How I work with AI: I do first what I sell you
Every week I use AI for my own work: reports, checks, product data and research. Here you can read how I do that, which rules I follow and what went wrong. If you want this set up for your marketing, see AI in marketing.

My five rules for AI workflows
I treat AI like a fast, enthusiastic new colleague: a lot gets done, but not everything goes live unseen. These rules sit above every task.
A trial run first
A new task first runs as a trial, without changing anything. Only then does it go live.
Why: That way I see what the AI would do before it touches money or a customer.
Every number has a source
A report shows where each number comes from, and it matches the table.
Why: A language model writes convincingly, even when a number is wrong.
Irreversible? I approve it myself
Sending, deleting, raising a budget or going live happens only after my approval.
Why: A mistake you can undo is a lesson. A mistake you cannot undo is an incident.
Check on live data
I check against what is in the account right now, not a file from yesterday.
Why: Files go stale. The account is the truth.
When in doubt, read only
I keep permissions small. Google does the same for its own agent connection: read by default, changes only as a proposal you confirm (Google).
Why: What an agent is not allowed to do, it cannot mess up.
Examples from my practice
These are the tasks I let AI do myself. Each task shows what gets checked.
Reporting agent: a check on every number
What the text says matches the table. If not, the report is not sent.
Creative assistant: catching price errors
An ad that names a different price than the store does not go live. The check happens before a human sees it.
Feed builder: checking every day
Prices and stock are compared every day. Deviations arrive as an alert, not as a silent change.
AI visibility tracker: a fixed set of questions
The same questions every month, so you see whether you are mentioned and whether that changes.
How these tasks are divided among the crew, and what each agent may do, is on AI agents.
What did not work, and what I learned
AI projects go wrong for me sometimes too. That is why the mistakes are here, along with the rule that came out of each. It is also the reason the rules above exist.
A check that depended on my laptop
- What happened
- A daily check worked, until my laptop went to sleep. Then nothing was checked and nobody noticed.
- What I learned
- What others depend on must not depend on my laptop.
- What is different now
- It runs in the cloud and reports when it has not run.
Writing results while a source failed
- What happened
- An expired login led to more than a hundred wrong rows in a file.
- What I learned
- A run that half fails should change nothing.
- What is different now
- A failed run changes nothing and sends an alert.
Judging prices from the AI's own browser
- What happened
- That browser often shows a wrong price, because a delivery address is missing.
- What I learned
- What the AI sees is not what a customer sees.
- What is different now
- A human checks prices with a real address.
Unclear instructions to my tools
- What happened
- The tools filled in data in a different form than I meant.
- What I learned
- A model does what it is told, not what you mean.
- What is different now
- The instructions now say exactly what I want, in plain language, with an example.
My checklist before anything goes live
Every task goes through these five questions. If I answer no to any of them, it does not go live.
OpenAI and Anthropic recommend the same approach in their guides to agents: start simple, and have a human approve what cannot be undone (OpenAI, Anthropic).
- Did I run it as a trial first?
- Does every number match the source?
- Can I undo it? If not: who approves it?
- Does the agent have only the permissions it needs?
- Does the client know what is happening? It is in the work log.
What I do not leave to AI
What I let AI do
- Going through lists: search terms, products, disapprovals
- Spotting deviations
- First versions of copy and reports
- Research and summaries
What I keep myself
- Strategy and the choice of where budget goes
- Everything that goes live or costs money
- The conversation with you
- Responsibility for the outcome
One more risk I take seriously: an AI that reads outside text can run into hidden instructions. This tops the OWASP risk list (OWASP). That is why agents that read from outside get no permission to change or send anything.
Frequently asked questions
Do you use AI for your own work?
Yes, every week: for reports, checks, product data and research. What I advise clients to do, I do myself first. Every result is checked before it counts.
Does AI write my copy?
AI makes first versions, and I review and adjust them. Nothing goes live without me reading it myself. For ad copy I also check that the price matches the store.
What happens when the AI gets it wrong?
That is why I work with trial runs, a source for every number and a human who approves irreversible steps. That way a mistake is found before it reaches a customer or your budget, and what went wrong ends up in the rules.
Which data does the AI see?
Only what the task needs, and read only by default. We agree up front which data is used for what. Personal data stays out when it does not have to be in.
Can I see what has been done for me?
Yes. Your dashboard has a work log in plain language: what I did and why. See client dashboards.
Sources
- Google, Google Ads MCP server
- OpenAI, A practical guide to building agents
- Anthropic, Building effective agents
- OWASP, Top 10 for LLM applications
The mistakes and rules on this page come from my own work. Updated November 18, 2026.
Further reading
AI agents
What is an AI agent, how does it work and when does it pay off?
AI in marketing
What AI already does in marketing, in ads, feeds and reports, what the platform handles itself and where a human is still needed.
Governing AI agents
Permissions, approval and an audit log for agents in your ad account.
Start with the AI automation scan.
Then you know what pays off.