Why small accounts train judgement
Running a small budget forces me to decide without all the data I would like, and that is where judgement gets trained. I was direct: I have managed accounts worth thousands of euros a month, but when spend is limited there is not enough statistical confidence to let the machine decide alone. There I decide myself.
When there is heavy spend, a simple rule matrix can guide the machine with little intervention from me. I call this the real training ground: what teaches fastest is the number of different challenges you expose yourself to, more than the amount of money passing through your hands.
I gave my own example: I ran a large volume of leads in the legal services vertical in Brazil for three years and became a specialist in that niche. Even so, it taught me at a different pace than running an agency portfolio with dozens of clients facing different problems at once.
Defining success before touching any tool
Judgement comes from deciding, before you start, what counts as a good result and what counts as a bad one for that project. Without that prior definition, I end up relying on generic averages the tool itself suggests, which rarely reflect the real economics of the business.
I explained the matrix I apply to AdSummit campaigns: an ad set can only be paused after spending more than sixty euros with a return below 0.8. These rules came from experience accumulated across seven editions of the event.
One point I stressed is that looking only at the ad's immediate return can badly mislead the reading. I described a 7-euro course: I spent more than 22,000 euros on ads, and the course generated more than 50,000 euros in direct sales, giving a positive result before other costs over two or three years. On its own that already looks reasonable. But the same customers later brought in more than half a million euros in other product sales. Looking only at the campaign's immediate return would have hidden all that value.
Building with explicit rules and room to test
I do not want artificial intelligence to think for the campaigns, I want it to flag deviations based on a matrix I wrote myself, with my own limits and goals. In the system I showed for AdSummit, an automation runs every hour, checks the return of each ad set, and adjusts budgets according to fixed rules.
I showed, live, that dashboard: an automatic report reaches my WhatsApp every hour, with the day's return and the changes the automation made. If something falls outside what I defined in advance, I get the alert and decide myself what to do next.
If you want to apply this logic to your own project, the starting point can be similar: before granting access to real budgets, write the stop and scale rules, and test them first against a shorter period or a smaller budget. It is worth letting things breathe with modest spend at first, as I described when explaining how I spread budget across dozens of creatives until identifying the survivors.
Explore this decision further: What to Prepare Before Your First Work Session with Claude Code.
Review and the accountability AI does not take on
I insisted several times on a simple idea: artificial intelligence can suggest, optimize, and even run much of the mechanical work, but the person answering for results and managing client expectations is me, or whoever is running the project. That responsibility includes checking the actual outcome.
I was blunt about excuses: if a campaign goes wrong, the responsibility falls on whoever configured the matrix or failed to understand the business context, whether that person is working manually or leaning on automation.
If you are testing a new project, asking someone experienced to look at your rule matrix before running it live is a simple way to catch reasoning gaps before they cost money.
Recording what you learn in each cycle
I called this sensitivity: the learning that remains after pausing a campaign or ending a partnership, which no AI tool hands me in an automatic report. It is this sensitivity, built across many small challenges, that later transfers to the next client in the same segment.
This learning does not come from asking the tool for a summary of what happened. It comes from looking carefully at what worked and what did not, and understanding why, which takes time set aside just for review, separate from time spent acting.
A simple way to apply this in practice, without inventing an elaborate process, can be to write, at the end of each test or each client relationship, three or four lines on what worked and which limit of cost, time, or spend turned out to be wrong. This note follows from the distinction I made between running campaign operations and actually learning from them.
To turn a test into reusable learning, I would prepare a record of the initial question, chosen rule, available data and expected decision. I would save that prediction before observing the outcome, so the explanation cannot be reconstructed afterwards to fit whatever happened. This makes it easier to distinguish a consistent criterion from a convenient retrospective justification.
In a hypothetical exercise, the rule recommends keeping a campaign running, but the person responsible stops it because the offer is no longer available. The record should capture that information and the human intervention. On the next test, I would check whether the system recognises unavailability or refers the decision. Learning then becomes a specific process change that can be evaluated again.
Repeating in another context to test whether judgement holds
The proof that I developed judgement, rather than just luck in one niche, is carrying what I learned into a different market with its own rules. I contrasted my three years in the Brazilian legal vertical, where I became a specialist managing large volume, with my agency experience, handling clients of very different sizes and difficulties.
Exposure to varied challenges, even small ones, teaches you to handle the unexpected faster than one comfortable, already-mastered big challenge.
If you are deciding where to invest your practice time, this points to a concrete path: after solving a problem in one context, look for a second context with different business rules before assuming your method is ready for any situation.
To prepare your next step, explore AI and business mentoring.
Explore this decision further: How to plan an app with Claude Code before you start coding.