An AI marketing agency does two things that a traditional agency does not: it uses AI systems to produce and optimise marketing work at a different cost and speed, and it makes your business visible inside AI systems where buyers now research. Most firms using the label do the first. Fewer do the second. Almost none separate them in a proposal.
The delivery-model question — agency, consultant, or in-house — turns on three variables that most comparisons never mention, and they are the ones that determine what you are left with in eighteen months.
What the category actually covers
Four capability areas sit under the label. A firm may credibly offer one.
Production. Content, creative and campaign assets made faster using AI tooling. The highest-volume claim and the lowest barrier to entry.
Performance. Media buying, bidding, audience and creative testing where model-driven systems make most of the granular decisions and the human work moves up to strategy, creative and measurement.
AI-surface visibility. Being retrieved, cited and described correctly inside AI Overviews, AI Mode, ChatGPT and similar. A distinct discipline with its own measurement unit — prompts rather than keywords — covered in our AI visibility tracking guide.
Internal enablement. Building your team's capability: workflows, prompt libraries, governance, review standards. Often sold as consulting rather than agency work.
A firm strong in production and absent in visibility is a different purchase from the reverse. Ask which of the four they actually staff for.
The ownership questions
Here is what the standard agency-versus-in-house comparison misses. Three variables decide this, and the third is the one nobody raises until it is expensive.
One — who carries the tooling spend
AI marketing runs on a stack, and someone pays for it. An agency absorbing tooling into a retainer is buying at their scale, not yours, and spreading it across clients — usually cheaper than you replicating it. An in-house team pays list price for everything and the subscriptions persist whether or not anyone is using them.
Get this explicit. A retainer that looks expensive next to a salary often includes a stack that would cost a meaningful fraction of that salary to assemble alone.
Two — model and vendor lock-in
Whose stack are your workflows built on, and what happens when it changes? Providers deprecate models, change pricing and alter behaviour. Work built rigidly on one provider's capabilities inherits that risk.
Ask any partner how they handle a model deprecation, and whether your workflows are portable. A firm that has lived through one will answer immediately. A firm that has not will tell you it is not a concern, which is the answer of someone who has not been through it.
Three — who owns the prompts, workflows and context
This is the one that matters most and gets asked least.
In AI-led marketing, accumulated context is the asset. The prompt library tuned to your brand voice. The workflows that encode how your category actually converts. The retrieval context holding your positioning, objections and proof. That is months of iteration, and it is worth more than any single campaign it produced.
If it lives in a vendor's account, you are renting your own institutional knowledge. When the relationship ends you do not just change supplier — you restart accumulation from zero, and the replacement partner spends the first quarter rebuilding what you already paid for.
Put it in the contract at the start, when it costs nothing. The three questions:
- Do we own the prompt libraries and workflows built for us?
- Are they documented well enough for another team to run?
- On exit, what is handed over and in what format?
Any partner that hesitates has told you the answer.
Agency, consultant or in-house — scored against those three
| Agency | Consultant | In-house | |
|---|---|---|---|
| Tooling spend | Shared across clients, usually cheapest per unit | Usually yours | Yours, at list price |
| Lock-in risk | Moderate — their stack, but they manage change | Low — advisory, you build | Low, but you absorb migration work |
| Context ownership | Contract-dependent — negotiate at signing | Usually yours by default | Yours |
| Speed to capability | Fastest | Medium — you still execute | Slowest |
| Depth of specialism | High, across accounts | High, narrow | Limited to who you hire |
| Cost shape | Retainer | Project or day rate | Salary plus stack plus ramp |
| Fails when | Scope is vague and ownership unnegotiated | Nobody internally executes the advice | The single hire leaves |
The honest summary: an agency buys you capability immediately and requires you to negotiate ownership deliberately. A consultant buys you direction and requires you to have execution capacity. In-house buys you permanence and requires you to survive the ramp — and to keep surviving it when that person is recruited away, which in this market happens.
Most businesses running seriously on AI marketing end up with a hybrid: outside depth for the specialist disciplines, inside capability for the parts that must not leave the building.
What each model looks like in practice
The comparison table is abstract. Here is how each actually runs, including the failure mode.
The agency engagement. A team with specialists across the four capability areas, running your programme with their stack and their accumulated process knowledge. You get capability immediately and depth you could not hire for at your scale. It works when scope is specific and ownership is negotiated at signing. It fails when the scope is "help us with AI" — vague scope produces activity reports, and activity reports are how twelve months disappear.
The consultant engagement. A senior practitioner who assesses, designs and advises, while your team executes. You get direction, decision quality, and usually retain ownership by default. It works when you have execution capacity and a specific decision to make — which tools, which sequence, what to build internally. It fails when there is nobody to execute: expensive advice becomes a document nobody actions, and both sides blame the other.
The in-house build. One or more hires who own the capability permanently. You get institutional memory, context that compounds, and nothing leaves the building. It works when the workload is continuous and large enough to justify a full role. It fails on two things people underestimate: the ramp is longer than planned, and a single specialist hire is a single point of failure in a market where those people are actively recruited.
