The strongest AI candidates are almost never actively job hunting, and when they do read an ad, they give it the same scan they give a pull request: thirty seconds, looking for concrete signal, primed to close the tab at the first cliché. "Fast-paced environment," "AI rockstar," fifteen required frameworks, no salary, no named problem: tab closed. A job description for an AI role is a technical document with a marketing function, and it wins by being specific where every competing ad is vague. This article covers what senior AI people actually scan for, the problem-first structure that holds their attention, and the red flags that repel them before you ever know they were there.
What senior AI people actually scan for
Eye-tracking an experienced candidate through your ad would show them skipping the intro paragraph entirely and hunting for a handful of concrete facts. If those facts are missing, the gaps get filled with the pessimistic default.
| What they look for | Question in their head | If it's missing, they assume |
|---|---|---|
| The problem | "What would I actually work on?" | There is no defined problem; I'd be the AI figleaf |
| Data & stack reality | "Is there real data and infrastructure, or slides?" | Data is a mess and I'd spend a year on plumbing alone |
| The team | "Who would I learn from and report to?" | I'd be the only technical person, unsupported |
| Salary range | "Is this even in my bracket?" | Below market, and they know it |
| Deployment evidence | "Has anything shipped here?" | Eternal proof-of-concept land |
The problem-first structure
This structure front-loads the scan targets and pushes company boilerplate to the end, where it belongs. It fits on one page.
- 1The problem (3-4 sentences, first thing on the page): the concrete technical challenge, its scale, and why it is hard. "Our demand forecasts run on a 5-year-old gradient-boosting pipeline that breaks on promotions; you will own its successor end to end."
- 2What you will do (5-6 bullets, verbs first): real activities in the first year, including the unglamorous ones, honesty here pre-sells the reality.
- 3What you bring (max 6 must-haves, outcome-phrased): taken directly from the requirements profile, plus a short, clearly separated nice-to-have list.
- 4Stack and data reality (honest, 3-4 lines): what is modern, what is legacy, what is greenfield, and what the data actually looks like.
- 5Team and reporting line: who they work with day to day, who they report to, how many engineers/scientists are around them.
- 6Salary range, location policy, process: the range, the remote/on-site reality, and the interview process with the number of rounds and total time commitment.
- 7About the company (short, last): two or three sentences, after everything the candidate actually came for.
Honest disclosure: stack, data, salary
Every ad competes with dozens claiming a "modern cloud-native stack" and "data-driven culture." Verifiable honesty is the differentiator, and it filters in the people who want your actual job, not the imaginary one. Salary belongs in this section too: ranges measurably increase qualified application rates, EU pay-transparency rules are making disclosure mandatory across member states, and senior candidates increasingly read a missing range as a lowball in waiting.
| Aspirational (reads as spin) | Honest (reads as signal) |
|---|---|
| "Modern, cloud-native ML stack" | "Models run on a managed cloud platform; feature pipelines are solid, experiment tracking is basic, you would help choose what we adopt next" |
| "Huge amounts of data" | "About 40M transactions across three systems; joined and cleaned for two core use cases, raw elsewhere" (illustrative figures) |
| "Competitive salary" | "EUR 85,000-105,000 depending on level, plus the criteria that decide where in the range you land" |
| "Work with cutting-edge AI" | "Two models in production, one LLM feature in pilot; your first project is taking the pilot to production" |
Red flags that repel senior candidates
These are the tab-closers. Each one is common, each one is avoidable, and each one costs you candidates you never see.
- 1The unicorn list: ten-plus required skills spanning research, engineering and DevOps signals the company does not know what it needs, seniors read it as guaranteed role confusion.
- 2"Rockstar," "ninja," "guru," "AI wizard": personality-cult vocabulary reads as a culture that rewards heroics over engineering.
- 3No salary range in 2026: increasingly read as "below market and hoping you won't ask," and soon simply non-compliant in much of the EU.
- 4Vague process ("several interviews"): seniors budget their time; an unspecified process signals a disorganized one.
- 5Mission-only ads: three paragraphs of vision with no named problem, stack or team tells a technical reader there is nothing underneath.
- 6Buzzword stacking ("GenAI, AGI-adjacent, blockchain-ready"): technology name-dropping without coherence signals leadership chasing trends, not solving problems.
A miniature example: the opening that survives the scan
An illustrative opening block for a fictional logistics company, the first thing a candidate reads. Note that it answers three scan targets in the first five lines.
| Element | Text (condensed) |
|---|---|
| Problem lead | "Our route-planning runs on hand-tuned heuristics that leave an estimated 8-12% efficiency on the table (internal estimate). You will own the ML system that replaces them, from first model to production rollout across 40 depots." |
| Stack honesty | "Python services on AWS; clean telemetry data for 18 months, older data patchy. No ML in production yet, you would be first, with a platform engineer dedicated to the rollout." |
| Team | "You join a data team of four (two engineers, one analyst, one PM), reporting to the Head of Data, who ships code herself." |
| Range & process | "EUR 90,000-110,000 by level. Process: intro call, technical deep-dive on a past project, half-day on-site with the team, offer, three weeks end to end." |
Common mistakes when writing the ad
- 1Writing the ad before the requirements profile exists, so the ad becomes the spec, backwards and unprioritized.
- 2Letting HR templates dictate structure: company history first, problem last, exactly inverted from how candidates read.
- 3Listing every tool anyone on the team uses as a requirement instead of describing the stack as context.
- 4Hiding the seniority level: "senior" in the title with junior scope in the bullets wastes everyone's process time at offer stage.
- 5Never updating the ad: a description that does not change after learnings from the first ten conversations is a template, not a document.
