What a lean agency actually costs: four years of books

I was wrong about my own books

Last week I laid out our entire marketing system. What people came back with was not how it works, but what it costs. Or rather: what it no longer costs.

So I opened my own books. Four years, in percentages, because the amounts are mine and not that interesting.

I knew the answer before I started. I had cancelled a pile of subscriptions, so that had to be it. It was not. The line I expected turns out to be a footnote, and the line that actually moved was not one I was watching.

What I thought the answer was

I paid for an absurd amount of software.

Typeform, for intake forms that looked like us. Google Forms would have worked, but a branded form genuinely changes how a discovery session lands. ClickUp, billed per person per year, and there were twelve of us in there. When we moved off ClickUp, Slack Pro took its place, also per person, also times twelve. Zapier, to tie Google Forms and ClickUp together. Google Drive. And WeTransfer, which I am only remembering now.

That list felt like the fat. Cancel it, and your agency gets lean.

I checked it against the books, and my memory was right: the software line was exactly that list. But as a share of revenue, the whole line came to under two percent. All of it together cost less than one average project.

That is the first lesson, and it is an annoying one. The subscription you watch leave your account every month feels like the cost, because you see it every month. The cost you do not see, because it arrives per project, is a multiple of it.

What the answer actually is

The line that collapsed is not software. It is bought-in services. Everything I used to outsource: freelancers, photographers, bought-in production.

As a share of revenue:

• 2023: 27 percent
• 2024: 24 percent
• 2025: 33 percent
• 2026 through September: 7 percent

That is the shift. Not my subscriptions, but the work I no longer have to hand out.

Product photography is the clearest case. We used to do a lot of it and now we barely do, because the image models have got good at it. Not at everything. An on-location brand shoot is still an on-location brand shoot. But a product shot that used to cost a studio day is now an afternoon's work for someone who knows what they want to see.

And the honesty has to come right here, or none of it holds: my revenue is lower this year than last. Roughly half of the same period in 2025, which was a record year. Less revenue also means less buying in. So part of that fall from 33 to 7 percent is arithmetic, not achievement.

What is an achievement is the ratio. Total costs as a share of revenue:

• 2023: 50 percent
• 2024: 41 percent
• 2025: 45 percent
• 2026: expected to land around 35 percent

That last figure counts in the purchase invoices I still have to book and the work that is all but signed. I am flagging it as an estimate. But the direction is solid, and it holds while revenue is lower. That is the number I can actually use.

Why I started building our own software

Every project management tool for agencies is built by people who do not run an agency. You get tasks, boards, timesheets. What you do not get is somewhere for a client to land.

So we built that. In our own software every client has their own channel. Files go into their own folder, not into a WeTransfer link that expires next week. Feedback arrives as annotations and boxes drawn straight onto the design, instead of an email saying "the second one from the top could be a bit bigger". Moodboards we build together, their brand kit lives in there, and the timeline is simply visible to them.

There is also a small app store in there with things a business owner genuinely uses: removing backgrounds, upscaling images, pulling a colour palette out of a photo, planning an Instagram grid, shortening links, making QR codes. Small things, but every one of them used to be a separate subscription.

And there is an agent running inside it, Dune, which knows the playbooks and services you put in. It can send invoices, post messages, get files ready.

Was I nervous when I cancelled the rest? I have asked myself that. We had built something ourselves and it had to hold up in practice. But it went fine, for a simple reason: it got better for the client, not harder. There has not been a single complaint about the platform. When we onboard someone we just show it in a video call or send a Loom, and within two minutes they understand where they are going to land.

The task that disappeared at eleven at night

Project management.

Not one task, all of them. Creating a project, putting people on it, walking through the brief, setting up the timeline, messaging the client about what is going to happen and how long it will take.

I do that by talking now. I send a voice note to my agent system: the invoice is approved, the client is in the software, create a channel, go through the brief, add the people, set up the tasks, show the timeline and send them a warm welcome. That system runs locally, knows my business, manages the website and the codebase, does the research, knows our positioning and our client types. It has its own account in our software, so it can talk to clients the way a member of staff would.

Do I still make invoices? Sometimes. Do I check everything before it goes out? Always. But project management is over, and that has changed my working day. I spend more time on concept, on design and on selling, and less on things that stop being important the moment a system can do them.

What I deliberately do not automate

The design.

When we build a brand, we believe taste is what makes the brand. That is not a slogan, it is a working rule. The first real contact with a client, the discovery, is exactly where you should be sitting and talking about where this brand is going to land, who it is for, what we need to ask. The more input there, the better everything downstream.

There are good design systems. But as humans we are still more creative and have better taste than the models available today. So that part is not going anywhere.

What has changed is how we build. You are not dragging things around in a website builder any more. You talk to systems and you check whether the taste actually made it through.

And a decent list is still done by hand: product and lifestyle photography where it matters, brand shoots, shooting commercials, most of the packaging we design, online manual structures, a lot of listing design, 3D modelling, and all custom Shopify development.

Of course we use AI to work faster. But faster does not mean better looking, and it does not mean right. Everything gets checked and validated on top. What we get back is that we can handle a bit more volume, and that we can verify whether the work holds up more quickly.

The line that went up

Because a story like this is only believable if something also gets more expensive.

AI credits and cloud. My system runs in the cloud and that costs real money, and that line is not shrinking. I also still pay for Adobe, because I want to be able to design by hand, and for a handful of image models. The last one partly out of curiosity: I want to know what is happening in that corner, even in a month when I do not use it.

That is the trade. I swapped a stack of fixed subscriptions for a bill that moves with how much I run. That is not a saving, that is a different shape.

Where it gets too thin

I have an honest answer here that does not flatter me: I do not know, because I have not hit it yet. Everything I removed, I removed because something better had taken its place, and so far I do not regret any of it.

But I also know why, and it is not cleverness. It is that I never cut anything just to cut it. Every time it was: this thing has been replaced, so it can go. Do it the other way round, cancel first and hope it works out, and you get to write a different article.

Three things I would take with you

If you are reading this with your own books in mind, these are the three that made the difference for me.

Look at what you buy in, not at what you subscribe to. You watch your subscriptions leave every month, so you assume they are the problem. What you buy in per project you never see as a line, because it sits scattered across invoices attached to jobs. Add a year of those up. For me it was a multiple of all the software put together.

Never replace something before its successor is running. Not one of my cancellations was a cost-cutting decision. Every time it was: this thing has been replaced, so it can go. Cancel first and build after, and you build under pressure, and it shows.

Automate the building, not the taste. That is not a matter of principle, it is a matter of results. Building can go faster without getting worse. Taste cannot. The moment you confuse the two you ship work you do not stand behind, and eventually that costs you the client.

What I say when a client asks

"So you let AI do my work."

No. I do not let AI do the work. I am the one doing the work.

I used to have to perform the actions myself. Now a computer performs them. But I am just as involved, and really more so, because I can test more, try more, push a concept further and deliver faster. I steer systems and taste in the right direction. If what comes out is not what it should be, we find another way.

We are a creative agency. We believe in a personal experience from start to finish: a lot of calls, a lot of contact, actually meeting up. Human to human.

So that people can keep working with people, in a world slowly being taken over by robots.