Where is AI reliable, and where isn’t it? After months of building a content pipeline, my answer fits in one rule: AI judges, code counts. It follows directly from the limits of generative AI, and it’s what makes a pipeline reliable and a report defensible in front of leadership. It’s also one of the pillars of verifiable AI content.
AI excels whenever you need to understand meaning. Is a comment a question, a complaint or a comparison? What’s the intent behind a search: learning, comparing, buying? Is the tone worried, annoyed, ironic? Does an objection rest on a generalization (“they’re all the same”) or on a specific fear?
These questions resist fixed rules. I measured it in a test: the task was to recognize subtle nuances in customer objections. Keyword rules caught one in six. AI caught all six, including sentences without a single telltale word, and those written in French. For judging, AI wins.
Code excels whenever you need to match, deduplicate, calculate or verify. Linking an item to a city, counting how many people ask the same question, adding up a total, checking that a reference exists: these are operations where you want the same result every time.
An AI doesn’t guarantee that. Ask it the same calculation twice and it may give two answers. And when it’s wrong, it’s wrong with the same confidence as when it’s right.
Here too, I measured instead of assuming. I tested AI on the geographic matching of the pipeline’s items, that is, knowing which region a comment relates to. It brought no gain over fixed rules, while costing more and varying from one run to the next. So matching stayed with code. AI keeps a role where judgment is needed, for example to say whether a local detail is truly distinctive.
How do you know whether a task belongs to AI or to code? I ask myself one question: could two competent experts, in good faith, give two different answers? If so, it’s judgment: the tone of a message, the intent of a search, the nature of an objection. That’s what AI is for.
If not, if there’s a single right answer, it’s calculation or verification: a total, a grouping, an exact match, the presence of a fact in a source. Code should handle it, because it will give the right answer every time, and because you can check how it got there.
The rule comes into its own in reports. A monthly SEO report contains dozens of figures: clicks, impressions, positions, contact requests. It also contains sentences that explain them. The temptation is strong to let AI write the whole thing.
In my pipeline, the split is strict. Code calculates every figure. AI receives those figures and writes the commentary, like an analyst presenting results they aren’t allowed to change. Then an automated check rereads the text: any figure that doesn’t exist in the calculated data is rejected, and the commentary is redone.
This check starts from a simple observation: an AI writing commentary tends to “round,” to extrapolate, or to invent a trend that sounds good. A 3% increase becomes “strong growth.” With the check, every figure in the text can be found in the data, and every trend statement stays faithful to the calculations. And a human always approves before sending.
The split between AI and code comes with another rule: throw nothing away. When the pipeline encounters the same information twice, it doesn’t delete the second one; it counts that it comes back. Code keeps that count exactly, and the count becomes a measure: a question that comes back fifty times weighs more than an isolated one.
Likewise, a report that shows only part of a list says so clearly (“the top 10 of 250”). It never suggests it’s showing everything.
This rule ties into a broader conviction. AI is a brilliant executor and a lazy strategist. Give it a clear frame, and it produces and scales remarkably well. Let it define the frame, and it picks the most ordinary answer. The “AI judges, code counts” split is precisely a frame: it tells each side what it does best.
For leadership, the benefit is immediate: a report where every figure holds up. If someone asks where a 12% drop comes from, you can show the calculation, not an AI hunch. If two successive reports are compared, the figures are produced the same way, so they’re comparable.
For your content, the benefit is the same: AI brings a fine understanding of your customers, and code guarantees rigor. It’s also what makes it possible to measure and reduce hallucinations rather than hoping they’re rare.
This division of labor is applied at every step of the method, from listening to the market to monthly measurement. To see the whole picture and what it can bring your business, discover the AI SEO pipeline.