Give the AI the rulebook before you give it to the marketers

A few weeks ago I wrote about what AI does to email volume. Writing used to be slow and the slowness was quietly keeping a lid on how much you sent. That lid is gone. This post is about the other half of the problem which is what the words say once anyone can produce them.

Try the AI assistant in Eloqua or Marketo and the draft you get back is fine. On topic, grammatical, a little eager. If you squint it reads like your brand. If you don’t squint it reads like everyone’s brand, because the model has no idea who you are beyond what the person typing the prompt happened to mention.

The rules are in a drop-down

Both platforms give the marketer some dials. In Eloqua you pick a purpose, a tone and a formality for each piece of text you generate. In Marketo you pick a buying group, a journey stage, a communication strategy and a tone and you can point it at reference content. Our guide to AI copy in Eloqua and Marketo goes through all of it.

Here is what those dials have in common. The marketer sets them. Every marketer, every time, from scratch. Nothing stops the field marketer in Munich choosing Urgent for a newsletter and the product marketer in Austin choosing Humorous for a security notice. Nothing tells either of them that the hero zone in your template holds about 180 characters before the layout gives up. Nothing knows that legal struck the word guarantee from your vocabulary three years ago after a conversation nobody wants to repeat.

You spent years getting the brand voice consistent. One team wrote everything, a style guide sat on the wiki and the two people who actually read it reviewed everybody else’s drafts. Now there are a hundred people with a button and the style guide is still on the wiki.

A brand voice that depends on which drop-down someone picked isn’t a brand voice. It’s a mood.

“Just put the brand guide in the prompt”

That is the first thing everyone suggests and it works for about a week. The conscientious marketers paste the voice section into every prompt. Then they get busy and paste a shorter version. Then they stop, because the whole point of the button was to save time and the pasting takes longer than the writing used to.

Even when people do it, the prompts drift. Each person summarizes the brand guide a little differently and the model obliges each of them a little differently. You end up with twelve interpretations of “confident but approachable” going out under one logo.

“We’ll catch it in review”

The second suggestion is to lean on approvals. Somebody reads every AI draft before it goes out so what does it matter what the model wrote?

It matters because AI made drafting fast and did nothing for reviewing. The same two people who used to review a handful of carefully written emails a week now get a flood of competent-looking drafts, each one needing a read for tone, for claims, for length, for the banned words. Review becomes the new bottleneck and the approval queue fills up with work that a rule could have done before the draft existed. Reviewers also get tired. By the fortieth draft “industry-leading” starts to look acceptable.

What a rulebook looks like

The fix is boring and it works: decide the rules once, centrally and have them applied to every prompt without anyone having to remember. Four kinds of rule cover almost everything.

Tone. Two or three sentences describing how your brand sounds and how it doesn’t. “Plain, confident, never pushy. We don’t use exclamation marks and we don’t call anything revolutionary.” Every draft starts there, whoever asked for it.

Length. A character limit per zone, matched to the template. The subject line gets one number, the hero gets another, the body gets a third. Copy that fits the first time doesn’t need trimming by hand and doesn’t break the layout when someone forgets to trim it.

Words that never appear. Legal’s list, brand’s list and the competitor names. Guarantee. Best-in-class. Any claim about results. Anything that sounds like a promise about pricing or availability. The model never produces them so nobody has to catch them.

Standing instructions. The things you would put at the top of every prompt if you could. Write in US English. Don’t mention pricing. Don’t invent dates or names. Don’t promise outcomes. Write them once and they ride along with every request.

Notice who writes these: marketing ops, with brand and legal in the room, the same people who already decide what a template allows and who approves what. AI copy is one more thing the template governs.

Two more things the rulebook should settle

Where the draft lands. A draft that goes into a free-form editor can still wreck the layout and the fonts. A draft that goes into a locked template zone can only be words. Pair the rules with templates that leave the right things open and the model has very little room to do damage.

Whose model it is. The platform assistants run on the vendor’s model under the vendor’s terms which your security team may or may not have read. If you would rather the requests went to the AI provider you already have a contract with, on your own key, under the data retention terms you already negotiated, that is a decision to make before the button is on rather than after.

How we do it in Campaign Launchpad

We built AI content generation into Campaign Launchpad the same way we built everything else in it: ops sets the rules once and the software applies them every time. Ops writes the tone, the character limits, the blocked words and the system prompt. A marketer in a campaign wizard types something like “write a reminder for people who registered for the partner summit but haven’t picked their sessions” and Launchpad adds the rules before the request goes out. The draft comes back in the brand’s voice, under the limit, without the words on the list and lands in a template zone it can’t break.

Requests go to Anthropic or OpenAI on your organization’s own API key so the agreement that governs the data is yours. And because the same rules apply whichever platform the campaign is built in, a team running Eloqua and Marketo gets one voice out of the AI across both.

Let them use the button

I’m in favor of the button. Faster first drafts are a real gain and marketers who spend less time staring at a blank text box spend more of it on who they’re writing to and why. The thing to insist on is that the rules exist before the first prompt and that they apply to everyone, including the people who were never going to read the style guide.

Write the rulebook. Put it where the model can’t miss it. Then let people write.