You don’t have an AI problem

You do not have an AI problem. Marketing automation in Australia fails for a simpler reason: the ChatGPT account exists, the system does not.

I am Adam Stewart. I run Digital Bond, and I work with founder-led teams that have already “tried AI” and still chase leads by memory.

I ran a room with about twenty business owners. Everyone had tried ChatGPT. Almost nobody had a system running.

They could draft an email. They could summarise a meeting. A few had a favourite prompt saved in Notes. Then the week started, the leads came in, the onboarding pack went out late, and the CRM stayed as honest as the last person who remembered to update it.

It was not a knowledge problem. It was a systems problem.

What did the workshop actually show?

I had expected more scepticism.

That is not what showed up. These people were not anti-AI. They had accounts. They had used the tool the way most of us used it for the first two years. Ask a question. Get an answer. Rewrite something. Brainstorm a few ideas.

Useful? Absolutely.

Useful is not the same as installed.

When I asked what happens after a new lead arrives, the answers got quieter. Someone checks the inbox when they can. Someone copies a name into a spreadsheet. Someone means to send the follow-up sequence they wrote in January.

Same pattern on content. Plenty of drafts. No calendar that a second person could run. Same pattern on onboarding. A good first week if the founder is in the building. A messy first week if they are not.

That is the gap I keep seeing in Australian businesses that are already “doing AI”. The tool arrived. The operating rhythm did not.

Why is ChatGPT not marketing automation?

A chat window is a place you go.

A system is a path that still works when you do not go there.

If your follow-up lives in a prompt you have to remember to open, you do not have follow-up. You have a good intention with a login.

If your CRM is a place you update after the fact, you do not have a CRM. You have a diary with extra fields.

If your content plan is a chat history, you do not have a content system. You have a pile of almosts.

Marketing automation in Australia is often sold as software. HubSpot, Mailchimp, a Zap, a new AI feature in the thing you already pay for. The software can help. It cannot invent the path.

Here is the definition I use with founders.

A working system has four parts: a trigger, a default action, a human check, and a definition of done. AI can draft the action. A person still approves anything that gets sent, published or spent.

Draft the first reply. Tag the lead. Update the record. Assemble the onboarding pack from the facts you already trust.

Do not let the model send the email, publish the post, or spend the ad budget because it felt confident.

Where the system usually breaks

I look at four places first, because that is where founder-led teams leak.

Follow-up. A lead comes in at 7pm. Nobody replies until the next afternoon. By then they have asked someone else. This is the most expensive “we use AI” story I hear, because the team used AI to write a clever website and then left the enquiry sitting in a shared inbox.

CRM. Contacts exist. Stages are fiction. Notes live in heads. Nobody can tell you which campaign paid for last month’s jobs without opening five tabs and guessing.

Content. There is a voice. There are ideas. There is no owned process for brief, draft, review, publish, and reuse. So the founder becomes the bottleneck again, or the output becomes generic.

Onboarding. The sale happens. Then a human has to remember fifteen steps. The customer feels the wobble immediately. Marketing gets blamed for “bad-fit clients” when the real issue is the first fortnight after they pay.

None of those is an AI problem.

Each one is a missing loop.

What does a working system look like?

Not a stack diagram.

A sentence you can say out loud.

“When a qualified lead arrives, they get a same-hour reply, a booking link or a next question, and a record that stays true without someone retyping it.”

“When we publish, the asset is reused in two other places, and we know which offer it pointed at.”

“When a new client pays, they get the right pack, the right intro, and we can see who is stuck.”

If you cannot say the sentence, do not buy another tool.

People skip this part. They ask ChatGPT to “automate follow-up” and get a plausible essay. The essay is not a system.

If your pipeline stages mean different things to different people, the model will pick one and sound sure about it.

Building the first version of a tracker or calendar can wait until the loop is named. I wrote about turning that conversation into a small tool in build a marketing tool from a ChatGPT chat. Do not start there if you cannot yet say what “finished” looks like.

Start with one low-risk recurring task

Do not start with the brand voice project.

Do not start with a fully autonomous agent.

Start with one low-risk recurring task. Something that happens every week. Something where a mistake is annoying, not fatal.

A first-reply draft for website leads.

A Friday summary of which leads went quiet.

A checklist that turns a won deal into an onboarding ticket.

A content brief that a second person can run without asking you what the offer is.

Then decide the human checks before you turn anything on.

What can the system draft?

What must a person approve?

What should it never touch?

That last list is the difference between a useful system and a mess you will quietly switch off.

Try this

I am going to describe one recurring marketing task in my business. It currently lives in [inbox / spreadsheet / someone's head].

Tell me:
1. The trigger that should start it.
2. The information it needs, and where that information already exists.
3. The steps a junior hire would follow.
4. Which steps are safe to draft with AI.
5. Which steps must stay human (send, publish, spend, anything customer-facing that cannot be undone).
6. What "finished" looks like, in one sentence I can measure.

Do not suggest new software unless the current tools cannot hold the information. Do not invent a 12-step platform. Give me the smallest system that would still work if I was offline for a day.

Paste a real example. A messy one. The polite version will get you a polite system that nobody uses.

How to measure it without fooling yourself

A lot of teams measure AI the way they measure software seats.

Who has a login?

How often did they open it?

How many prompts did they send?

That number can go up while the business stays exactly as stuck.

Measure finished work.

Ready to use: the draft went out with no rewrite.

Needs correction: someone had to edit it.

Needs escalation: a person had to do the job from scratch.

Then look at time-to-first-reply, records updated the same day, and onboarding steps completed without a chase. Those are system numbers. Logins are vanity.

I will take a boring metric that moved over a pretty dashboard that did not.

If someone is trying to sell you a new channel before this loop exists, that is the 2010 menu I argued against in agencies are still selling the White Pages. Fix the handoff first.

If the first leak is lead capture, response or CRM ownership, start with the marketing automation consultant view. If HubSpot is the system that should hold the truth, the HubSpot consultant page shows the narrower implementation path.

My honest read

My honest read: if everyone on your team has used ChatGPT and nothing in the business runs without a reminder, you do not need another prompt pack. You need one named loop with a human check. Do not expand AI until that loop finishes work without you standing next to it.

Keep the gate narrow on purpose. Widen access only after the first loop is boringly reliable.

What to do this week

Pick the leak you already know you have. Follow-up is the usual one.

Write the sentence for what should happen. Run the prompt above. Install the smallest version that would survive you being offline for a day.

That is the week’s work. Do not buy another seat until that loop finishes without you.

Before you build, the Digital Bond scorecard can show which part of the marketing system is constraining the next step.

You already know enough AI.

The question is whether the work still happens when you close the chat.