.jpg)
Hiring for AI breaks the usual engineering playbook, because the resume can't tell you whether someone built a production system or finished a course.
So how do you tell? You change what you ask for in the job description and how you score the interview.
This page gives you job description templates and weighted interview scorecards for the five kinds of people who build AI inside a company. It also lays out the order most companies should hire them in. The full kit covers all five families plus forward deployed engineers, and it's a free download.
If you'd rather have someone else find the people, the AI and ML engineer staffing team at MSH recruits all five.
-------------------------------------
Get Our AI Hiring Kit
--------------------------------------
The Short Version
- "AI engineer" covers five different jobs. Architects, applied scientists, MLOps engineers, AI product leads and AI-enabled software engineers need different job descriptions and different interviews.
- Start with an architect who codes and a couple of engineers who ship, and save applied scientists for last.
- Structured interviews, where every candidate gets the same questions scored the same way, were the strongest single predictor of job performance in a 2022 meta-analysis (Sackett et al.). A weighted scorecard is how you get there.
- Require AI in the build or review round and score what they catch it getting wrong.
The Five AI Role Families
An AI engineer job description only works if you know which of five jobs you're hiring for. The families below come from the roles MSH recruits across the AI stack. Each one needs its own experience bar and its own interview. A posting that asks for two families at once tends to attract people who half-fit both.
Family A and the forward deployed engineer side of Family E already have dedicated MSH pages, AI and GenAI architect recruiting and forward deployed engineer recruiting.
The Five Families Hiring Order
Most mid-market companies plug into AI models someone else built. That changes who you hire first, and most AI team advice skips it because it was written for companies that train their own models. Nobody signs the specialist who only matters in the playoffs before they've got a point guard who can run the offense. An AI team works the same way.
Stage 1, an architect plus builders. Start with a hands-on architect (Family A) and one or two AI-enabled engineers (Family E). That's enough people to get one use case live in front of real users.
Stage 2, an adoption owner. Somebody has to make sure people use what got built. Plenty of AI work runs fine technically and still dies at this point because nobody owned the change in how people work. That's the job of your AI product and adoption lead (Family D).
Stage 3, platform help. Add MLOps and platform engineers (Family C) once two or three AI systems are in production and your architect can't carry the platform load alongside everything else.
Stage 4, applied scientists, if the model is your edge. Hire applied scientists (Family B) when fine-tuning or custom models give you an advantage worth paying for, because a scientist hired before anyone else on this list tends to build prototypes nobody in the business ever runs.
For the wider org design, see how to build an AI team and how to build an AI center of excellence.
Why AI Candidates Are Hard to Read on Paper
AI now writes a lot of the resumes you're reading, so polish tells you less than it used to.
Oz Rashid, MSH's CEO, likes to say it this way. "The resume is only 10% of the story. Just like the job description is only 10% of the story." Now candidates use AI to write the resume and companies use AI to read it. So a hiring manager ends up comparing two polished documents nobody fully wrote, and that 10% just got a lot smaller, right?
Gartner predicts that by 2028, one in four job candidate profiles worldwide will be fake. In the same release it reported a survey of 2,918 job candidates in which only 26% trusted AI to evaluate them fairly (Gartner, July 2025).
The first screen below works for every family. The second is for the engineering families, A, B, C and E.
- The production walk-through. Ask them to take you through a system real users relied on, then ask what broke after launch and what changed because of it. Push on the boring parts, like who got paged when it went down and who approved the cloud bill, because that's usually where notebook-only candidates go quiet. The quality of that story matters more than whether they can reverse a linked list on a whiteboard.
- Hand them AI-written code to fix. Give them a short piece of it with a few planted problems and ask them to make it production-ready. Strong engineers spot the import of a package that doesn't exist and the endpoint that returns another customer's records. If they paste it back into the assistant and hand you whatever comes out, you've learned what you needed.
Check That the Person Is Real
In June 2025 the Justice Department announced searches of 29 known or suspected laptop farms across 16 states, tied to North Korean schemes that used stolen and fake identities to land jobs at more than 100 US companies (US Department of Justice). The FBI has warned that those workers use AI and face-swapping technology in video job interviews to hide who they are (FBI IC3, January 2025).
Most of this fits inside the loop you already run. Keep cameras on and make sure the face at the offer matches the ID, and the kit has the rest of the checklist.
How the Weighted Scorecards Work
Every scorecard opens with three pass-or-fail knockout checks. A candidate who misses one doesn't go further. After that come five or six competencies, weighted to total 100 and scored 1 to 5 against written anchors. Each family also gets an interview loop that tells you which round covers which competency, plus a work sample you can run in about an hour.
Each interviewer only scores the competencies their round covers. Nobody should pretend a 30-minute screen told them how someone handles a production incident. At the debrief the hiring manager averages each competency across the interviewers who scored it and multiplies by the weight divided by 5, so a competency weighted 25 that averages 4.0 contributes 20 points.
