AI is already starting to hit the agency business, and the expectation is that it will hit much harder. Will this be the twilight of the business or the birth of new opportunities? Will it damage the bottom line, or will new revenue lines make up for the ones that disappear?
Our first estimate is that the most likely outcome is neither extreme. It is a slow erosion, hard to spot in a year-over-year comparison, and one that in our view is not corrected with more technology but with the right handling of three variables: price, costs and structure.
This piece does not promise that AI will improve an agency's profitability. It tries to answer something more useful: what you have to do so that it does not damage it, and where real room to grow is left.
What the evidence already shows
The consensus among the consultancies is not that AI has yet to arrive, nor that it has already changed everything. It is something more uncomfortable: adoption is massive and value capture is still uneven. Almost every company uses it; very few can show what they gained from it, although the impact is starting, little by little, to be tangible.
McKinsey finds that 88% of organisations report regular use of AI in some function, but only 39% report an impact on EBIT and close to two thirds fail to scale it. BCG is harsher: barely 5% obtain value at scale, while 60% get minimal or no material value. Deloitte qualifies this in the other direction, with 66% reporting productivity gains, although only 20% can already attribute revenue growth to AI.
What we are beginning to see is that the marketing budget is not growing to pay for AI: it is being reallocated. And part of that reallocation comes out of agency fees.
In the other knowledge industries
The industries that went digital before ours give early signals that contradict one another, which is itself a data point: nobody has, so far, a clean measurement of the effect.
Stanford observed a 19% drop in employment among 22 to 25 year olds in exposed occupations compared with their less exposed peers, linked above all to lower hiring. It is a descriptive finding: it does not prove causation.
Across 1.39 million freelance postings, Demirci, Hannane and Zhu found a 21% drop in exposed writing and coding assignments, and 17% in image creation. The important detail is in what remains: the assignments that survive are more complex and pay more.
That last finding is the one that most resembles our near future. Work does not disappear. What disappears is simple, isolated work, the daily production that today generates thousands of dollars in any agency, and what rises is the threshold a client has to cross before deciding that expertise is worth paying for.
The bullets are landing close
On the client side, no inference is needed any more. The moves have been announced and, in several cases, quantified by the companies themselves.
58% of the advertisers surveyed by WFA and MediaSense, more than 80 multinational organisations, link a wider AI rollout to a reduction in agency fees, and 74% want to align compensation with results.
Klarna says it cut spending with external marketing suppliers by 25% across translation, production, CRM and social, and attributes some 10 million dollars annualised to AI. Unilever says it produces product images twice as fast and at half the cost. And Coca-Cola committed to five years of cloud and AI with Microsoft for 1.1 billion dollars, marketing uses included: a figure that neither proves a fee cut nor is all advertising investment, but that tells you where the money is going.
None of these data points proves that fees are about to collapse. All of them point the same way: the client can do more alone, buys differently, and is already asking out loud why it should keep paying the same.
Productivity is not in dispute. Who gets to keep it is.
A simulator, to stop arguing in the abstract
To turn the argument into something concrete we built a simple model, with the most important and general variables of the agency business, that lets you simulate future scenarios and see their effect on operating margin (EBITDA). It exists so we can argue afterwards about which variables actually move the needle: the ones we will have to defend as an industry.
Open the simulator and run your own scenario →
It has seven assumptions and a single convention: today's fees are worth 100. That makes the figures comparable across agencies of any size, and everything that follows reads as «for every $100 you bill today». Two assumptions describe where you stand. The other five, what you think will happen by 2030.
A. Where you stand today
- Your EBITDA margin. Of every $100 you bill in fees, what is left after all costs.
- How much of your cost is fixed. What you pay even in a slow month: salaries, rent, software. The rest is variable. It is a share of your total costs, not of what you bill.
B. What you think will happen by 2030
- Change in what you charge. The average price of your mix, in real terms. Not your total revenue.
- Change in how many jobs you sell. Projects, fees, retainers: units of work, not money. The price change goes separately; entering it here counts it twice.
- Change in how many deliverables you produce. Everything you have to deliver. It includes what you produce because you sold more, and what clients ask of you for the same fee.
- Change in variable cost per deliverable. What you pay outside for each deliverable: freelancers, production, licences. Your own team working faster does not go here: it does not cut salaries.
- Change in your fixed costs. Real changes in payroll, rent and subscriptions, plus whatever AI adds to your structure. Working faster, on its own, does not lower salaries.
One clarification, because this is where intuition usually fails. «AI makes the work 35% faster» is productivity, and productivity does not cut costs on its own. It depends on who does the work.
- If freelancers do it, you pay less for each deliverable. That is a cost that falls: variable cost per deliverable.
