AI made agencies cheaper to run. Labor-based fees hand that to the client.
Here is the trap I keep watching agencies walk into. They spend months and real money making a workflow faster. The work gets done with fewer hours. And then the client, quite reasonably, asks why the fee hasn't come down.
Nothing in that sequence is irrational. If an agency is paid for hours, headcount, or visible production effort, efficiency is a discount it hands out for free. The software subscriptions and token bills arrive first. The pricing question arrives later, if anyone asks it at all.
Digiday's 2026 Media Agency Report, published on October 5 by Catherine Wolf and Li Lu, never calls it a trap. But read its numbers side by side, and that is what they describe. Agencies are reporting savings. They are reporting costs. They are trying several ways to make money from AI. What I can't find anywhere in it is a settled answer to what the client should actually pay for.
A word on the evidence before going further. The survey covers 30 agency professionals, fielded in the third quarter of 2026, plus a focus group of five senior executives from Go Fish Digital, Horizon Media, Novus, Tinuiti and UM. Digiday doesn't publish question-level bases, geography, or weighting. This is not a market estimate. It is a snapshot from people who sit close to the money, and I find it useful for exactly that reason: they aren't speculating about AI in the abstract; they are describing their own P&L.
Savings everywhere, prices nowhere
Seventy percent of respondents said AI has saved costs by making tasks more efficient. Seventeen percent said it has saved costs by replacing roles. Digiday reports the two separately, so you can't add them up, and it puts no dollar figure on either.
The cost side is less comfortable. Forty-five percent said implementation cost about what they expected, 31% said it cost more, and 24% said less. That is self-reported experience rather than audited spend, but nobody in that sample is describing AI as free.
The costs are also the recurring kind. Digiday's anatomy of an agency AI bill runs through tool subscriptions, tokens, and the employee time spent using them. David Dweck, president of Go Fish Digital, said his agency builds dashboards that track token spend for every single employee, down to the cent. When someone builds something useful that is getting expensive, they move it onto native software so it stops eating tokens. Olivier Pepin, CTO at Novus, described the other half of the problem. The strategy team builds a tool, the investment team builds a tool, and eventually somebody in the center has to work out which of them are real and which were a waste of everyone's time.
I like both examples because they treat AI as an operating expense that needs a manager, not as a capability that justifies itself. An experiment can be clever and still fail the test of repeatable economics. Most of them will.
Then there is revenue. Fifty-nine percent said their company monetizes AI through new products. Thirty-four percent sell strategic roadmaps or audits for clients' AI products, and 31% charge an extra fee for AI added to existing services. Those three figures add up to well over 100%, so plenty of agencies are trying more than one route at once.
That is what experimentation looks like before a business model settles.
Value the client never sees
The most revealing example in the report isn't sold as a product at all. Dweck described using AI as client memory, feeding it meeting notes, call recordings, and task lists so that when someone leaves and an account changes hands, the client barely notices a dip in service.
That is genuinely valuable. It is also precisely the kind of value that never appears on an invoice, because its entire purpose is that the client doesn't notice anything happened.
The agency pays for it. The client benefits. The fee stays the same, or comes under pressure because the team behind the account now looks smaller.
Compare that with what Go Fish actually sells: generative creative, AI-built avatars, and feed optimization that pushes video and imagery into e-commerce and retail media environments to lift performance. Those have an obvious unit of value, because the client sees the result in a channel it already measures. UM's chief investment officer, Marcy Greenberger, described yet another use: synthetic audiences for quick simulated focus groups, which sits closer to strategy than to production.
Here is the pattern I take from this. AI that moves a number the client already watches can be priced. AI that protects quality, continuity, or speed behind the scenes has to be argued for, explicitly and in advance, or it becomes a free upgrade.
Clients want flexibility, except in one place
This matters more because clients are loosening their commitments. In the survey, 36% said client media budgets held flat in 2026, 32% said they rose, and 18% said they fell. For 2027, 39% expect flat budgets, another 39% expect growth and 21% expect a decline. Those are respondents' expectations, not a spending forecast.
The level isn't the story. The timing is. Greenberger told Digiday that more clients held money back at the May upfronts. Not necessarily because they were spending less, but because they wanted fewer long-term commitments and room to react during the year. Sarah Duffy, who runs integrated investment at Horizon Media, said budgets can now swing day to day and month to month depending on how a client's business is going.
Put that next to an AI cost base. The agency commits to subscriptions, token spend, and integration work up front. The client wants the freedom to move money at short notice. Someone carries the risk of that mismatch, and right now it is the agency.
One area where clients are clearly adding money is AI search. Eighty-nine percent of respondents said clients increased AI search spending this year, tied with streaming and CTV at the top of the list. Seventy-seven percent plan to raise AI search and GEO budgets over the next 12 months.
Where that money comes from is the telling part. Among those planning an increase, 40% said it would come from general marketing and 20% from traditional search. Digiday describes the 40% as a majority. That isn't true, and the other 40% remains unexplained. But even the partial answer shows that AI discovery has no natural home in the media plan. It competes across marketing and pulls from search at the same time.
Greenberger described clients working two problems in parallel: how to show up in AI search as advertisers, and how to get their own sites pulled accurately into the answers an LLM generates. An agency that ties those together, and can say which spend is moving and how success will be judged, has an offer. An agency selling GEO as a content service has a line item the next budget review will question.
Agents need more than press releases
I would be most careful with agentic buying. Harry Browne, who leads TV, audio and display innovation at Tinuiti, told Digiday the topic has "a lot more press releases than substance." In the same breath, he said Tinuiti is building agents for planning and audience discovery and setting up MCP connections to DSPs and SSPs while it figures out where the real value sits.
That second detail is the one I would underline. An MCP connection into a buying platform means an agent can reach systems that move client money. At that point, the commercial question and the governance question become the same. What is the agent allowed to decide? Who approves what it can't decide? What happens when it gets something wrong?
An agency that can answer that in writing has something to charge for. An agency that can't is selling a demo.
What I'd want to see in a 2027 plan
If I were reviewing an agency's AI plan for next year, I wouldn't start with adoption numbers. I would ask for one real workflow and its full cost, meaning subscriptions, tokens, and people. I would want a named owner and a defined point where a human reviews the output. Then I would ask what improvement the client agreed to in advance, whether that is continuity, performance, speed, or risk control, and how the agency gets paid for it: inside the retainer, as an added fee, as a product, or tied to an outcome.
Vague answers there mean the agency is funding a client benefit with no plan to recover the cost.
Governance belongs inside that delivery model, not on a slide at the end. Approved data, access rights, decision limits, and escalation paths are what turn a clever capability into accountable execution. They are also the strongest argument for the fee.
None of this depends on agency size. UM, part of Omnicom Media since Omnicom closed its acquisition of IPG last November, was the only holding-company agency in Digiday's focus group. A holdco's shared infrastructure matters commercially only if it produces a repeatable method a client can buy, with evidence it works. An independent's speed only holds up if a bespoke build survives past one client and someone has costed its maintenance. Either way, clients can license the tools themselves. What they can't buy off the shelf is an operating method that reliably produces a result.
Agencies have spent two years proving AI makes them faster. Next year they have to prove what faster is worth, and to whom. The ones that can name the capability, show its full cost, point to the client outcome, and explain the price will get paid for their efficiency. Everyone else will keep giving it away.