Why AI cannot analyse data without prior context
An AI model reads tables and metrics at a speed no human can match, but it knows nothing about my business model or the psychology of my buyers. Without that backdrop, it applies generic rules to a market it doesn't understand, treating every campaign as interchangeable with any other.
In the class I taught, I put it directly: the tool "doesn't know your ICP. It doesn't have the context and business history that you, as the person embedded in this business, are supposed to have." This is a structural limitation, not a minor detail solved with a cleverer prompt.
If I only feed it a metrics file, the assistant assumes a higher acquisition cost is always a problem to fix. It doesn't know whether that campaign was designed to attract a more demanding buyer profile, nor does it grasp the pain points the creative addresses for that audience. It lacks what only comes from following the operation day by day.
I was clear about where each side's role sits: "I don't want artificial intelligence looking at my campaigns and thinking. I want artificial intelligence looking at my campaigns and merely detecting." The decision about what counts as a good or bad result stays mine.
Giving context doesn't mean writing a giant manual or uploading dozens of loose documents. You can build a simple brief: who buys, how much you can pay per customer, what has already been tested, and what questions the assistant should ask before drawing hasty conclusions.
Offer and ideal customer: what the tool needs to know
The assistant needs to understand the offer's core promise and who it targets before judging clicks or conversions with any seriousness. Without that, it tends to favour surface-level metrics that attract curious browsers with no real buying intent, mistaking click volume for genuine interest.
In the class, I insisted the strategic piece comes down to "knowing better than anyone who your customer's customer is, who that person is, who the ICP is." That means clarifying the problem the product solves, the most common objections raised by hesitant buyers, and the arguments that actually create identification with those who already decided to buy.
Whether a campaign is running cold traffic to validate messaging, or re-engaging an already warm audience with a direct offer, is a distinction that changes the whole reading of the numbers. A higher cost per click can be a sign of good audience qualification rather than budget waste, but you only see that difference if you know which funnel stage the campaign is in.
Automation changed a campaign. Can you explain why?
How to Give AI Context Without Dumping Cluttered Information
How to use AI in business beyond chatbot prompts
Economic limits: what you can afford to lose and expect to gain
No assistant should evaluate ads without knowing what you can afford to pay per customer and what value you expect generated over time. Without those explicit limits, it risks recommending you cut campaigns that look weak short term but actually sustain part of the business.
I shared a concrete example from my own operation in the class: a seven-euro course generated over 50,000 euros in direct sales, with ad spend of roughly 22 to 23 thousand euros. Before other costs, that left something like 30,000 euros over two or three years, an amount I described in class as not enough, on its own, to "pay the bills" when looked at in isolation. However, those same customers kept buying over the following months and years, and by my own account the same funnel brought in over half a million euros on a lifetime-value basis. As I summed it up in class: "even if the ROAS on the campaigns had initially been negative, it would have been worth it, because afterward the ROAS would be much more absurd, looking at it through an LTV lens."
If you hand only the immediate numbers to an assistant without giving it this time perspective, it will suggest cutting exactly what is quietly working well over the medium term. It's worth recording your acceptable acquisition cost, your margin per sale, and what you already know about your customers' repurchase rate.
Recent changes and testing history
The assistant needs to know what changed recently in the account and what has already been tested, so it doesn't blame the wrong cause for a conversion drop. If you changed pricing, checkout, or billing in recent days, that's information to supply before requesting any analysis.
Without that record, the tool may suggest producing new creatives when the real problem is a broken page or a tracking error that has nothing to do with the ad itself. And without knowing what your audience already rejected in previous cycles, it risks repeating suggestions you already know, from experience, don't work for that market.
A minimum spend threshold before any decision
In class I described a rule I apply, in my own AdSummit operation, before letting any system touch a budget: "ads below a 0.8 ROI get paused, but only once spend has passed 60 euros. So the rule is, no ad set gets paused before it has invested 60 euros."
This threshold exists because accumulated spend below that level provides insufficient data for a reliable read. Sixty euros is a practical baseline designed to prevent rushed decisions on thin data, rather than a threshold carrying formal statistical significance.
In the account I describe, this automation runs as a rules-based system rather than a real-time conversational assistant. I used AI assistance during the design phase, but the system now executes my predefined conditions directly without calling a language model on each run.
This distinction matters for anyone thinking about using an AI assistant differently: when you ask a chatbot to weigh in on an account, you can apply the same logic of minimum spend and time thresholds before accepting any suggestion it makes to cut or scale.
Questions the assistant should ask before offering an opinion
To avoid rushed diagnoses, it can help to ask the assistant to confirm certain things before recommending any change: whether conversion tracking is working, whether landing pages load without errors, whether leads generated are being followed up promptly by the sales team.
An acceptable cost per acquisition in the ads dashboard can still look fine while hiding a serious problem if the funnel downstream fails to convert those leads into customers. I was explicit about this in class: "often the bottleneck is in the ads, true, but often the bottleneck isn't there, it's in the offer itself, other times it's the follow-up, the missing email after a lead goes cold." An assistant without this instruction will tend to always blame the campaign, even when the real problem sits elsewhere in the sales process.
Source
This article draws from what I taught in Sextas Ímpares #151, "O papel do gestor de tráfego na era da IA," recorded while on holiday in the Algarve. Worth watching from around 13:59, where I talk about knowing your customer's customer, around 22:39 on the lifetime-value logic of the seven-euro course, and from 35:41 where I show the rules I apply to AdSummit campaigns. Full video: https://www.youtube.com/watch?v=SYe0ebyVZ5c