AI Won’t Kill Agencies. It Will Kill the One-Off Production Model.

01 / Ask the better question

Will AI replace creative agencies? The evidence says the question is too simple.

The extinction story makes a good headline. It is less convincing when clients are asked directly.

In a 2026 IPA and Tracksuit study of 200 senior brand-side decision-makers, nine in ten said agencies were central to commercial success or contributed meaningfully to it. Nine in ten also expected their agency use to increase or remain stable over the following three years.

Quality of asset production was valued by 83%, strategic brand expertise by 80%, and specialist channel knowledge by 79%. Clients still want the work and expertise. They are less likely to value a production model merely because many hands touched the file.

AI adoption is already too broad to treat as an optional agency specialty. The IAB’s 2025 digital video report found 51% of buyers already using generative AI to build or enhance video ads, with another 34% planning to use it.

The client does not need an agency merely to open the same model. Agency value sits around the model: turning an incomplete brief into a viable route, protecting brand context, choosing useful variation, catching the expensive exception, and remembering what worked when the campaign returns.

02 / Price the whole route

The AI discount is real—and usually calculated badly.

Creative demand is expanding into more channels, formats, audiences, and markets while parts of execution become cheaper.

Adobe’s 2025 survey of more than 1,600 marketers found 70% working at organizations that produced at least 1,000 assets a year. Nearly half said creating, reviewing, approving, and activating one piece could involve 51 to 200 people across the workflow. Another 58% said reviews and approvals consumed more than 40% of their time.

That is not only a generation problem. It is a coordination problem wearing a creative hat.

Yet the IPA’s The Price Isn’t Right found time-and-resource pricing still dominant. Only 27% of agencies believed they were paid fairly, while 58% reported little or no progress in changing commercial agreements.

If an agency sells 100 hours and AI reduces the work to 60, efficiency can look like a reason to invoice less. The agency improves its operation and weakens its own scope at the same time.

Hiding the efficiency is not the answer. Neither is billing imaginary hours. The repeatable route, accumulated brand context, and the ability to handle exceptions can create value beyond execution time—if the agency can show them.

That does not require one universal pricing revolution. Depending on the risk and scope, an agency may use output pricing, a subscription, a resource pool, workflow licensing, an outcome component, or a hybrid. First expose the economics. Then choose the contract.

03 / Compare operating models

Modern agencies are already selling more than labor.

Look at what the newer creative-service companies package. Superside’s public pricing starts at $15,000 a month on an annual term: at least $180,000 over twelve months. Its dedicated AI-native team begins at $30,000 a month, plus a $1,000 platform fee—an annualized minimum of $372,000.

These are calculations from published minimums, not average customer budgets. What matters is the package: recurring production budget, team, collaboration platform, brand context, project management, and AI workflows. The client is not buying a pile of AI credits.

Monks takes the idea further by positioning proprietary orchestration and human governance as part of a managed service. In its Hatch case study, Monks reports three videos and 60 ad variants with 50% fewer production hours and 97% lower costs than its legacy approach.

Those are vendor-reported results. Monks does not disclose the absolute production budget or enough detail to reproduce the comparison. The percentages are evidence of one case, not an industry promise.

A subscription alone does not make a production model workflow-native. Design Pickle, for example, scales creative servicesfrom two reserved design hours per business day to twelve or more. Access changed; the underlying unit is still human time.

04 / Fix the denominator

750 generated images are not 750 delivered assets.

One Superside project makes the measurement problem unusually clear.

While exploring a visual direction for Independence Pet Group, the process generated more than 750 illustrated pet images. A separate account of the project lists the final set: 14 illustrations, eight virtual backgrounds, and four sticker designs. Superside reports 11.5 hours of work, 90% less design time, and an estimated $15,000 to $20,000 saved against traditional illustration.

The impressive number is 750. The commercially useful number depends on the question: generated candidates, final components, an approved visual system, or one completed project.

Cost per generation cannot defend an agency scope because the client is not buying generations. The client is buying work that survives direction, selection, refinement, brand checks, and delivery. Our guide to review-ready output covers that production denominator; the agency question is how much of the route can improve and run again.

05 / Test the operating asset

A prompt library is not workflow capital.

Prompts, templates, and checklists can help. Saving them in a shared folder does not make them an operating asset. A connected AI ad workflow becomes capital only when the agency can run it again without losing the boundaries, numbers, and decisions around it.

Here, capital is an operating metaphor—not a balance-sheet category or automatic legal IP claim. The test has three parts.

Portable means the structure travels, not the client data

Reuse how an approved input enters, which steps run, what gets checked, and where a person steps in. Keep each client’s source material and brand context inside its own boundary. Copying last month’s folder and hoping nobody left the wrong logo inside is a habit, not an asset.

Attributable means the economics are visible

Trace model cost, elapsed time, human intervention, revisions, failed runs, and client-ready output. Without attribution, “AI made us faster” is office folklore. Pawook can expose production cost; an agency still has to compare it with the contracted fee in its financial system before calling the workflow more profitable.

Governable means the exception has an owner

A routine variation can continue while a changed claim, right, or client promise may need to stop. The workflow should preserve what changed and who made the consequential decision. The detailed approval route can vary; visible ownership cannot.

Some valuable work should fail the portability test. A new strategy is not defective because it cannot be turned into a reusable route.

Workflow capital belongs around recurring production: adaptations, localization, controlled variation, assembly, and other processes where the route repeats. It should not flatten the original judgment that gives the route a reason to exist.

06 / Keep the consequence human

Agencies may survive while roles disappear.

There is a comforting story where AI handles dull work and everyone floats upward into strategy. Real organizations will be messier.

The IPA Agency Census 2025found agency employment down 6.8% year over year and creative and non-media employment down 14.3%. Eight percent of agencies said they had reduced staff directly because of AI; 24% expected to do so in the following year. Employment among people aged 25 and under fell 19.2%.

If repetitive production was where junior people learned to read a brief, spot weak craft, and understand why a senior rejected an option, removing the task can also remove the apprenticeship. Agencies need deliberate training in direction, evaluation, workflow design, and exception handling.

Shared tools can also raise the floor without creating a distinct point of view. In a randomized experiment involving 293 short stories and 600 evaluators, Anil Doshi and Oliver Hauser found AI assistance improved average novelty and usefulness ratings while making stories more similar to one another. It was a fiction experiment, not an advertising-effectiveness study, but the warning is useful.

Human advantage is not moving pixels more nobly than a machine. It is deciding what should exist, recognizing when polished work is wrong, challenging a convenient brief, and accepting responsibility for the consequence.

07 / Run one commercial test

Test three comparable briefs before changing the agency.

Skip the company-wide AI transformation deck. Choose one recurring, low-consequence production route and run it for 30 days or three comparable briefs, whichever comes later.

Establish a baseline first. Then encode the repeatable route, preserve the client boundary, and assign named owners to exceptions. The scorecard should mirror the Workflow Capital Test.

The result may show that localization is reusable but concept development is not. It may expose an “automated” route that still needs an hour of cleanup. Good. A failed assumption is cheaper than an agency redesigned around it.

Pawook can keep client projects separate, connect briefs and source assets to reusable multi-model workflows, preserve runs and outputs, and expose production activity and cost. It is infrastructure an agency operates, not a Superside- or Monks-style managed service. It does not supply the talent, set the price, approve legal risk, or decide what good looks like.