A robot called Jo

Under the hood of Jo, my AI Growth Manager.

Jo is what I named my AI growth manager, which was probably a mistake because now I talk about it like a colleague who occasionally needs a quiet word.

Jo runs client campaigns. This means it does all of this; list intake, audience and ICP selection, messaging, setup, launch checks, monitoring, paid support, reporting, and a bunch of boring rules that stop a campaign turning into a polite disaster.

I’ve built a few of these systems now. AI writers with messaging frameworks and voice rules. A qual researcher that can call your mobile and interview you. An AI ad engine that reads direct response craft files before writing a headline. Same pattern every time. Don’t give it a prompt. Give it a system. Documents it reads before it does any damage.

When Jo gets a new campaign, it does not start by writing. It starts the way any decent marketer would. Who is the client? What do they sell? Who buys from them? What has worked before? What geography are we in? What channels are even allowed? Cold email in Ireland and the EU is not the same game as cold email in the US. Jo checks geography before suggesting a channel, which saves you from building a beautiful email sequence you can’t legally send.

Jo creates and manages what I keep calling a database. Technically it is a very large CSV file. Not glamorous. But works.

At the beginning of any session, it reads my notes from years spent in direct marketing. Bob Stone, Drayton Bird, John Caples, direct response writers who understood attention long before everyone started saying “performance” in meetings.

Before writing, Jo has to answer two questions. What is the strongest benefit to the reader? What would make a stranger stop scrolling? Only then does it write. It then goes through a slop audit. Em dashes, zero tolerance. Corporate language. Triple adjective stacks. Paragraphs with the same shape and length. Sentences that sound like they were written by someone trapped inside HubSpot trying to win a procurement tender.

If it reads like AI wrote it, Jo rewrites it before the client see it. Ironic I know. But not all writing is the same. Jo is not being asked to sit in a corner and have a point of view on culture. It is being trained for a narrow job: persuasive campaign writing, for a specific client, to a specific audience, inside a specific channel, with claims we can stand over.

AI has a confidence problem. Not emotionally, obviously. I don’t think Jo is lying awake worrying about its Q3 forecast. The issue is overconfidence. So I’ve trained it to treat claims differently. Some things are facts. Some are benchmarks. Some are hypotheses. I often download reports for Jo to read and upskill. But it knows that a vendor report saying “Campaigns that combine LinkedIn plus email produces 287% more responses” does not get treated as a fact. It might be true. It might also be a sales deck wearing a lab coat. Jo can use it as a hypothesis, but if we want to build a recommendation on it, we test it ourselves.

The same applies when results start wobbling. Jo does not immediately rewrite the copy just because reply rate is low. It checks the stage first: list, channel, deliverability, offer, message. Jo will not call a winner after 40 results because one version is slightly ahead and everyone is bored waiting. It will not call one bad week a trend when the list changed, or the copy changed, or it happens to be August and half the country has mentally left the building. Sometimes the answer is not “the campaign is broken”. Sometimes the answer is “too early”.

Over time, all of this turns into campaign memory. Every campaign adds a file. Every experiment adds a learning.

Jo can write a cold email, obviously. So can every other AI tool. The email is just the visible bit. The real work is the legal check before the channel recommendation, the craft file before the headline, the source rule before the claim, the diagnostic table before the rewrite, the campaign file before the analysis, and the learning log before the next brief.

That is what I mean by building systems. Not a chat window with a job title. A working system with enough process, craft, evidence, and memory inside it that the output has somewhere to come from.

The system still needs me for taste, judgement, and the times when Jo randomly decides to ignore the rules I sweated over to create. But for the most part, the repeatable checks, the nagging, and the boring bits that slip when people get busy now sit in the system.

Which is handy, because apparently I need managing.