Where the hundred thousand actually goes
Start with overhead. The 2026 agency benchmarks put it at 30 to 40 per cent of revenue, with the top quartile holding below 25 and anything past 45 treated as a firm carrying too much fixed cost for its revenue base. Overhead is rent, finance, HR, leadership, new business, and the three pitches the agency lost before it won yours. On your hundred thousand that is thirty-five to forty thousand pounds, spent before anyone opens an editor.
Then margin. The average digital agency made a 13 per cent net margin after tax in 2025, down from a long-run average of 15 since 2015. Firms over fifty people averaged 8 per cent. Studios under ten averaged 19. Call it thirteen thousand.
What is left, roughly half your budget, is direct delivery cost. Except delivery is not quite what you think it is. The benchmark for a healthy professional services firm is a ratio of at least three billable staff to every one non-billable. Account managers and project managers count as billable and sit at 60 to 70 per cent utilisation, because their work splits between your project and internal coordination. Senior leadership bills 30 to 40 per cent. Designers and developers, the people actually producing the thing you are buying, are targeted at 80 to 85.
And they miss. Median agency utilisation in 2026 is 68 per cent against a 75 per cent target. That seven-point gap is unbilled time which still has to be paid for, and it is paid for out of your rate.
Work it through and the money reaching a keyboard is somewhere around a third of what you signed for. You will see the claim that it is ten per cent. It is not, and the exaggeration does the argument a disservice. A third is bad enough, and a third is verifiable.
Three billable staff to every one non-billable is not a warning sign. It is the benchmark for a healthy agency.
The coordination layer is the product
The other two thirds is not theft. It buys something, and it is worth being precise about what.
Put five specialists on a project (a designer, a backend developer, a frontend developer, a project manager, an account manager) and you have ten communication paths between them. Add QA and a DevOps engineer and you have twenty-one. Every path is a briefing, a ticket, a Figma comment waiting on a reply, a Slack thread, a stand-up, a handoff where context is lost and then half recovered the following day.
The account manager exists because seven specialists cannot each hold the whole project in their head, so somebody has to. The project manager exists because twenty-one handoffs need sequencing. The overhead exists because those people need hiring, managing, paying and replacing.
None of that is padding. It is the genuine and unavoidable cost of splitting one piece of work across enough people that no single one of them understands all of it. Agencies have been selling that structure for decades and it was the right structure, because for decades nobody could hold design, backend, frontend, content and infrastructure at a professional standard at the same time.
So you have never really been buying hours. You have been buying the translation between people who each know one part.
AI did not make developers faster
Here is where most arguments like this one go wrong, and mine would too if I were careless.
In July 2025 METR ran a randomised controlled trial on exactly this question. Sixteen experienced open-source developers, 246 real tasks in their own repositories, repositories averaging over 22,000 stars and a million lines of code. Each task randomly assigned to allow or forbid AI tooling. The developers predicted a 24 per cent speed-up. Afterwards they believed they had got 20 per cent. Measured, they were 19 per cent slower.
METR repeated it with late-2025 tools and published again in February 2026: 57 developers, 143 repositories, more than 800 tasks. The original cohort was still 18 per cent slower, the newly recruited group 4 per cent slower, both with confidence intervals wide enough to cross zero. METR are scrupulous about the weakness of their own result. Developers were declining to take part rather than work without AI, and between 30 and 50 per cent said they withheld tasks they did not want to attempt unaided. METR describe their own estimate as a lower bound.
So if my pitch were that AI lets me type four times faster, the best randomised evidence available would be against me, and I would deserve it.
That is not the claim. The claim is narrower and much harder to argue with. AI did not compress the work. It collapsed the handoff.
When I build something now there is no ticket, because there is nobody to write one for. No Figma comment waiting on a reply. No sprint boundary between noticing the content model is wrong and fixing the content model. No briefing, because the person who decided the design language is the person writing the migration. I run agents across the schema, the frontend and the copy at once, and the work that used to happen between people now happens inside one context.
Twenty-one communication paths become zero. Not faster. Absent.
Production got a little cheaper. Coordination stopped being necessary. Only one of those was ever a line item you could see.
This site did not exist on a Sunday morning
No design. No content model. Not a line of code, not a comma of copy.
By that evening it was live on its own domain and populated: a Strapi 5 backend with the full content model, a SvelteKit frontend, a design language, the services pages, the case studies, and the writing you are reading this in. Deploying to Vercel and Strapi Cloud, with accessibility tests running in the pipeline.
The honest limit, since you would find it anyway: I was the only stakeholder. No procurement, no brand guidelines to interpret, no legal review, nobody to persuade. That is the easy version of the problem and I am not going to pretend it was the hard one.
But look at what was absent that would have been absent regardless of the client. Nobody needed briefing. Nobody waited on anybody. There was no point at which the work stopped because the person holding the next piece was in another meeting. Those delays are not caused by difficult clients. They are structural, and a difficult client makes them worse rather than creating them.
What took the time was deciding what to say, and being straight about which evidence I actually had. That part does not automate, and it is the part worth paying for.
The market has already worked this out
WPP went into 2024 with 108,044 employees and came out of 2025 with 98,655. That is 9,389 people in roughly twelve months, with further cuts expected through 2026. The holding company structure is being dismantled into four units, Creative, Media, Production and Enterprise Solutions, unified around an internal AI platform. The programme is called Elevate28, it targets £500 million of savings by 2028, and it costs £400 million to execute. The company has dropped out of the FTSE 100 after around thirty years on the index.
Underneath the holding groups the same thing shows up in smaller numbers. Wow Company's BenchPress survey recorded the first notable decline in average agency hourly rates in 2025. The ANA has had more than 80 per cent of major brands running in-house agencies since 2023, and reporting since suggests brands trimming agency spend by up to a quarter while keeping the strategic relationship.
When the largest holding group in the industry restructures around an AI platform and removes nine thousand roles, that is not a forecast about the model. That is the model repricing itself in public, and it is repricing faster than most of the quotes you will be sent this quarter.
Three questions for your next agency quote
Before any pitch of mine, three questions. They are more useful than anything I could tell you about myself, and they work on me too.
What proportion of this number is delivery labour? Ask for it as a percentage. An agency worth hiring will give you one. A weaker one will explain why the question is naive, which is itself the answer.
Who writes the code, and what utilisation are they carrying? If the target is 75 per cent and the industry median is 68, you are funding the gap between the two. You are also entitled to know whether the senior name on the pitch deck appears anywhere in the delivery plan.
If the work takes half the hours, what happens to the price? If the answer is nothing, you are not buying the work. You are buying the structure, and the structure is the part that has changed.
Then the alternative, which is the reason I wrote this. One technical lead, hands on your codebase, working with you directly. No account manager in between, because there is nothing left to account-manage. A small number of engagements at a time, so the one you have is the one getting the attention. Knowledge transfer rather than dependency: when the engagement ends your team owns the thing, and I would rather be re-hired than be necessary.
I do not publish rates, because the honest number depends on the work, and a figure written before anything has been scoped is a marketing decision rather than a price. Ask, and I will tell you. Bring the quote you are comparing me against while you are at it. That conversation will be more useful to you than any capability deck.