Most AI hiring requests we see arrive as one line: "we need an AI engineer." Sometimes that is exactly right. More often the engineer is hired, builds something impressive in six weeks, and then discovers nobody decided what it was for, nobody can get it the data it needs, nobody is on the hook to keep it running, and nobody in the business changes how they work. The engineer was the right hire. It was just the third hire, not the first.
We staff AI work as four pairs that follow the life of a real use case: Plan decides what gets built and why, Build makes it work, Run keeps it working and safe, Adopt makes it legitimate and actually used. The twenty questions below walk that lifecycle. Answer them honestly and the gaps tell you which role to hire next. A stage where you cannot answer three or more questions is a stage with nobody owning it.
Plan: decide what gets built and why
- Can you name the three AI use cases you are not doing this year, and why?If the list of candidates has never been ranked, the first one that got funded was the loudest, not the best.
- For the use case you are building, what number moves, by how much, and who signed off on that number?A value case without a baseline and an owner is a hope. Finance will treat it as one.
- What will the system cost to run per thousand requests at the volume you expect in year two?Pilots are cheap. Production at scale is where model choice, caching and routing decisions either pay for themselves or quietly eat the business case.
- Who decides between building, buying and waiting, and what evidence would change their mind?If that person is the vendor, you do not have a decision process. You have a sales process.
- What does "done" look like, in a sentence a business user would recognize?"Retrieval precision above 0.9" is an engineering target. "An adjuster finds the right policy clause in under a minute" is a definition of done.
Missing answers here point to an AI Product Manager or an AI Solutions Architect.
Build: make it work
- Does an evaluation set exist before the first prompt is written?Without one, every improvement is an opinion and every regression is a surprise reported by a user.
- Which systems of record does the use case need, who owns each, and has anyone actually been granted access?The single most common reason a six-week build becomes a sixteen-week build is a data owner who was never asked.
- How will the system know what a given user is allowed to see?Retrieval that ignores permissions is a data breach waiting for its first curious employee.
- When did a real business user last see the working prototype?If the answer is measured in weeks, the build is drifting toward what engineers find interesting.
- What happens when the model is wrong?Every AI feature needs a designed failure path: a confidence threshold, a human handoff, a "not sure" state. Bolting this on after launch is expensive and public.
Missing answers here point to an AI Engineer or a Data Engineer.
Run: keep it working and safe
- If the model provider changes a model version on a Tuesday, how would you know by Wednesday?Quality drift is silent. The only defense is an evaluation harness that runs on a schedule, not a person who "keeps an eye on it".
- Can you trace one bad answer back to the prompt, the retrieved context and the model version that produced it?If not, every incident becomes a guess.
- What is the monthly inference bill today, and who gets paged when it doubles?Token costs scale with adoption. Success without a cost owner is a budget problem in disguise.
- Has anyone tried to make the system do something it should not?Prompt injection through a retrieved document, an agent with more tool access than its task needs, a user extracting another user's data. If nobody has tested for these, assume all three are possible.
- What is the rollback plan for a prompt change?Prompts are code. They need versions, reviews and a way back.
Missing answers here point to an LLMOps Engineer or an AI Security Engineer.
Adopt: make it legitimate and used
- Is there a written inventory of every AI system in use, with its purpose, its data and its owner?Regulators, auditors and your own board will ask for it. The organizations that have it built it before they were asked.
- Which risk tier does this use case fall into, under whichever framework applies to you, and what controls does that tier require?A customer-facing decision and an internal search tool are not the same category and should not carry the same paperwork.
- Who in the business is the named champion, and what do they get for being one?Adoption spreads through people, not memos. Champions without time or recognition stop championing.
- What changes in the daily workflow the day the system goes live, and what gets retired?If the old path stays open, most people will take it.
- How will you know, ninety days after launch, whether people are still using it?Active use, task completion and time saved are the numbers that justify the next use case. Nobody measures them by accident.
Missing answers here point to an AI Governance Advisor or an AI Enablement Lead.
How to use the result
Score each stage by how many of its five questions you can answer with a name and a number. A stage at four or five is covered. A stage at two or three needs a fractional specialist for a few weeks. A stage at zero or one is an empty chair, and hiring a Build engineer will not fill it.
The pattern we see most often: Build is well covered, Plan is partly covered by an executive sponsor, and Run and Adopt are empty. That is the profile of a company about to ship a pilot that works and then stalls. It is also the easiest profile to fix, because the missing roles are the ones that can start part-time.