Every few days at the moment someone declares the agency model dead. AI does the work in a fifth of the time, clients are building the capability in-house, the whole billable model falls over. And I think a fair bit of that is real. Not all of it, but enough that I wouldn't wave any of it away.
It's just not the thing I'd be most worried about if I ran an agency right now.
What I'd be worried about is much less interesting than AI, and much closer to home. After a lot of years now sitting inside the systems creative businesses use to run themselves, the most common problem I see still isn't technological. It's that a lot of agencies are quietly lying to themselves about how their work is actually going. And the numbers they'd need to survive the next few years are the ones they've spent the last few years bending.
This isn’t meant to land as accusatory or up on my high-horse. I say it almost mechanically, and moreso with hope… because I don't think most of it is done cynically at all. But it's happening, and it's happening a lot.
Some of the shapes it takes, all of which are more common than you'd probably guess:
Hours getting moved off one job and onto another, because the first one is over budget and the second one still has room in it.
A percentage-complete figure that gets nudged to whatever makes the WIP report look survivable this month.
Timesheets edited after the fact — sometimes by the person who filled them in, more often by someone in charge of them — so the job lands roughly where everyone said it would.
Selling a client one thing in a quote and then delivering something completely different, and never going back to reconcile the two. The quote said four rounds. You did eleven. The job closed out at four. It promised Nikita’s time, but James did the majority of the work.
So much of this is actually just human emotions and behaviour. Almost every instance of it I've come across has a fairly human reason underneath. Nobody wants their name on the job that lost money. Utilisation gets read as a judgement on a person rather than a measure of a job, so people protect themselves from it, which is a completely rational thing to do. Account leads don't want to be the one who tells a client the scope moved, so they absorb it and hope it evens out. And somewhere above all of that, someone needs the monthly number to look like the number they promised, so the pressure runs downhill and gets paid for in small adjustments.
It's not dishonesty so much as a kind of collective politeness. Everyone knows the number isn't quite true. Nobody says it out loud.
And that works fine, right up until the moment you need your numbers to tell you something you don't already know.
Which is, I think, exactly the moment we're in.
Look at what the AI conversation is actually asking agencies to decide. What do we charge when the work takes a fraction of the time it used to? Do we move to outcomes or value pricing? Do we bill for the time it would have taken? Which parts of the work should we automate, and which parts are the reason a client pays us at all? Holding companies are already shifting a real chunk of their fees toward outcome-linked models, and the independents will get asked to follow whether they've worked it out or not.
Every single one of those questions is unanswerable without a truthful history of effort against revenue. You can't price an outcome if you don't know what it costs you to deliver one. You can't work out which work to hand to a machine if you don't know which work was eating your margin. You can't tell a client that the value you deliver is worth more than the hours it takes if you've never once measured the hours honestly.
So the scary version isn't AI making your work cheaper to produce. It's AI making your work cheaper to produce and you having no way at all of telling whether you kept any of that saving or handed the whole lot to the client for nothing. If your history is fiction, the future is guesswork. And guessing has been survivable for a long time in this industry, mostly because growth covered it. I'm not sure it will be for much longer.
The more hopeful way round, and the reason I keep going on about this, is that the agencies who do know are going to be in a much better position than the doom posts suggest.
Because once you can see effort against revenue truthfully, the same technology stops being a threat and starts being a lever. You can see which work is genuinely low-value and repetitive, and automate it without guessing. You can see which clients are actually paying you properly for the difficult thinking, and which ones have been quietly subsidised by the easy stuff. You can find the thing your team is brilliant at that you've never charged for, because it was buried inside a line item called "production." You can go to a client with something added rather than something discounted. Maybe that's the actual opportunity here — not doing the same work faster, but finally being able to see where the value was sitting all along.
None of which is possible if the timesheets are a work of fiction.
I'd be a bit of a hypocrite if I made this sound easy, by the way. I’ve seen these things myself, even done some of them at times. Not at Streamtime — before that, when I was on the agency side. That's nine years ago now, so I'm not going to pretend I'm reporting from the front line. But I’ve been in meetings where decisions were made to move hours between jobs. I told myself a job had landed about right when I knew full well it hadn't. I sat in month-end conversations where the number on the page was the number we needed it to be, and no-one put their hand up because they didn’t fancy having that conversation with the overlords.
And the thing that gets me now, nine years on, is how little of it has changed. The tools are much better. The habit is identical. Which probably tells you it was never really a tooling problem in the first place.
It's uncomfortable to fix, too. The first honest month usually looks worse than the dishonest one before it, and that's sort of the point – the first accurate report you run is meant to be a bit of a shock (reminder: that report is only accurate if the inputs are). Have a read of this post, and then this one on how to approach that conundrum.
But it's a much smaller shock than finding out in two years' time that you can't answer the only question that matters.
If your numbers can't tell you how your effort relates to your revenue, that's the thing I'd fix before I worried about anything else. It's less exciting than an AI strategy, and more likely to be the thing that decides how this goes for you.
Happy to talk about it properly if it's useful — this is more or less the whole reason Streamtime exists, and I'd rather have the conversation than sell you anything. Drop me a message.
.png)
