Your NPCs Can Talk Now, So What Does A Game AI Designer Actually Direct?

What Breaks First When NPCs Can Say Anything?

The demo goes well for ninety seconds. A guard answers a question nobody scripted, improvises a rumor about the north road, and the room grins. Then the same guard offers to open a gate the level does not contain, forgets that the player killed his captain an hour ago, and keeps chatting while a fire spreads behind him. Nothing in that collapse is a dialogue problem.

Three separate systems failed. There was no rule about what a character is permitted to promise, so the model wrote a quest the game cannot honor. There was no memory that survives a save and a reload, so continuity lasted exactly as long as the conversation buffer. And there was no arbitration saying that a burning building outranks small talk, so a threat and a chat both won at once. Every one of those is behavior design, and all three predate generative models by twenty years.

Studios have started naming the role rather than quietly absorbing it into gameplay engineering. Electronic Arts lists a Senior AI Designer for Battlefield in Montreal, role ID 212456, on its own careers site, and Riot Games carries a Manager for Software Engineering, Gameplay AI on VALORANT among its open roles ¹. Both sit on the gameplay side, distinct from the machine learning platform postings alongside them. The category is still forming, which shows up in the postings themselves: they disagree about whether this person writes production code, and you will have to decide that for your own studio before you write yours.

Which Backgrounds Produce Someone Who Can Direct Agent Behavior?

No degree program produces this title. The dependable feeders are combat and encounter designers who have spent years tuning enemy aggression, gameplay programmers who own the behavior tree and the navigation mesh, and technical designers fluent in whatever node graph the engine ships. Each arrives with half the job. Designers can say why a fight feels cheap; programmers can change it without filing a ticket.

The unexpected backgrounds are often the stronger ones. Animators and puppeteers have spent careers making intent readable in silhouette, which is the entire skill behind an enemy that telegraphs before it lunges. Tabletop game masters improvise inside hard constraints for four hours a week and know exactly how a fiction breaks when a promise cannot be kept. Robotics and simulation behavior engineers bring state machines, sensor budgets and graceful degradation as reflexes. QA leads who have written reproduction steps for nondeterministic bugs already think in seeds and traces, which is what makes a generative character debuggable at all; if that reflex is the gap on your team, it may be a testing hire rather than a design hire.

What all of them share is comfort with the idea that a character is a budget. So many decisions per second, so many tokens, so much latency before the illusion dies, so much authority to hand a system that will occasionally be wrong. The people who struggle are strong prompt writers with no shipping history: they produce a character that converses beautifully and cannot be tuned, staffed, profiled or cut down when the frame budget arrives. Where the work is mostly about the environment agents learn and act inside rather than the felt behavior, that is a different specialty with a different test.

Test Candidates On A Scene That Is Already Lying

Skip the portfolio walkthrough and skip the take-home asking for an NPC from scratch. Build one exercise from a slice of your own game: a character with a language model attached, no persistent memory, and a combat system that politely yields to conversation. Ninety minutes, an AI assistant allowed and expected, and one written page at the end saying what to change and in what order.

The tells separate fast. Strong candidates ask what the player is supposed to conclude from this character before touching anything, because that answer decides everything downstream. They write a refusal list, the things this character may never promise, before they write a single line of prompt. They ask what happens at the ninety-fifth percentile of model latency, and they have an answer for it that is not a spinner. They reach for the boring cases first: the reload, the second visit, the player who says nothing at all for a minute. Performed expertise rewrites the system prompt, adds lore, and comes back with a lovely conversation that answers none of it.

The AI-use beat is worth watching directly rather than asking about in the abstract. People who got good at this used models as a content pump behind a validator, never as the behavior itself. A practitioner will tell you about generating several hundred bark variants and cutting most of them because the tone drifted off-character, and will describe the check that caught the drift. They will have opinions about caching, about which decisions must stay deterministic for QA to survive, and about the moment they stopped trusting a model's own confidence. Ask for one instance from their own work where a model handed them something plausible and they caught the flaw. Practitioners answer with details; people repeating conference talks answer with a principle. If measuring that kind of drift is the whole of your problem, the measurement itself is a job.

Where Do Game AI Designers Leave Evidence Of The Craft?

Look for artifacts rather than titles, because the title is barely three years old on most resumes. The people doing this work leave behind talks, behavior trees, mod projects and postmortems about an enemy that felt unfair. Credits are a weak filter here: half the qualified candidates are credited as gameplay programmer, technical designer or combat designer instead.

The reliable venues are small and long-established. GDC's AI Summit has run for years and its speaker list is close to a census of the discipline. The Game AI Pro book series collects the field's working techniques, and its contributors are almost all currently employed practitioners. Studio careers pages are worth reading directly rather than through aggregators, because that is where these roles appear first and with the honest scope attached ¹. Beyond that, modding communities around simulation-heavy games produce people who have spent unpaid years tuning schedules, factions and pathing, and itch.io prototypes reveal more about behavior taste in ten minutes than a resume does.

Adjacent titles that convert well: combat designer, systems designer, technical animator, simulation engineer, and the AI programmer at a studio that never gave the work a design title. Sourcing move with the highest yield: find the person who gave the unglamorous talk about an enemy's failure state or a companion's pathfinding, and write to them. That talk was almost never given by a manager. If your actual bottleneck is coordinating this work across teams rather than doing it, that is a program role, not a design role.

What Does This Role Cost, And Does The Work Happen On Site?

No public wage series carries this title yet, and no credible survey isolates it, so a point estimate would be invented. Price it against the band you already pay a senior gameplay or systems designer at the same studio and seniority, because that is precisely who you will be hiring away, and expect to sit at the top of that band rather than the middle.

Say the proxy out loud when the requisition goes to finance, because "top of the senior gameplay design band" is a sentence you can defend and a scraped average is not. The studio hires against that band because it is the pool the candidate is leaving, and it pays the top of it because this person also carries a runtime system with a latency budget and a failure mode players can see. Name both halves in the requisition and the number stops being arbitrary. One broad market figure is worth having in the room and nothing more: PwC's 2026 AI Jobs Barometer measured a 62 percent average wage premium for roles requesting AI skills across roughly one billion job advertisements ². That is a direction across every industry at once rather than a band for one studio in one city, and quoting it as one will cost you the offer.

Money is rarely what loses these candidates. What closes them is authority: ownership of the behavior budget, a standing seat in encounter review, and a written rule that no model-driven character ships without a deterministic fallback. What kills the offer is discovering that the job is writing prompts for someone else's design, or that the character's behavior is decided in a marketing meeting. Ask any serious candidate what would make them quit in year one and listen for a version of that answer.

On location, this leans on site more than most AI-era roles, and honestly so. Tuning how a fight feels is done in a room with a build, a capture rig, dev kits and other people watching the same screen, which is why the Electronic Arts posting names Montreal at all ¹. Hybrid is the realistic norm. The systems half of the work, the arbitration rules, the memory schema, the evaluation rig, travels remotely without much loss; the feel half does not. Say which half your job is in the first recruiter conversation rather than in the final round, because candidates in this pool have strong and mostly reasonable opinions about it.

See a sample report >