There is no AI marketing department to hire. There is a stack you assemble.

Every quarter somebody walks into a dealership marketing meeting with a slide that says AI marketing department and a per-seat price. The slide is not lying about the price. It is lying about the noun. A department is not a license. It is five jobs done by people who talk to each other, and no single vendor on the market today does more than two of those jobs well enough to hand over unsupervised.

Five jobs, not one product

Strip the category back and the work underneath it is always the same:

1. Research and monitoring. What competitors are saying, what the manufacturer changed, what the category is being asked in an assistant.

2. Drafting and adaptation. Turning one approved idea into thirty channel-correct versions without losing the claim.

3. Creative production. Imagery, video, and the licensing trail that lets legal sign it.

4. Execution inside the buying surface. Bidding, audiences, sequencing, the actual send.

5. Review and compliance. The gate between a draft and a liability.

The rest of this page is the honest version of that table, vendor by vendor, ending with the one arrangement we would actually run.

Claude is the reader, not the campaign

Claude, from Anthropic, is the model family we would put in charge of the fifth job and a large part of the second. Not because of a benchmark, but because of the shape of the work. Marketing's hardest prompt is not write me a headline. It is here is a forty-page manufacturer advertising standard, here is the price substantiation file, here is the brand book, here is the draft, tell me every line that violates something.

That is a long-document reasoning task with an instruction-following requirement, and in our own use Claude holds a long rule document better than the alternatives we have tested. It quotes the clause it is worried about. It flags the claim it cannot substantiate. It does not quietly smooth over the paragraph it did not understand, which is the failure mode that costs dealers money.

Two capabilities make it practical rather than impressive:

• Long context. You can put the entire OEM guideline in front of it rather than a summary of the guideline. Summaries are where compliance dies, because the exception that matters is always in the paragraph somebody cut.

• MCP, the Model Context Protocol. Anthropic published an open standard for connecting a model to tools and data, and it is now supported well beyond Claude. Practically: the model reads your actual fitment feed, your actual service menu, your actual price file, instead of inventing plausible values. A model with no access to the parts catalog will hallucinate a part number with total confidence.

Where Claude is genuinely weak for a marketing department is the same place every model is weak. It has no audience graph, no bid strategy, no send engine, and no idea which of your twenty headlines won. It cannot place a dollar of media. If a vendor tells you their Claude wrapper is a full marketing department, they are selling you a chat interface and a monthly invoice.

The correct home for it is upstream and downstream of the platforms: reading the rules on the way in, checking the claim on the way out, and writing the structured copy that feeds a campaign you will then measure properly. Measurement is a separate discipline with its own denominators, which is what our report page exists to demonstrate.

Google, which owns two jobs and half of a third

Google is the only vendor with a credible claim on both research and execution, and it is the one most dealers already pay for without using.

• Gemini is the general assistant, competent at drafting and analysis and increasingly woven into Workspace, so the department's documents are the context rather than a paste buffer.

• NotebookLM is the underrated product in the stack. Point it at the manufacturer's policy documents, the service menu, the objection-handling notes, and it answers from those sources with citations to the exact passage. For a service department that has to quote policy correctly on a phone call, this is worth more than another copywriting tool.

• Veo and Imagen cover video and image production with Google's own licensing posture.

• AI Max and the rest of the Ads AI stack are where execution actually happens, and where autonomy has to be capped deliberately rather than by hope.

The trade is obvious and worth stating: Google's execution layer optimizes for Google's auction. That is fine if you understand it as a commercial relationship and not as a neutral advisor.

OpenAI, the generalist default

ChatGPT Enterprise is the tool most marketing departments will end up with by default, and the default is defensible. Custom GPTs let you package a brief and a tone guide for the team. Deep research produces a usable category or competitor scan from public sources. The agentic browsing features handle the repetitive portal chores that eat an afternoon: pulling spec sheets, checking inventory pages, reconciling a co-op claim against a portal export.

It is a strong second reader and a fast first drafter. What it does not give you is governance. A general assistant will happily write a claim your manufacturer forbids, because nobody handed it the manufacturer's document, and it has no opinion about whether the underlying number exists.