The hybrid, which is where most serious programmes land. Outside depth for specialist disciplines with genuine learning curves, inside capability for brand judgment, measurement and anything that must not leave. The decision is not which model, but which parts belong where — and that is a question about permanence, not about cost.
The cost worksheet
Blank inputs. We are not publishing invented market rates, and anyone quoting you a benchmark without your inputs is guessing.
| Line | Agency | Consultant | In-house |
|---|---|---|---|
| Fees or salary | |||
| Employer costs | n/a | n/a | |
| Tooling and subscriptions | |||
| Internal management hours × loaded rate | |||
| Ramp to productive output (months) | |||
| Cost of the capability gap during ramp | |||
| Recovery cost if it ends in month 6 | |||
| Total, 12 months |
Two lines decide most of these comparisons and are almost always left out: the cost of the capability gap during ramp, and the recovery cost if it ends early. An in-house hire that takes five months to reach productive output has a real cost attached to those five months. An agency engagement where you own nothing has a recovery cost that only appears at the end.
The handover checklist
Whichever model you choose, these five items belong in writing before work starts. They cost nothing to negotiate at the beginning and are expensive to retrofit — and in an AI-led engagement they are worth more than the deliverables.
| # | Item | What good looks like |
|---|---|---|
| 1 | Prompt libraries | Documented, exportable, yours on exit |
| 2 | Workflows and automations | Built on your accounts, legible to a third party |
| 3 | Brand and retrieval context | The positioning, objections and proof corpus is yours |
| 4 | Tooling accounts | Subscriptions in your name where practical, or a documented migration path |
| 5 | Exit handover | Defined format, defined timeline, defined in the contract |
The test is not whether a partner agrees to these. It is whether they have a process for them already. A firm that has handed over cleanly before will have templates. One that treats these as unusual requests is telling you what their last exit looked like.
Where the data disagrees
There is a genuine measurement dispute here that determines whether any of these models can be evaluated at all: how should AI marketing capability be measured — by output or by outcome?
Position A — output velocity. Measure assets produced, tests run, campaigns shipped, time-to-launch. The argument: in a system where the algorithms make the granular decisions, the controllable variable is how many quality inputs you feed them. More tested concepts means more chances to find the winner, and velocity is the honest proxy.
Position B — outcome only. Measure pipeline, cost per acquisition, contribution margin. The argument: output is an activity metric, and activity metrics are how agencies stay busy while results stay flat. A team producing forty assets a month and no incremental revenue is expensive, not productive.
CONTESTED, and both camps have real practitioners.
Our read. The dispute dissolves once you separate the timescales. Output velocity is a leading indicator and outcome is a lagging one, and holding an engagement to lagging indicators alone means you cannot steer for the first quarter. So measure both, with a rule: velocity is only a valid metric when paired with a win rate. Forty assets with a documented hit rate is a functioning testing system. Forty assets with no measured win rate is just volume, and it is the most common way marketing budget disappears without anyone being able to say what happened.
The practical version: agree a hit-rate definition before work starts. What counts as a winner, measured how, over what window. An engagement without that definition cannot be evaluated by either camp.
What would change our position: platform-published data linking creative or content volume to incremental outcome at fixed quality. Nobody has isolated it, because volume and quality are entangled in every real account.
What to do in the next ninety days
First thirty days. Decide which of the four capability areas you actually need — most businesses need one or two, not all four. Then settle the ownership terms before you settle the model. Ownership is nearly free to negotiate at signing and expensive to retrofit.
Days thirty to ninety. Run one workstream properly with whichever model you chose, with a defined hit rate. Resist expanding scope until you can read the first result.
Beyond ninety days. Decide what has to move inside the building permanently. In our experience it is usually brand judgment and measurement, and rarely production.
Nobody can guarantee a ranking, a lead volume or a revenue outcome, and a partner who offers to is telling you how they sell. What can be committed to is a hypothesis, the evidence that settles it, and the date you will know.
Where this goes next
Two shifts worth planning around.
The retainer model is going to split. As production costs fall, paying a monthly fee for output volume becomes harder to defend, and the market will separate into low-cost production and high-value judgment. Firms trying to charge judgment prices for production work will be squeezed from both sides. That is good for buyers who know which they are purchasing.
Context becomes the moat — yours or theirs. The accumulated prompt libraries, workflows and brand context that make AI marketing work well are compounding assets. Over the next two years the businesses that insisted on owning them will have a genuine advantage, and the ones that did not will discover their institutional knowledge sits in someone else's account. That is a decision you make in a contract clause today, not a strategy you can adopt later.
Take the ownership questions into your next partner conversation — including ours.
If you want the three variables scored against your own situation before you commit to a model, that is what an AI marketing consultant engagement is for — and if the answer is that you need delivery rather than direction, our agency work picks up from there. WKND has been running growth programmes since 2018 across more than 400 brands, with 20+ years of collective DTC and B2B experience behind the work.
