What an AI marketing system actually does, and what it never should.

What counts as an AI marketing system?

The test is simple. Can you say what it watches, what it produces and who acts on the output? If yes, it is a system. If the answer is "it uses AI to transform our marketing", it is a slide.

A working system has three parts. An input it reads on a schedule: ad accounts, analytics, a CRM, a product catalogue. A job it does every time without being asked: check, compare, summarise, draft, route. And an output that lands somewhere a person will see it: a list, a report, a message, a ticket. The AI sits in the middle, doing the reading and the explaining. The judgement about what to do next sits with you.

What jobs do these systems do in a real business?

Five patterns cover most of what small and mid-sized businesses need. They are the five I build most often.

1. A campaign agent that checks pacing and flags the unusual

Every morning it reads spend, conversions and cost per result across your active campaigns. It compares them with the plan and with the last few weeks. Anything drifting gets a short note: what moved, by how much, and what it thinks caused it. Output: a prioritised action list. Nobody logs into four ad platforms before their first coffee.

2. A reporting engine that explains, not just displays

Dashboards show numbers. They do not tell you why the numbers changed. A reporting engine pulls the channels into one view and writes the paragraph a good account manager would write: what changed, why it matters, what happens next. Output: a written report your directors will read.

3. A trend monitor that arrives before the planning meeting

It tracks searches, conversations and competitor movement in your category, then surfaces the signals worth a decision. Output: demand signals, not a firehose of alerts.

4. A pre-sale chatbot that answers from your own material

Not a generic assistant. A narrow one, built on your product information, delivery terms and past customer questions, answering at the moment a buyer is deciding. Output: clear answers, and a log of what people asked that your FAQ does not cover.

5. Workflow automation that moves work between the tools you already use

Leads into the CRM with the right tags. Briefs into the project tool. Approvals routed to the right person. Fewer handoffs, and a process you can inspect when something goes wrong. Output: work in motion.

What should AI never decide on its own?

Three things, in my experience, and I hold the line on all of them.

Money. A system can calculate that a campaign is overspending and recommend moving budget. It should not move the budget. Every reallocation gets a named human approval, with the before and after recorded.

Customer-facing claims. A chatbot answers from approved source material. It does not invent a delivery date or a discount. If the answer is not in the material, it says so and hands over to a person.

Anything you cannot inspect. If a recommendation cannot show the evidence behind it, it does not ship. Deterministic code calculates the figures. AI explains them. That split is what keeps the numbers trustworthy.

How do you know if your business is ready?

You do not need a data team or a big budget. You need three things.

A process that repeats. Weekly reporting, or the same product questions arriving by email every day. If it happens on a schedule and eats attention, it is a candidate.

Data you broadly trust. Not perfect, but real. If your conversion tracking is broken, fix that first. A system built on bad data produces confident nonsense.

Someone to own the output. A report nobody reads and an action list nobody works through are wasted. Name the person before the build starts.

Where should you start?

Not with the most impressive idea. With the one process that costs the most attention each week. For most businesses that is either campaign monitoring or reporting, because both happen constantly and both are dull enough that people skip them.

Build that one. Run it alongside the manual process for a few weeks so you can compare. When you trust it, hand it the job for real and move on to the next one. One system at a time beats a platform nobody finishes.

Every system I build is delivered one of two ways: handed over to your team with documentation, or run as a managed service where I keep it maintained and improving. Which one suits depends on whether you have someone who wants to own it.