TL;DR
Hire your first AI engineer after the anchor metric is set and after at least one pilot has shipped. The role is a production AI engineer, not an ML researcher or a data scientist. The skill stack is production deployment, observability, evals, and cost control. Contract first. Avoid hiring on research credentials alone. Plan a 12-to-20-week search.
- Hire after the anchor and the first shipped pilot.
- Production engineer, not researcher or data scientist.
- Skill stack: deployment, observability, evals, cost.
- Contract first when possible. Pay the market premium.
In this article
When to hire the first AI engineer
The most common mistake I see is hiring the first AI engineer too early. The pattern: a CEO reads about AI, decides the company needs an AI hire, posts the role, hires someone, and then asks them to figure out what to work on. Six months later the engineer has built a few prototypes nobody uses and is looking for the next job.
The right time to hire is after two things have happened:
- The anchor metric is defined. One P&L line, owned by a named human, that the AI program is going to move.
- One pilot has shipped to production. Using contractors, a consultancy, or existing engineering, the first AI capability is live and producing measurable output against the anchor.
Both of those happen without a full-time AI engineer. The reason to do them first is they produce the artifacts (anchor, shipped pilot, working production environment) that make the AI engineer role definable. Without them, the role is a guess.
For more on what the anchor and the first pilot look like, see The AI Transformation Playbook for Consumer Brands.
Hiring an AI engineer before the anchor is hiring a person with no scoreboard. They will spend two quarters building scoreboards before doing any work.
What the role actually is
There are three distinct roles people call "AI engineer," and they are not the same:
The ML researcher
Trains models. Publishes papers. Optimizes for novel architectures. Comes from a PhD background, often from a research lab or a frontier model company. Useful at a company whose product is AI itself. Almost never the right first hire at a consumer brand.
The data scientist
Builds analytical models, runs experiments, produces insights. The role most companies hired in the 2015 to 2022 era. Skilled at statistical modeling, less skilled at production deployment. Some data scientists transition into production AI work. Many do not.
The production AI engineer
Integrates models into production systems. Ships APIs, pipelines, evals, observability. Owns the operational reality of running AI in front of real users. This is the role you are hiring for at a consumer brand. The job is not training the model. The job is shipping the system that uses the model and keeping it accountable in production.
Confusing these three roles is the single most expensive mistake in AI hiring. Research credentials do not predict production output. Production experience does.
The skill stack that matters
For a first production AI engineer at a consumer brand, four skills matter more than anything else:
Production deployment
The candidate has shipped AI capabilities to real users, on real traffic, in production environments. They understand integration with existing systems. They have written API contracts. They know what failure modes look like at 3am.
Observability
They have built or run logging, tracing, and monitoring for AI workloads. They know what to log (inputs, outputs, latencies, errors, costs) and what to alert on. They have debugged production AI incidents and know what the telemetry should have told them.
Eval design
They understand that the model is only as good as the eval. They have built golden datasets, regression suites, and ongoing eval loops. They can answer the question "how do you know the new model version is better?" without hand-waving.
Cost control
They have managed unit economics on AI workloads. They know how to track cost per call, cost per user, cost per outcome. They have made decisions about which model to use where, and they can defend the decisions with numbers.
These four together are the bar. Candidates strong on three of four can be hired if they are honest about the gap and curious to close it. Candidates weak on all four should not be the first hire, regardless of pedigree.
The interview loop
The interview loop for a first production AI engineer has four stages.
Stage 1: Screening conversation (45 minutes)
What have they shipped in production? What is the AI system they are most proud of? What broke and how did they fix it? Most candidates have not shipped real production AI. Screen aggressively here.
Stage 2: Take-home or live exercise (2 to 3 hours)
Give them a small, realistic problem: integrate an LLM into a specific workflow, design the eval, set up observability. Look for clean code, clear thinking about the eval, and explicit handling of cost. If they cannot do this exercise, they cannot do the job.
Stage 3: Architecture conversation (60 minutes)
Walk through your actual stack. Have them propose an architecture for the next thing you would ship. Look for production thinking, not theoretical optimality. Push on edge cases, failure modes, cost.
Stage 4: Cross-functional and reference (60 minutes plus references)
The candidate will work with marketing, CX, ops, product. Have them meet two of those leaders. Look for the ability to translate business problems into technical work. Then take references seriously. The reference call is where you learn whether they ship.
Skip the trivia questions. The candidate who can recite transformer architecture details but cannot debug a production incident is the wrong hire.
