AI cuts the cost of agency work. Clients and platforms keep the difference.
Adobe has a label for the moment an agency stops knowing what its work costs. When a Firefly Graph workflow can't be priced in advance, the credit estimate next to the Run button gives way to two words: "Credits vary." Pressing Run counts as the user's acknowledgment that the final total may come in higher or lower than a comparable fixed-cost workflow. Adobe also suggests keeping an eye on a run once it starts, to catch a workflow looping more often than planned.
I find that label more honest than most of what gets said about AI and agencies. It admits that the cost of AI work is set at runtime, and nobody knows the number until the machine is done.
The price of agency work is set somewhere else, in a negotiation that already happened or is coming back at renewal. The agency's margin lives in the gap between those two numbers. AI makes the work faster, and that part is real. Whether the agency keeps any of the savings depends on two parties who want them too: the client who pays the fee, and the platform that meters the work.
AI can make an agency faster and poorer at the same time. The agency is usually last in line to collect.
Twenty percent of what?
Forrester's 2027 predictions put numbers on the pressure. Its October 21, 2026, release predicts that rising AI costs will force B2C marketing leaders to scale back a third of their AI initiatives.
The reason is pricing: as martech vendors move to usage-based models, "every agent interaction, generated asset, and automated workflow becomes a billable event."
Forrester's agency report goes further. As MediaPost reported, Predictions 2027: Marketing Agencies expects 20% of the world's estimated 50,000 independent agencies to face acquisition, downsizing, or closure in 2027. The full report sits behind Forrester's client paywall, and the public accounts don't show how the percentage or the population estimate was built. These are forecasts, and I read them as a claim to test against how agency economics actually work.
MediaPost's headline says the sector will shrink by 20%. That's tidier than the forecast. Downsizing leaves a smaller agency that is still in business. An acquisition takes a firm off the independent list but can preserve its work and its people. Only closure ends the business, and nobody has said how much of the 20% belongs to each outcome.
For the person who built the agency, those differences are the whole story. A sale can protect every job and still hand the future returns to someone else. A smaller firm can survive while its owner loses the freedom to choose clients, or to decide how much time the work deserves. What is at stake is independence, and the headline number doesn't measure it.
The bill arrives after the work
A demo ends when the image appears on the screen. An agency's job ends when the client approves something it can use. Most of the cost sits between those two points.
Salesforce shows the shape of it. Under the Flex Credits model on its Agentforce pricing page, which I checked on October 28, credits sell at $500 per 100,000, and the worked examples price a standard action at 20 credits, about ten cents. Each example comes with the same disclaimer: it excludes other charges, Data 360 credits among them. Salesforce also sells per-user add-ons with unmetered employee usage, so what a buyer pays depends on the arrangement it actually runs. Either way, the unit Salesforce bills is the action, and no client ever bought an action.
Adobe shows the same thing from the creative side. On eligible plans, standard generations are unlimited, but inside Firefly Graph they stay metered.
Adobe gives two reasons some workflows can't be priced in advance. A step can produce variable output, like a language model whose cost depends on how long its answer turns out to be. Or a workflow can run a step a variable number of times, for example once for every object a segmentation step happens to find. A process that fans out into more work burns more credits than it looked like it would.
Now run a product launch through that system. Some generated variations are technically fine and wrong for the brand. Others need another round after the client changes the offer. Where generations are metered, credits are spent when the asset is made, so the rejected candidates cost money too. The unit that matters for the agency's own costing is the approved asset, and it carries the price of everything that didn't make it.
Integration adds a second layer. Connecting a client's product catalog is a one-time build if the connection gets reused across many assignments. With a client-specific system that changes every quarter, it turns into a standing maintenance job. Both get called automation, but only the first one stays cheap.
Then there is review. AI can cut drafting time enough to pay for the supervision it needs. It can also push work toward a senior reviewer whose hours cost more and whose capacity doesn't scale. The people checking the output are part of production, and leaving their judgment off the bill is the easiest way to make AI savings look bigger than they are.
Forrester makes the same point, more bluntly. According to MediaPost, the agency report argues that most agencies buy AI as SaaS, which keeps the actual spend on tokens, usage, and GPUs out of view. Once providers shift to usage-based pricing, Forrester expects agencies and CMOs to find the AI bill higher than the payroll it was meant to cut. It hasn't published an agency-level calculation behind that. An agency doesn't need the prediction to come true to get squeezed. It only needs to give up more in fees than it saves in delivery.
Ten percent
Here is how that happens, with invented numbers. An agency charges $20,000 for an assignment and spends $12,000 delivering it, leaving $8,000 for overhead and profit. AI, including the new tools and the extra review, brings delivery down to $10,000. If the fee holds, the agency keeps $10,000.
The client knows AI is involved, though, and Forrester reports that cost efficiency is the top benefit clients ask agencies for. Suppose the client negotiates the fee down to $16,000. The agency now keeps $6,000. It cut its delivery cost and lost a quarter of the money that pays for everything else.
