The quiet death of the dev shop

There are roughly twelve thousand digital agencies in the UK, depending on how you draw the boundaries. Most of them have a margin model that depends on the cost of producing software holding steady. It has not held steady. The reckoning is, by our estimate, about eighteen to twenty-four months out.

This is not a prediction in the gloomy sense. It is a structural argument from how the numbers actually work, which we will lay out plainly, because almost no one inside the industry seems willing to.

The basic arithmetic

A standard mid-sized UK digital agency in 2024 looked something like this. Thirty-five staff. Blended cost-per-head, fully loaded, in the region of £75,000 a year. Day rate to clients in the range of £700 to £1,100 depending on seniority. Utilisation target around 70%. Gross margin on the build, before overhead, of about 35-45%.

The model is hours-in, hours-out, with a margin on the spread.

The model works if, and only if, two things stay roughly stable. One: the cost per useful unit of code produced stays roughly proportional to the headcount. Two: the price clients are willing to pay per hour stays roughly above the cost of producing the hour. Both have been stable for so long that most agency owners do not think of them as variables. They are variables.

Both have moved.

The cost side has moved because a single competent developer with a Cursor licence and good prompting habits now produces, on a sustained basis, what two or three of the same developer produced two years ago. We measure this internally. We see it in the work coming out of the Academy. The number is conservative; the actual upper bound is higher and shows up in spurts. The mean has roughly doubled. Call it 2.2× on a typical week.

The price side has moved because clients are not stupid. They can see, in adjacent markets, what AI-augmented teams charge. They are reading the same essays we are. The price they are willing to pay per hour of an unaugmented engineer has, in our anecdotal but consistent sampling across UK markets, dropped about 15-20% over the last twelve months for new contracts. Existing contracts are sticky for now but are repriced at renewal.

Put the two together. Cost-side productivity up 2×. Sale-side hourly price down 1.2×. The agency that used to make a healthy margin on the spread is now being squeezed at both ends simultaneously. The margins that were 40% on the build in 2023 are 12% on the build in 2025 if you do nothing.

12% is not a viable margin for a service business carrying a 35-head overhead. It is barely break-even. Most agencies are not yet running the maths on this honestly because the lag in their existing contract book is masking the trend. The reset will arrive at renewal cycles in 2026 and 2027.

Why the obvious response will not save most of them

The obvious response, and the one we hear most often in the bars at industry conferences, is: we will just adopt the AI tools ourselves, do the same work in fewer hours, and pocket the productivity gain.

This is correct in principle and unworkable in practice for three reasons.

First, the contract structure. Most agency contracts are still priced and scoped on an hours basis, either explicitly (day rate × estimated days) or implicitly (fixed price with a built-in margin assumption that bakes in hours). If you ship in half the time, the client sees the gain, not the agency. The agency has to reprice every contract to capture the productivity, and most agency owners do not have the appetite for the renegotiation. Even if they do, the next agency in the procurement process will undercut them. The race to the bottom is the equilibrium, not the exception.

Second, the headcount problem. An agency at thirty-five staff cannot, in the short term, ship the same work with seventeen people. Even if the productivity gain mathematically permits it, the organisation does not. You have line managers, client relationships, account leads, dependencies. You can do twice the work with the same headcount in theory; in practice you will get perhaps 1.4× because the seams between people slow you down. Meanwhile your fixed cost stays the same.

Third, and most importantly, the talent geometry. The productivity gain from AI tools is not distributed evenly across your team. The top decile of developers, the ones who already had good engineering taste, get a 3-4× lift because they can now command tools that match their judgment. The middle of the distribution gets perhaps 1.5×. The bottom of the distribution, the people who were always struggling to keep up, get almost nothing — sometimes they get worse, because they cannot tell when the model is wrong. The top of an agency’s payroll becomes much more valuable. The middle and bottom become harder to defend at any price.

Net effect: the agency has to compress, but the compression hits the politically hardest parts of the firm. The owners who have been with the same team for ten years will not, in our experience, run that compression cleanly. They will defer. They will hope the market comes back. The market is not coming back.

What is happening instead

Three patterns are showing up in the agency landscape, and we expect all three to accelerate.

Pattern one: small, sharp, partner-led. A growing number of two-to-six person teams of senior operators are picking up the work that used to require a thirty-five-head firm. The economics are exceptional. Five very senior operators with the right tools can deliver what twenty middling staff used to, at higher margin, on faster cycles, with no client relationship dilution. We are seeing these teams price in the £200k-400k range for projects that would have been £150k-£250k from a traditional agency, and the clients are happily paying the premium for speed and seniority. This is the future of bespoke build for the next five years. Most of these teams have started in 2024-2025 and are growing.

Pattern two: product-shop pivots. Some agencies are quietly retooling toward becoming product shops. They take a slice of equity in everything they build, hold the IP where they can, and use the agency work to fund a small portfolio. This is the path Moonlabs sits on. It is hard and slow and requires capital tolerance, but it is structurally robust. The output is a balance sheet, not a backlog.

Pattern three: AI-flavoured commodity. A larger group is doubling down on volume. They market themselves as “AI-native agencies,” price on hours, push to do twice the work with the same staff. This works for about eighteen months until the market price for AI-flavoured commodity catches up with the rest of the market and the margin collapses again. It is the longest tail and the most painful to watch.

The agencies that will survive into 2030 are almost entirely in pattern one or pattern two. The pattern three firms will, in the main, be acquired by larger consultancies for their client lists and their senior staff, with the rest of the headcount written down.

What this means for buyers

If you are on the buy side — a CTO, a head of digital, a startup founder — the implication is direct. The hourly day-rate is no longer a sensible unit to procure software in. You are paying for an input that has been disconnected from the output by AI tooling. You should be procuring on outcomes, with sharp time-bounded scopes, with a heavy preference for very small senior teams over large mid-weight teams.

Cheap warning sign: any vendor whose first response to your brief is to scope it in days × roles. They are using a unit of measurement that has stopped working.

Better signal: a vendor who scopes the work in “a working version of X by date Y, paid against delivery, with named senior operators on the build.” You want that vendor. Pay them more. The total cost will be lower and the outcome will be better.

What this means for the people inside agencies

If you are reading this and you are inside one of the agencies that will not make it, the time to plan is now, not in eighteen months. The talent market for the top decile of agency operators in 2027 is going to be unusually liquid in a bad way: lots of mid-career people landing in the market at once, depressing wages for everyone except the very top.

The move is to get yourself into the top decile while the rest of your firm is still pretending nothing has changed. Get fluent on the modern toolchain. Pick up the commercial and fundraising skills that the AI tools do not yet automate. Start writing publicly so you have a network independent of the firm’s client list. Build optionality.

The agencies that will be quietly acquired in 2027 do not yet know they are the agencies that will be acquired in 2027. The people inside them do not have to share the firm’s fate. The career path runs through the firm, not in it. Use the next eighteen months accordingly.

Louis O’Connell-Bristow is co-founder of Moonlabs, the operator-led AI incubator and Academy. He spent the early years of his career on the inside of UK digital agencies before founding the businesses that became Homemove, home.co.uk and homedata.co.uk.