Comp and the market reality
Production AI engineers are in shortage. The candidate pool is smaller than the demand. Comp reflects this. Two principles:
Pay the market premium. A senior production AI engineer commands a premium over a senior generalist engineer. Trying to hire at generalist comp gets you generalist resumes or a long search ending in a worse hire. The premium is not large in absolute terms compared to the cost of a missed quarter on the anchor metric.
Think total comp. Equity matters more for early candidates because they are betting on the company. Base alone does not close the right hire. Build a total package that includes meaningful equity and a clear path to scope expansion.
I am not publishing comp numbers here because they age in months. Use Levels.fyi or the public Anthropic and OpenAI bands as reference points for the high end. Adjust for your stage, market, and remote policy. Then add 10 to 20 percent for the production AI premium versus generalist comp.
Contractor vs full-time
Contract first whenever the option exists. A three-to-six month engagement with a known consultancy or experienced contractor does three things:
- It ships the first pilot. You have a working system in front of users.
- It clarifies the hire. You learn what the AI engineer at your company actually does, day to day.
- It de-risks the search. You hire from a position of operational strength, not from "we need an AI hire" panic.
The contractor is often more expensive per week than the full-time hire would be. That is fine. The information value of the engagement is worth multiples of the cost difference. Many great full-time hires have come out of a contractor relationship that converted.
This is how I structure most of the engagements through Automatic. Ship the pilot, define the role, then either hand off to a full-time hire the brand recruits or stay on as a fractional anchor while the team grows.
What to avoid hiring for
The patterns that produce the worst hires:
1. Research credentials alone. A PhD from a top program is impressive. It also does not predict production output. The question is what they have shipped, not what they have published.
2. The "AI evangelist" personality. The candidate who is most excited to give the company-wide talk on AI is rarely the candidate who is most ready to ship the next system. Hire builders.
3. The candidate who has never debugged a production incident. Production AI breaks in specific, surprising ways. Candidates who have not lived through this will not anticipate it.
4. The candidate who hand-waves on evals. If they cannot describe how they know a model change is an improvement, they have not really shipped production AI. They have shipped prototypes.
5. The fancy-tool tourist. Watch for candidates who name-drop every AI tool but cannot describe what they shipped with any of them. The tooling churn rewards depth, not breadth.
The right first AI engineer ships systems in front of real users with monitoring you can audit. Everything else is signaling.
The bottom line
Hire your first AI engineer after the anchor metric is set and after one pilot has shipped. The role is a production AI engineer, not a researcher or data scientist. The skill stack is production deployment, observability, evals, and cost control. Run a four-stage interview loop. Pay the market premium. Contract first when the option exists. Avoid hiring on research credentials, AI evangelism, or fancy-tool tourism. Plan a 12-to-20-week search.
The right hire shows up in week one and shortens the cycle on the next shipped capability. The wrong hire shows up in week one and rebuilds the things that were already working. The difference is mostly in the hiring discipline before the offer is extended.
FAQ
When should you hire an AI engineer?
Hire your first AI engineer after the anchor metric is defined and after at least one pilot has shipped to production using outside help or contractors. Hiring before the anchor produces a person with no scoreboard. Hiring after the first shipped pilot produces a person with a clear mandate.
What skills matter most in an AI engineer?
Production deployment, observability, eval design, and cost control are the four skills that matter most. Model training is overrated for most consumer brands in year one. The work is integrating models into production systems with monitoring and unit economics that hold up.
Should I hire an AI researcher or an AI engineer?
Hire an engineer, not a researcher, for the first AI hire at a consumer brand. Researchers optimize for novel model architectures. Engineers ship working systems. Most consumer brands need shipped systems, not novel architectures. The exception is at a company whose product is AI itself.
How much does an AI engineer cost?
AI engineer compensation varies by market, level, and the specific role definition. The senior production AI engineer commands a premium over a generalist senior engineer because the role is in shortage. Budget for a meaningful premium and consider total comp including equity, not base alone.
Should I contract before hiring full-time?
Contract first whenever possible. A three-to-six month contract engagement with a known AI consultancy or experienced contractor lets you ship the first pilot, learn what the role actually needs at your company, and validate the hire definition before committing to full-time comp.
How long does an AI engineer search take?
A serious search for the first AI engineer takes 12 to 20 weeks from job opening to start date in most markets. The shortlist is small, the candidates are in demand, and the right hire is worth waiting for. Plan the timeline at the start of the program, not at the end.