AI agent cost vs salary: the real maths for a UK business
Owners keep comparing a subscription price to a salary and getting the sums wrong on both sides. Here is what an AI agent costs per unit of work, what an employee really costs once the on-costs are in, and the comparison that actually holds up.
"Is it cheaper than hiring someone?" is the first question most owners ask about an AI agent, and most answers are wrong because both sides of the sum are priced incorrectly. A person isn't priced at their salary, and an AI agent isn't priced like a Netflix subscription. Get the units right and the comparison is straightforward.
How much does an AI agent actually cost?
Behind almost every AI agent sits a language model billed by usage, not by seat. Anthropic's published pricing for Claude, for example, runs from $1 per million input tokens for its smallest model, Haiku, through $2 for its mid-tier Sonnet model, up to $10 for its most capable model, with output tokens billed at five times the input rate. A token is a few characters, so a million tokens is roughly 700,000 words: far more than one task consumes. Anthropic's own worked example puts processing 10,000 customer support tickets, each averaging about 3,700 tokens, at roughly $37 in total on its cheapest model. That's a fraction of a penny per ticket.
Scale that up: an agent that reads incoming enquiries, drafts a reply and flags anything unusual for a human to check might cost a UK business somewhere between a few pence and a few pounds a day in model usage for a few hundred items, depending on how much text it reads and writes and which model does the work. On top of that sits the cost of building the thing and keeping it running: integration with your inbox or CRM, monitoring, and fixing it when a workflow upstream changes. That build-and-run cost is usually the larger number in year one, and it's a fixed or retainer cost, not a per-task one: the marginal cost of the agent doing one more task is close to zero.
What does a UK employee actually cost beyond salary?
A salary is the headline, not the bill. In 2026/27, UK employer National Insurance is 15% on earnings above £5,000 a year, though eligible employers can claim the Employment Allowance, which takes up to £10,500 a year off the bill. Auto-enrolment adds a minimum employer pension contribution of 3% on earnings between £6,240 and £50,270. On a £30,000 salary that is £3,750 of National Insurance and about £713 of pension, before any allowance. Add sick pay, any bonus or overtime, plus the softer costs (management time, training, and the output lost to 28 days' statutory holiday, sickness or someone leaving), and a "fully loaded" cost somewhere around a quarter to a third above gross salary is a reasonable working estimate for a typical UK role. A £30,000 salary is closer to £38,000 to £40,000 once you count what it actually takes to keep that person in the seat, before recruitment costs if you have to replace them.
How should you compare an AI agent's cost with a salary?
Comparing a subscription fee to a salary is the wrong sum. The right one is cost per unit of output. A person doing repetitive text work (reading enquiries, drafting standard replies, extracting data from documents) might process a few hundred of them in a working day, at a fully loaded cost of somewhere around £150 to £250 a day depending on seniority and sector. An AI agent doing the same category of task at the volumes above costs a small fraction of that once it's built, because each extra item adds only pence to the model's usage bill, while a person's cost rises with every extra hour the work takes.
That doesn't make the agent a straight replacement. It's cheap at the specific thing it does well: high-volume, well-defined, text-shaped work with a clear right answer most of the time. It's expensive, or simply the wrong tool, for anything that needs judgement, a relationship, or a decision nobody wants to see get made by software unsupervised. The honest shape of the comparison is that an AI agent competes with the repetitive slice of someone's job, not with the person. Most roles are a mix, which is why the systems M22 builds keep a human approving anything that matters rather than running the whole job unattended.
When does an AI agent beat a salary on cost?
If a task is high-volume, text-based, and has a mostly-consistent right answer, price it in pence per item and compare that to what a person's time on the same task actually costs once the on-costs are counted. If it needs judgement or a relationship, don't try to cost it that way at all: that's not where the saving is, and pretending otherwise is how automation projects lose trust. For what a build to do this actually costs to commission, see our guide to AI consultancy pricing in the UK, and for the wider "should I build one at all" question, M22's audit process starts by pricing exactly this trade-off for your business before anything gets built.
How does M22 help with this?
This is exactly the sum M22 Consultancy runs before recommending a build: the AI Audit prices the hours a task currently costs against the expected return of automating it, so the agent-versus-salary comparison is answered with your own numbers rather than a rule of thumb. M22 builds the workflow automation or agent on whichever model fits the task, Claude, ChatGPT or Microsoft Copilot, and a monthly retainer then measures the time it actually saves. The AI Audit starts at £1,500, a fixed fee agreed before day one, with no obligation to build afterwards. Book a thirty-minute call to price your own numbers.