The billing model decides how fast this happens. Under hourly billing, doing the same work in fewer hours simply means billing fewer hours. A fixed fee lets the agency keep the gain for the life of the agreed scope, and then renewal hands the client a fresh chance to claim its share. Value-based pricing leaves more room, but only for an agency that can show a contribution the client accepts as worth paying for.
Freeing time is also different from saving money. Salaried people stay on the payroll. Hours saved on one assignment become cash only if they go into more profitable work or replace an expense the agency would otherwise have paid. A lighter workload and better creative decisions have real value.
The same hours just can't be counted as a payroll cut, new revenue, and higher quality all at once, and agencies have to decide which of those they are actually buying.
The platform takes its share
Forrester's answer for agencies is orchestration. Analyst Jay Pattisall's October 28 blog post opens with "No, AI won't kill agencies, but it will divide them." On one side are orchestrators that help CMOs redesign how marketing runs. On the other are executors with specialist skills. Forrester predicts that proprietary marketing operating systems will carry half of global holding company revenue and service delivery in 2027. Midmarket and independent agencies, it expects, will partner with or white-label third-party systems from Adobe, Canva, Google, Hightouch, Pixis, and Zeta.
Read that list next to the Firefly Graph button. Adobe is one of the platforms Forrester expects independents to build their orchestration on. It is also the vendor whose most flexible workflows can't be priced until they finish running.
The appeal of orchestration is obvious. An agency that shapes how work moves through a client's organization matters more than a supplier taking production orders.
The responsibility grows with the role, though. Connecting a workflow means owning how it behaves when inputs change or approvals fail, and selling that responsibility takes people who can maintain the system and a contract that pays for the ongoing work.
The incentives need a hard look too. An orchestration offer that earns its money by adding tools and passing through more consumption has little reason to simplify anything.
A client looking for savings should ask how the agency benefits when the workflow gets cheaper, and the agency should have an answer.
Dependency is where the economics turn into a question of power, and it runs in both directions. If campaign histories and approval practices live inside the agency's environment, moving the account means more than hiring a new creative team. The client would have to rebuild how its marketing gets done. That gives the agency leverage, and the leverage can sit right alongside genuinely better service.
Upstream, the same logic works against the agency. An independent running a white-labeled operating system owns the client relationship while depending on someone else's pricing and technical choices. It can become essential to its client without gaining any matching freedom from its supplier. Owning the interface doesn't settle who controls the economics underneath it.
So the agency ends up in the middle. If the client can replace the agency easily while the agency can't replace its platform, the agency carries the responsibility and has little room to defend its return.
An agency with durable expertise, portable records, and real choices about its suppliers sits in a different position. Adopting AI tells you nothing about which of these an agency is.
Who still gets paid
The October 28 post also leaves room for execution-focused agencies whose specialist capabilities cover work AI can't automate or can't execute authentically. In the same breath, Forrester predicts that enterprises will bring up to 60% of templated content production in-house: blog posts, social posts, image and video variants. That one rests on more than a forecast. In Forrester's 2026 marketing survey, 67% of B2C marketing decision-makers already rank bringing more creative and production work in-house as a high priority for the coming year.
That changes the question. If the client can bring routine production inside, cutting the agency's cost doesn't save the assignment. What matters is whether the client still has a reason to buy from outside.
For a production business, the reason can be reliability, or a total cost lower than the client can reach on its own. That model can still work. But its technology has to serve work the client keeps paying for, and its price has to carry everything that sits around the automation.
Forrester doesn't settle the creative question. According to MediaPost, the report puts creative-only agencies alongside generalist and production shops as the groups most exposed to adtech platforms and to clients automating their own work. The October 28 post names differentiated creative thinking as one of the things the winners deliver. Both can be true. If AI makes variations nearly free, deciding which ones deserve to exist becomes worth more to the buyer. Whether the buyer pays for that judgment is a commercial test, and expertise alone doesn't pass it.
The generalist is in the worst spot. It keeps the cost of broad service coverage while losing the production volume that used to pay for it. Forrester's report, as MediaPost describes it, expects midsize generalists to see their margins squeezed further. A technology subscription doesn't fix that structure. It adds a metered line to it.
Before renewal
The scorecard I would want an agency owner to keep is short and uncomfortable. Could the agency take a lower fee at renewal without cutting the people responsible for quality? How much of its contribution comes from its largest client, and could the technology commitments tied to that account be unwound if the client left? Could it switch its critical platform without losing the ability to deliver?
Pricing power shows up as work clients keep buying when cheaper options exist. Independence shows up as the ability to change suppliers. Both are harder tests than showing that a team uses AI.
Forrester's forecasts will be graded in 2027, and agency owners have to make hiring and investment decisions before then. Faster work is real. It could pay for more ambitious creative work and a more resilient small agency.
But after the client has taken its discount and the platform has taken its credits, the question is whether enough is left to pay for the judgment the agency exists to provide.