The part you used to pay for got cheap
If you are still buying a growth agency the way you bought one in 2022, you are paying a premium for the thing that got cheap. Production got cheap. A landing page variant, twenty ad headlines, a first draft of a launch email: minutes now, not days.
What did not get cheap is knowing which of those twenty headlines deserves budget behind it, and being able to tell six weeks later whether that call was right. That work has not sped up at all. It got harder, because there is now far more output to judge and the same number of people qualified to judge it.
Where the money used to go, and where it goes now
Making the thing. Copy, layouts, variants, decks, first drafts of almost anything. This is the line item most rate cards are still built around.
Deciding what to make, refusing the rest, and reading the result honestly six weeks later. No tool has compressed this.
The practical effect is that the share of an engagement spent making things is falling and the share spent deciding and measuring is rising. Very few rate cards reflect that yet.
The hours model breaks first
An agency that bills you for time has a problem it did not choose. Its people got faster. If the day rate holds, the same work bills fewer days and revenue falls. So the rate goes up, or the scope grows quietly, or the monthly report gets longer. None of that is dishonest. It is what a time-based contract does under those conditions.
Your interest runs the other way. You want the outcome for less and you do not care how many hours sat behind it.
Four ways to be charged, and what each one prices
We price the third way, so read that as disclosure rather than neutral advice. The test still applies to us: ask what the number is, and ask what happens if we miss it.
What is actually scarce
Every agency you talk to has the same models you do. Worth saying plainly, because a lot of current agency positioning depends on you not noticing. The model is a utility priced per million tokens, and your competitor is buying from the same shelf.
Four things stay hard to copy.
The four things a model does not hand anyone
One number from our own work, since the argument needs one. On a Berlitz UK offer campaign, 21% of leads became paid enrolments. No model decided the two things that earned that: the closing date sitting in the page bar, and a WhatsApp route beside the form. Both came out of an argument about what the page was allowed to promise.
Seven questions worth asking now
The old agency selection questions still work. Ask who is actually on your account, ask to speak to a client whose contract ended, ask for the notice period. These seven are the ones the last three years added.
1. Which parts of this work will a model do, and who checks it before it reaches us? "None of it" is not the good answer. No answer at all is worse. You are testing whether anyone has decided where the review step sits.
2. Show me the same brief run through your process twice, six months apart. One good output proves very little now. Consistency across time is the thing left to test.
3. What will you know about our brand after month one, and can we take it with us? The tone rules, the prompts, the component library, the audience notes. Ask whether that is yours on exit. Most contracts are silent on it, which is not the same as it being yours.
4. What did you stop charging for in the last two years? Possibly the most useful question on the list. An agency that has repriced something has genuinely absorbed the change. An agency with an unchanged rate card took the productivity gain as margin, which is their right, and now you know what you are buying.
5. What do you think we should not spend money on? This one survived from the old list unchanged, and it is still the fastest read on whether an agency will push back once the contract is signed.
6. Who is accountable if a model writes something false about our product and it goes live? You want a name and a review step. A reassurance is not an answer.
7. How do you know an AI-assisted piece of work was good? If the answer is about time saved, that is their margin, not your result. Ask again for the number on your side of the relationship.
Buy a small piece of work instead of running a pitch
The competitive pitch was always a weak evaluation tool and AI finished it off. A deck costs almost nothing to produce now. Choosing between agencies on the strength of their decks is close to choosing on nothing at all.
Pay two or three of them a small fee for a real, narrow piece of work instead. One page, one campaign, one month. Four weeks of that tells you more than four rounds of presentations, and it usually costs less than the internal time a full pitch process burns.
A four-week trial that beats a pitch process
Where this argument is weak
If your real problem is volume, none of the above applies to you. Four hundred product pages in nine languages is a production problem, and buying production capacity by the hour is an honest way to solve a production problem.
The view here also comes out of a narrow slice of work: language education and events across Canada, the UK and Spain, where a lead converts in weeks. Enterprise software on an eighteen-month cycle behaves differently and I would not claim these conclusions carry over cleanly.
And the models keep improving, so any specific claim about what they cannot do yet has a short shelf life, including the ones written above. The part I expect to survive is the pricing argument, because incentives move more slowly than capability.
Magic Pulpo
A growth agency × startup studio, Budapest