Applied AI is not one seat.

Applied AI has become a convenient label for several fundamentally different engineering jobs. Treating them as one seat makes the search wider, the interview loop noisier, and the mandate harder to explain.

The current NXT desk includes AI product and agent builders, AI infrastructure engineers, forward deployed engineers, agent platform engineers, and applied ML engineers. These are confidential individual searches, not a market sample. They do show why the work must be defined before the title.

Start with the production boundary

The fastest way to separate the seats is to ask where the engineer takes ownership and where the system can fail.

• AI product and agents: owns user-facing behavior, product loops, and the path from model capability to a usable feature.
• AI infrastructure: owns the systems that make AI workloads reliable, observable, scalable, and cost-aware in production.
• Forward deployed AI: owns technical delivery with customers, moving between the codebase, the deployment environment, and the room where requirements change.
• Agent platforms: owns runtimes, tool use, evaluations, APIs, SDKs, and the backend systems behind agent behavior.
• Applied ML: owns the path from data and models into measurable product behavior, rather than research in isolation.

Write the mandate before the profile

A useful brief names the first consequential outcome, the systems the person will own, and the tradeoffs they will make without a larger team around them. It should also say what the role is not. That exclusion is often more valuable than another line of preferred experience.

• What must this person ship in the first 90 days?
• Where does failure show up: in the user experience, infrastructure, customer deployment, model behavior, or developer workflow?
• Who is the closest working partner: founders, product, platform, customers, or research?
• Which decisions must this person make independently?
• What attractive background would still be wrong for this seat?

Interview the work, not the label

Once the boundary is clear, the interview can follow it. A product mandate should test product judgment and shipped behavior. An infrastructure mandate should test reliability and systems tradeoffs. A forward deployed mandate should test technical depth and customer judgment together. An applied ML mandate should test the full path from data to production.

The title should summarize the mandate. It should never substitute for one.

The practical consequence

A narrower mandate does not necessarily produce a smaller search. It produces a more coherent one. Candidates can understand the actual work, interviewers can evaluate the same job, and founders can compare tradeoffs instead of comparing titles.

Before opening an applied AI search, decide which production boundary the seat owns. The right title usually becomes obvious after that.