three founders, three functions handed over
Fern & Fathom, Bristol — customer service
A homeware brand doing around £2.6m in revenue with five full-time staff. Two years ago, founder Rosie Aldridge was personally answering 60 emails a day and losing weekends to it. Now an agent handles first-line support across email and web chat, drafting responses from the product catalogue and the returns policy, with anything involving a damaged item or a complaint routed straight to a human queue.
First-response time fell from roughly half a day to under an hour. The more interesting number is the one Aldridge tracks monthly: the share of contacts resolved without a human touching them sits at 61%. That's the equivalent of nearly two full-time roles she hasn't had to advertise — and two salaries' worth of gross margin she's been able to push back into stock and paid acquisition.
Sablewell, Manchester — pipeline
A B2B software firm at £6.9m ARR with eleven employees. Sablewell's growth team is one person plus a set of agents that research target accounts, enrich records, draft personalised opening emails and book meetings into the founder's calendar.
The agents don't close anything. They just make the top of the funnel wide enough that one human can work it properly. The company books between 35 and 40 qualified meetings a month. The stack costs under £14,000 a year all in, including the founder's own oversight time. A comparable human SDR in Manchester — salary, employer National Insurance, pension, software, recruitment fee — lands somewhere between £45,000 and £55,000 fully loaded, and that's before you factor in the three months it takes to get them productive.
Morton Vale, Glasgow — delivery
A compliance consultancy with seven people and about £3.4m in fee income. Consultants were losing entire days to first-pass report drafting: reading source documents, pulling out the relevant clauses, formatting. An agent now produces the draft, complete with source citations, and a qualified consultant reviews, corrects and signs off.
Partner Callum Doig reckons each consultant gets back around eleven hours a week. The firm hasn't added a single fee earner in eighteen months and has still grown revenue by a quarter. Doig's line is worth repeating: "We didn't buy a robot. We bought back the parts of the job nobody went into compliance to do."
The unit economics nobody puts in the deck
Strip out the hype and the maths is unglamorous but compelling.
- A fully loaded UK admin or support hire costs roughly £34,000–£42,000 once employer NI, pension, kit, software and management time are counted.
- A supervised agent stack for one function typically runs £4,000–£18,000 a year, depending on volume and how much orchestration and evaluation tooling sits around the model.
- Oversight is the hidden cost. Budget one to three hours a week of a senior person's time per agent. If that person is on £70,000, that's real money — but nowhere near a salary.
- The payback period is usually weeks, not years, which is why founders are comfortable experimenting and killing things that don't work.
The trap is treating the subscription as the whole cost. The genuine cost is the process redesign, the knowledge base you have to clean up first, and the senior attention required to keep quality steady as the business changes.
What the stack actually looks like
You don't need to be technical, but you do need to know the layers. Every founder we spoke to described the same five-part shape:
- A system of record — the CRM, helpdesk or ERP everything ultimately writes back to. If this is messy, nothing downstream works.
- An orchestration layer — the thing that decides which agent runs, in what order, with what tools.
- A grounded knowledge base — policies, price lists, product data, past tickets. Garbage in, confident nonsense out.
- Guardrails and evaluation — automated tests that check the agent's output against a set of known-good examples before it goes anywhere near a customer.
- A human review queue — the escape hatch, and the most important part of the whole design.
The founders who struggled were the ones who bought the model before they fixed the knowledge base.
The compliance bits that bite
This is the part that separates a grown-up deployment from an expensive experiment.
- UK GDPR and the ICO. You need a lawful basis, a data processing agreement with every vendor, and a clear retention policy. Know where your data is processed.
- PECR and the TPS/CTPS. Outbound calling and emailing still have rules. An agent doesn't get you out of them — it just helps you break them faster if you're careless.
- Sector rules. If you're anywhere near financial services, mortgage advice, insurance or consumer credit, FCA expectations around Consumer Duty apply to your automated channels exactly as they do to your people.
- Advertising standards. Don't let an agent make claims about your product the ASA's CAP Code would object to. Test your prompts against your marketing copy.
- Employment law. If automation removes roles, remember collective consultation obligations kick in at 20 or more redundancies. Handle it properly and early.
- Disclosure. Tell customers when they're talking to an automated system. It's good manners and increasingly good practice.
Where a human still wins
No founder we spoke to was remotely tempted to automate these: negotiation, pricing decisions, the first ninety days of a new client relationship, anything that touches reputation, and every situation where the answer is "it depends."
Agents are brilliant at the repeatable eighty per cent and quietly terrible at the twenty per cent that actually determines whether a business is any good. Design your workflow around that split and everything else gets easier.
The solo-scale checklist
If you're weighing up a first hire versus a first agent, work through this before you spend a penny:
- Map the job function into tasks and mark each one repeatable or judgement-led.
- Clean the underlying data and knowledge before buying anything.
- Choose a single function for the first deployment, not three.
- Set a monthly budget ceiling, including oversight hours, and track cost per outcome.
- Write the escalation rules in plain English before you write a single prompt.
- Run a two-week shadow period where the agent drafts and a human sends.
- Build an evaluation set of 30–50 real cases and re-test monthly.
- Get the data processing agreement and retention policy signed off.
- Name a human owner for each agent. No orphan bots.
- Review at 90 days with one question: has this removed work, or just moved it?
The founders getting this right aren't anti-hiring. Several plan to add people this year — in sales, in product, in account management. They've simply decided that their next headcount should go to the work only a person can do. That's not a headcount freeze. It's a headcount upgrade.