Name the job you actually want

“I want to move into AI” is not a target. It is a direction, and you cannot apply to a direction. The applications it produces read as unfocused because they are, and the reader can tell in fifteen seconds.

The standard advice is to get clear on what you want, which is useless: if you knew, you would not be reading this. Here is a way to get specific without knowing the answer in advance.

Work backwards from eight real postings

Find eight job postings you would genuinely take. Real ones, currently open, at organizations that exist. Not roles you admire from a distance and not the job you would want after two more years of preparation — eight you would accept if the offer arrived on Monday.

This is harder than it sounds and the difficulty is the point. If you cannot find eight, one of three things is true. The target is too narrow, and you need to widen the title or the sector. It is too vague, and you are searching a word rather than a job. Or it is aimed at something that is not actually hired for.

That third case is the most common in AI. Plenty of people are aiming at a role that sits between two real jobs: part data scientist, part product manager, part internal consultant. It is a fair description of work that needs doing, and in most organizations it is three people. Learning that from a failed search costs an afternoon. Learning it from nine months of unanswered applications costs nine months.

Count the requirements instead of reading them

Now stop reading the postings as aspiration and read them as data. One row per requirement, one column per posting, a mark where they intersect. Count the underlying skill rather than the wording — “stakeholder management”, “partnering with business teams” and “translating technical work for non-technical audiences” are one row, not three. That collapsing is most of the judgment in the exercise, and it is why you do this by hand.

Then sort by the count. The requirements appearing in six or more of the eight are the real ones. Those appearing once or twice are somebody’s wish list, written by a hiring manager thinking out loud on a Friday.

The output tends to surprise people. Run it for AI product roles and the top is rarely the technology. It is things like: has shipped something to real users, can read a model evaluation without being frightened of it, has worked alongside a data team, can write. Named frameworks and model versions sit far lower, because every company names the stack it happens to run.

It is dated evidence, not a law. The ranking looked different eighteen months ago and will look different next year.

That is the opening. The rest of the lesson — the remaining sections, the exercise and the check-your-understanding questions — is behind the gate below.