Anchors spell out what a 1, a 3 and a 5 look like for each competency, so a 4 from your staff engineer means the same thing as a 4 from your product lead. Without them you get the debrief everybody's sat through, where two people both wrote down a 4 and meant different things. It's like calling balls and strikes with no strike zone. A 2014 review of structured interviews reports a meta-analysis of 19 past-behavior interview studies in which anchored rating scales predicted performance at .35, versus .26 without them (Taylor and Small, 2002, as reported in Levashina et al., 2014).
Evaluation rigor carries the most weight for applied scientists, and adoption and change management carries it for AI product leads. AI-enabled engineers put 25 of the 100 points on AI-assisted development judgment, which mostly comes down to knowing which parts of the AI's output to trust, because once the whole team has the same coding assistant that judgment is most of what separates one engineer from another.
Free Sample JD and Scorecard for AI-Enabled Software Engineers
The other four families and the forward deployed engineer variant are in the full kit.
Five Interview Questions That Separate Shipped From Studied
Here's one question for each family. Listen for a named system with a number attached, and if you get a textbook definition back, ask again and add "on your last project" to the end.
Each family in the kit has five questions with strong and weak answers written out. For the full sourcing and interviewing playbook, read how to recruit and hire AI and ML engineers.
Should Candidates Use AI in the Interview?
Yes, where it tells you something. You want to know they understand their own past work, so the tools stay off for the screen and the deep dive. The build or review round flips that and requires AI, because that's where you score how well they steer the tool and catch its mistakes. In a design round they can use it to look things up.
Some big engineering orgs moved first. Canva said in June 2025 that it expects backend, machine learning and frontend engineering candidates to use AI tools in technical interviews, and that candidates with minimal AI experience struggled with judgment more than with code (Canva Engineering, June 2025). Meta started testing something similar that summer, according to 404 Media's July 2025 report (404 Media) and a company spokesperson quoted by HR Grapevine (HR Grapevine, August 2025).
What AI Talent Costs, and Where the Premium Sits
How big the AI pay premium is depends a lot on whose data you read. It grows with seniority and shrinks for early-career and management roles.
Start with PwC's 2026 AI Jobs Barometer, which looked at job ads and found an average 62% wage premium for workers with AI skills (PwC, June 2026). Pay data tells a smaller story. Ravio's compensation data on more than 400,000 employees in engineering, IT and data roles at 1,500+ tech companies found 12% for individual contributors and 3% for managers (Ravio, December 2025), and 55% of the 3,413 employers Payscale surveyed pay no AI premium at all (Payscale, February 2026). Levels.fyi broke engineer pay out by level, and the premium climbs the whole way up, from 6.2% at entry level to 11.9% at mid, 14.2% at senior and 18.7% at staff (Levels.fyi, July 2025). PwC's figure is global, and none of these studies cover the same countries or roles. At 6.2%, a junior with AI in the title shouldn't cost much more than a junior without it.
For a baseline, the Bureau of Labor Statistics puts the median annual wage at $135,980 for software developers and $140,300 for computer and information research scientists, the federal category closest to applied AI scientists (BLS, May 2025 data, BLS). Those medians cover every industry and seniority level. Treat them as a rough reference for Family E and Family B, then add the level-based premium above if you're hiring senior or staff people.
Put your pay range in the posting no matter where you're hiring. A growing list of states requires it, including California, Colorado, Illinois, New York and Washington (California DLSE, Colorado CDLE, Illinois DOL, New York DOL, Washington L&I). The size thresholds run from any employer in Colorado to four employees in New York and 15 in California, Illinois and Washington. California's rule follows the job, so a remote role that could be filled there counts.
Frequently Asked Questions
What should an AI engineer job description include?A clear role family, the production systems the person will own, the cloud and model stack they'll work in, the must-have experience and a pay range. Add a short "how we'll interview you" section that states your rule on AI tools.
What's the difference between an AI engineer and a machine learning engineer?It depends on who wrote the job post. The usual split is that an ML engineer builds and trains models, while an AI engineer builds products on top of models someone else trained, usually by calling them through an API. In MSH's five families that product work is Family E, with model work leaning toward Family B and infrastructure toward Family C.
What interview questions should you ask an AI engineer?Start with "what did you ship and what broke after launch," then ask how they checked accuracy and when they wouldn't use AI at all. Add a review round where they harden AI-written code. The kit has five questions per family with sample answers.
Should candidates be allowed to use AI tools in technical interviews?Most teams split it by round, with tools off for the past-work deep dive and required for the build or review exercise. Tell candidates the rule in the invite.
Who should be your first AI hire?Usually a hands-on architect who still codes, plus one or two engineers who can ship. That architect can carry the platform work until two or three AI systems are live.
How do you spot a fake AI engineering candidate?Keep cameras on in every round, compare the ID at offer with the person you interviewed and verify past employers directly with the company. Ship equipment only to the address on the ID and hold system access until background checks clear, both steps the FBI recommends (FBI IC3, July 2025).
What is a weighted interview scorecard?It's a sheet that scores every candidate for a role on the same five or six competencies, on the same 1 to 5 scale, with a written description of what each score looks like. Each interviewer scores only the competencies their round covers. The weights decide which ones count most, and the total lands in one of four bands, from strong hire to no hire.
.png)
.jpeg)
.png)