- If your own team does it, you do not pay less: the salaries are the same. What you gain is free time, and with that time you can do two things. Produce more with the same people, or do the same with fewer people and cut fixed costs.
Both are your decisions, and neither happens by itself. If the model counted your team's speed as a cost that falls automatically, it would tell you AI doubles your profitability. It does not, and that is what this piece is about.
Every change is an average across the whole business, not across the part AI touches most. And the result always reads in two different numbers: EBITDA in money, how much is left for every $100 of today, and EBITDA margin as a percentage, how much is left out of what you bill in 2030. They can move differently, and in the scenarios below they do.
Two scenarios
The starting point is the same in both and it is an assumption of ours: a 20% EBITDA margin and fixed costs that are 75% of total costs. That is, for every $100 you bill today: $60 of fixed costs, $20 of variable costs, $20 of profit.
1. Base scenario
The one the model suggests as most likely reading the evidence above: price gives way, you sell a little more, you produce considerably more for the same fee, and AI makes each deliverable cheaper. Structure is untouched.
| Projection to 2030 | Change |
|---|---|
| Price | −15% |
| Jobs sold | +15% |
| Deliverables produced | +40% |
| Variable cost per deliverable | −30% |
| Fixed costs | no change |
Result. EBITDA goes from $20 to $18.15, which is 9.3% less net cash in your pocket. The margin goes from 20.0% to 18.6%, a fall of 1.4 points.
The arithmetic: you bill 97.75 (100 × 0.85 × 1.15); fixed costs stay at 60; variable costs go to 19.6 (20 × 1.40 × 0.70). You produce 40% more deliverables and each one costs 30% less: the extra volume ate almost the entire AI saving.
That is exactly the problem. A fall of a point and a half in margin sets off no alarm. It gets explained away by the economy, by a client that left, by the exchange rate. An agency can spend five years feeling that things are going more or less fine while the business comes apart underneath.
2. Pessimistic scenario
Our pessimistic estimate, based on the impact AI already has in software, commercial production and illustration: price falls further, you sell less, clients keep asking for more work at the same fee, and the agency reacts by cutting structure.
| Projection to 2030 | Change |
|---|---|
| Price | −20% |
| Jobs sold | −10% |
| Deliverables produced | +10% |
| Variable cost per deliverable | −35% |
| Fixed costs | −20% |
Result. EBITDA goes from $20 to $9.70, which is 51.5% less net cash in your pocket. The margin goes from 20.0% to 13.5%, a fall of 6.5 points.
The arithmetic: you bill 72 (100 × 0.80 × 0.90); fixed costs come down to 48; variable costs go to 14.3 (20 × 1.10 × 0.65). Notice that the margin «only» loses six and a half points, but the money is cut in half: when revenue falls, the percentage hides the size of the hole.
The variables do not weigh the same
With the arithmetic in place you can ask something more interesting than the result: which part of the result each assumption produced. Start from today and move each variable by 10%, one at a time, with everything else still.
Today: fees 100 · fixed 60 · variable 20 · EBITDA $20 · margin 20.0%. The sales row moves two variables on purpose: selling 10% more jobs forces you to produce them.
| If only this happens, 10 points | Fees | Fixed | Variable | EBITDA | Margin | Effect |
|---|---|---|---|---|---|---|
| Your price falls 10% | 90 | 60 | 20 | $10 | 11.1% | −8.9 pp |
| You cut fixed costs 10% | 100 | 54 | 20 | $26 | 26.0% | +6.0 pp |
| You sell 10% more jobs, and produce them | 110 | 60 | 22 | $28 | 25.5% | +5.5 pp |
| Clients ask 10% more work at the same fee | 100 | 60 | 22 | $18 | 18.0% | −2.0 pp |
| Each deliverable costs you 10% less | 100 | 60 | 18 | $22 | 22.0% | +2.0 pp |
pp = percentage points of margin against today's 20%.
A 10% fall in your price takes 8.9 points of margin away; a 10% cut in your fixed costs gives 6 back. They are comparable magnitudes, and that makes the real tug of war visible: one force pushes down and the other pushes up.
The two things that move your margin most are your price and your structure. And selling more without adding fixed structure also weighs: selling 10% more jobs, and producing them, adds 5.5 points. Fees go up and only variable costs go up with them; the fixed ones get diluted.
Now look at the pair at the bottom. Being asked for 10% more work at the same fee costs you 2 points, and each deliverable costing 10% less gives you exactly 2 back. Throughout our history, that first movement is the main source of margin decay in agencies: more deliverables at the same fee.
If the production saving is the defence against falling prices, it is worth knowing how much saving would be needed. The rule is simple: to hold the margin, total costs have to fall in the same proportion as revenue. If you bill 15% less, you have to spend 15% less. With 75% fixed costs, an AI that cuts the cost of each deliverable by 30% contributes 7.5 of those 15 points; the other 7.5 can only come out of structure: a 10% cut in fixed costs. If revenue falls 25%, structure has to come down 23%.
