An LLM reads the CV. The moat is what it doesn’t see.
Anyone can wrap an LLM around a job board this weekend. The CV read is commoditised. What can’t be copied with a better prompt is the layer learned from your outcomes — and that’s the whole game.
Let’s be honest about what’s easy. Reading a CV against a job description and producing a fluent paragraph about fit is now a commodity. A capable model and an afternoon gets you most of the way. If that’s the product, there is no product — your competitor ships the same thing on Sunday with a slightly better prompt.
So the question that matters isn’t “can an AI read the CV?” It’s “what does the CV not contain?” And the answer is everything that actually decides the placement: which of your clients accept which profiles, where your deals close on rate, which stack a particular lead has quietly distrusted, who’s really available against a start date. None of that is on the page. All of it lives in your outcomes.
The defensible layer is yours, not ours
A prompt is copyable. A weight learned from a year of your own accepted-and-rejected history is not, because nobody else has the history. That’s the moat, and the uncomfortable, honest part is that it isn’t something we hand you on day one — it’s something the tool accumulates with you, from your desk, and treats as yours.
How the read is actually built
Under the surface, a submission is scored on six dimensions, not one fuzzy “match.” Each is a distinct kind of signal:
- Revealed preference — what this client has historically accepted and rejected.
- Rate — does the number close the margin and clear the client’s real ceiling?
- Two-sided fit — not just “will the client take them,” but “will they take the client.”
- Timing — availability against the actual start date.
- Technical strength — the part the CV does speak to, including equivalent stacks (Spring is Java experience, even when the word isn’t repeated).
- Confidence — how much of the above is read from the document versus inferred, stated plainly.
Two design rules hold it together. First, gates before weights: some dimensions are pass/fail — if the candidate isn’t available in time, no amount of technical brilliance rescues the submission, and the model shouldn’t average that away. Only once the gates clear do the weights decide ranking. Second, those weights are learned from your outcomes, not hard-coded by us. Your desk teaches the engine what to lean on.
The LLM reading the CV is table stakes. The product is the layer it can’t see — and that layer is built from the one thing a competitor’s better prompt can never reach: what happened on your desk.