Adobe and the licensing question

Adobe GenStudio for Performance Marketing and Firefly answer the question that legal actually asks: where did the pixels come from. Firefly was trained on licensed and Adobe Stock material, which is why it clears review in organizations that would stall on a model with murkier provenance. The output still needs an art director, but the trail exists, and a trail is what turns an image into an asset.

For a dealer group, this matters most in exactly the place people relax: stock photography that is close enough, a lifestyle image generated for a model-year campaign, a re-cut of a manufacturer asset. Provenance is cheaper to buy than to defend.

Salesforce and HubSpot, the execution end

Agentforce inside Salesforce and Breeze inside HubSpot put the model where the lead record already lives, which is the only place an agent can do anything consequential. That is genuinely different from a chat window: the agent can see that the equity is real, the service history is current, and the customer has already declined the alignment twice.

The constraint is always upstream. An agent working off four duplicate customer records will send four emails to one person with three different offers. The data problem is the AI problem, and it is the same problem described on this desk months ago.

Jasper and Writer, the marketing-specific pair

Both sell what a raw model does not: a brand voice that persists across a team, templates that enforce a structure, and approvals that sit between draft and publish. Writer leans toward regulated industries, where the value is the compliance posture rather than the prose. Jasper leans toward speed of production for teams who need volume with a guardrail.

The honest caveat is that you are paying for the wrapper around a model you could access directly. That is a fair trade if your team is large enough that consistency is the bottleneck, and a waste if your team is two people who never write the same page twice.

Microsoft, the ambient option

Copilot is already in the applications where the department types, which makes adoption nearly free and ceiling correspondingly low. It is a competent document and spreadsheet accelerator. It is not a marketing specialist, and treating it as one produces decks about marketing rather than marketing.

What fails, repeatedly

None of these tools fail by being stupid. They fail in predictable ways that a review gate catches and enthusiasm does not.

• Invented substantiation. Ask a model to make a claim punchier and it will add a percentage. There is no such percentage. It exists because you asked for emphasis.

• Fitment hallucination. A part number that looks entirely correct and does not fit. Only a source of truth prevents this, never a better prompt.

• Policy drift. The model writes a claim the manufacturer's advertising standard forbids. It did not read the standard because nobody gave it the standard.

• Volume without a gate. Three hundred variants of an unapproved claim is three hundred liabilities, produced in an afternoon.

• Attribution theatre. AI-generated content credited with clicks it cannot honestly claim, measured against a denominator nobody wrote down.

The gate stays shut

No creative ships unsigned. That rule does not weaken when a machine drafts the creative, it tightens, because the draft volume goes up by an order of magnitude and the reviewer's attention does not. The gate is four questions, asked in order, every time:

1. Is the source of truth attached? Feed, menu, price file, or manufacturer document.

2. Is the claim substantiated on file, in writing, by somebody with authority?

3. Does the manufacturer's own standard permit this language, and has someone who has read that standard confirmed it?

4. Who is the named approver, and did they actually look?

An AI department that cannot answer those four questions is a content generator with a department's electricity bill.

What we would actually assemble

Three sizes, three stacks, no fantasy about replacing headcount.

One rooftop. Claude or ChatGPT for two seats, NotebookLM holding the manufacturer documents, the ad platform's own AI for execution, the email platform's built-in tooling. The binding constraint is the operator's time, not the tooling budget. Someone has to own the prompts, the review gate, and the numbers.

Group, two to ten rooftops. Add a shared knowledge base that every rooftop's copy draws from, a single review workflow with a named approver per brand, and one person accountable for the data plumbing between the DMS, the CRM and the feeds. At this size the second largest risk after compliance is four rooftops producing four different versions of the same offer.

Distributor or aftermarket brand. Add licensed-provenance creative production, agent execution inside the CRM with hard caps on autonomy, and an evaluation set: twenty fixed briefs scored monthly, so you learn whether the model changed, the prompt changed, or the person did.

Claude sits at the center of all three, but as the reader and the reviewer rather than the department. It reads the rulebook, it checks the draft against it, it writes the structured copy that feeds the campaign. Then the campaign runs where campaigns run, and it gets measured on booked work instead of on the confidence of the prose.

The short version

There is no AI marketing department to hire. There is a stack you assemble, and the assembly is the skill. Put a long-context model that respects instructions in the review seat, put the execution inside the platforms that own the auction and the customer record, and keep a human with authority at the gate. Everything else on the invoice is a seat license with a better slide.