In other words: no single variable is enough. Holding the margin requires moving everything at once, price, sales, cost per deliverable and structure, and each of those levers does part of the work. If one does not move, the others have to move more than is realistic.
Three decisions
Three decisions follow from that ranking. Each figure measures what the margin gains against the base scenario (18.6%), that is, against letting things happen. They do not add up: each one is the effect of moving that lever alone.
1. Change the unit you charge for · +8.6 pp
If you charge by the hour or by the deliverable, AI is, and will increasingly be, a machine for destroying your own price: every productivity gain automatically turns into a discount you hand to the client.
The way out is to charge for value: by result, by access or by licence. It is the only way for productivity to stay on your side of the counter, or at least to defend a part of it to offset the falls.
Defending the price from −15% to −5% is worth 8.6 points. Aligning what you produce with what you sell, so production grows as much as jobs sold and no more, adds another 3.2: when every deliverable is billed, the gap between production and sales stops existing.
2. Sell the freed capacity, do not give it away · +12.1 pp
Going from +15% to +35% in sales dilutes a fixed structure that represents 60 of every 100 you bill. It is exactly the same capacity that in the base scenario goes into producing a fifth more work for every peso billed.
The extra work is not free even if AI produces it fast. Either it gets billed, or it turns into new revenue. Giving it away is the quietest way to lose margin.
In an industry used to pitching and giving work away for free, this is a central shift in mindset for its survival.
3. Attack the structure, not the production · +6.1 pp
It may sound counterintuitive, but it is what the numbers say: cutting fixed costs by 10% is worth 6.1 points, against 2.9 for squeezing another ten out of the cost per deliverable. Variable costs are 20 of every 100 you bill; structure is 60, and in the base scenario it does not move.
It is not necessarily about layoffs. It is about turning fixed into variable, and about growth no longer meaning automatically adding structure.
With one warning that continuous numbers hide: structure moves in steps. With fixed costs of 60 against 100 of revenue and a team of nine people costing 45 of those 60, each person is worth 5. There is no −5%; there is 0% or −8.3%. If AI frees up the equivalent of 2.6 people, you let two go and keep paying for 0.6 of idle capacity, or you let three go and come up short at the first peak of work. The rounding goes down and it arrives late. Placing the freed capacity, by contrast, moves in small increments and is reversible: it needs someone to buy it, but it does not force you to bet in blocks of a whole person.
The levers interact and no agency executes all three at full size. The first two are twice the third in impact, but they are precisely the two that do not depend on you alone: price is negotiated and freed capacity needs a buyer. The only one you can execute on Monday without asking anyone is the one that gives back least, and that does not make it any less important. It is not about picking one: you have to move all three. None of them, alone, holds the margin; together, they do.
No lever is enough on its own. You have to move all three.
The game is there to be played. In 2030 we will be able to see how we played it.
Sources
Information cut-off: September 2026. Each entry states what was taken from the document and its declared limits.
- McKinsey, The state of AI in 2025. Survey of 1,993 respondents. Source of the 88% reporting regular use, the 39% reporting EBIT impact and the two thirds that fail to scale. It is self-reported intent and adoption, not a measurement of results.
- BCG, The Widening AI Value Gap, 2025. The 5% obtaining value at scale, the 35% scaling and the 60% with minimal material value.
- Deloitte, State of Generative AI in the Enterprise, 2026. The 66% reporting productivity or efficiency gains and the 20% already attributing revenue growth to AI.
- Stanford Digital Economy Lab, Canaries in the Coal Mine?, August 2026. Using ADP payrolls: −19% in employment of 22 to 25 year olds in exposed occupations against less exposed peers, linked above all to lower hiring. Descriptive; it does not prove causation.
- Demirci, Hannane and Zhu, Who Is AI Replacing? (CESifo). 1.39 million freelance postings: −21% in exposed writing and coding assignments, −17% in image creation. Surviving assignments are more complex and better paid.
- WFA and MediaSense, Future of Media Agency Remuneration. More than 80 multinational organisations: 58% link the AI rollout to a reduction in agency fees, 74% want to align compensation with results. These are declared expectations, not executed cuts.
- Klarna, press release. A 25% reduction in spending with external marketing suppliers, across translation, production, CRM and social, and some USD 10 million annualised attributed to AI. Result declared by the company itself.
- Unilever, press release, 2025. Product images at twice the speed and half the cost with digital twins. Declared result; specific cases show wider variation.
- The Coca-Cola Company and Microsoft. A five-year strategic partnership in cloud and AI for USD 1.1 billion, with marketing uses included. The figure neither proves a fee cut nor is all advertising investment.
