An AI Simulation Learning Designer Is Hired to Write the Wrong Answers Well

Everyone Passed the AI Role-Play and Nobody Got Better on the Phone

Ninety-four reps completed the new AI role-play module in its first month. Average score 88. Then a manager sat in on live calls and heard the same collapse she heard before the module existed: the customer says the price is too high, the rep drops to a discount inside twenty seconds. The simulation was not broken. The simulated customer accepted the first reasonable answer it was offered, because nobody had told it not to.

That gap is the role. An AI Simulation Learning Designer authors the practice: the situation, the character on the other side, the criteria for handling it, and how the difficulty moves as a learner improves. Vendors report that role-play inside a learning platform produces objective per-attempt metrics and faster onboarding, on the order of thirty to fifty percent ¹. Those gains come from the authorship, not the generation. The generation is a weekend. What is being practiced is not incidental either: customer operations sits among the highest-value areas in McKinsey's estimate that generative AI could add $2.6 trillion to $4.4 trillion annually ², and those are the same conversations ninety-four reps just failed on the phone.

The first trait to screen for is specificity about failure. Ask a candidate to describe the hardest customer in the domain you are training. A performed answer says difficult, resistant, emotionally charged. A real answer gives you a person: a facilities manager who already priced three competitors, who is not actually price sensitive but has been burned by a rollout that missed a deadline, and who will say the word expensive when he means unreliable. That distinction is what the persona prompt has to encode, and a designer who cannot say it out loud cannot write it.

The second is willingness to write a persona that stays hard. Language models are agreeable by default, and an untuned character rewards persistence rather than skill. Ask how a candidate would stop a simulated buyer from capitulating. Good answers get concrete fast: give the character a hidden goal and a private constraint, define what evidence would actually move it, cap how much ground it can concede per turn, and test the scenario against a deliberately bad learner to confirm the bad learner fails.

The third is assessment discipline. A designer who writes beautiful scenarios and then grades them on a five-point confidence scale has built theater. Press on how they would define one criterion, watch whether they describe observable behavior rather than an impression, and watch whether they define what a failing attempt looks like as carefully as a passing one. Rubrics that only describe good performance produce feedback nobody can act on.

Which Backgrounds Produce a Simulation Designer Who Can Write a Difficult Customer?

The obvious feeder is instructional design, and it converts unevenly. An ID who has spent a decade on compliance modules brings storyboarding, learning objectives and assessment vocabulary, all of which transfer. What sometimes does not transfer is dialogue. Writing a branching conversation where a character has motives is a different muscle from sequencing a slide deck, and it shows within one writing sample.

The strongest single feeder is anyone who has run live role-play for a living and watched it work. Sales enablement leads who have coached objection handling in a room know exactly which moment learners fail at and why the room version does not scale. Clinical and nursing simulation faculty are further ahead than almost anyone: standardized patient programs have used scripted human actors, structured debriefs and observable checklists for decades, which is this job with a person instead of a model.

The unexpected feeders are worth chasing. Improv teachers understand that a scene partner who agrees with everything kills the scene, which is precisely the persona-tuning failure above. Narrative and interactive fiction writers have shipped branching dialogue with state, and they think in terms of what the character knows rather than what the script says next. Crisis-line and mediation trainers have written escalation ladders. Former call center quality reviewers arrive already able to say what a good handling of this call sounds like, which is the rubric problem solved.

What almost nobody arrives with is all three halves. Expect to hire strong in two and build the third. A scenario writer can learn rubric construction in a quarter with a good assessment partner. An assessment specialist can learn persona tuning faster than a writer can learn to think in criteria. The profile that most often disappoints is the platform administrator whose experience is configuring the vendor's tool: real, useful, and not the scarce part. Ask for a scenario they wrote, not a system they launched.

One organizational note, and the opening scene is the reason for it. A scenario that punishes a twenty-second discount will lose to a comp plan that pays on closed volume every time, whatever the rubric says, so this designer ends up across the table from whoever does sales compensation design well before the first module ships.

Ask How They Got a Simulated Customer to Stop Being Nice

Ask the candidate how they personally got good at this with AI in the loop, and listen for iteration rather than prompt craft. The answer you want describes a specific scenario that failed in a specific way. The buyer folded. The patient volunteered the symptom the learner was supposed to elicit. The angry caller de-escalated on turn two. Then it describes what they changed, and how they knew it worked.

The strong answers share a method. They ran the scenario themselves as a deliberately incompetent learner to see whether incompetence was punished. They ran it five or ten times to see the spread, because a scenario that behaves differently on each run is not an assessment. They kept the transcripts of the runs that went wrong. Several will tell you they keep a file of persona instructions that failed, which is the closest thing this discipline has to a craft notebook.

The second thing to probe is how they use a model to draft and where they refuse its output. A designer who generates fifty scenarios overnight and ships them has produced volume, not practice. Good candidates describe using generation for breadth, first drafts, variant phrasings, edge cases they had not considered, and then hand-writing the two or three things that decide whether the scenario teaches anything: the hidden motive, the concession rule, and the criteria.

The skill underneath is checking a plausible thing against something outside the conversation. A model will happily produce a clinical scenario with a symptom pattern that does not occur, or an objection no buyer in the segment has ever raised. The designers worth hiring take drafts to a subject-matter expert and come back with the list of what was wrong. Ask for that list. If a candidate has never had a scenario corrected by a practitioner, they have been writing fiction.

Be careful about the interview format itself. This conversation rewards vocabulary. Someone saying branching state, persona constraints, behaviorally anchored rubric may have shipped four programs or read one article, and the transcript looks identical either way. Give them a real situation from the job and ninety minutes: write one scenario, the persona instructions, and a three-criterion rubric, then have them walk you through how a weak learner fails it. The same problem shows up in every judgment-heavy hire, and the same fix applies, which is putting the work in front of the person instead of asking about it, an approach also covered in hiring an AI oversight director.

Where Do You Find Someone Who Has Already Built Practice Environments?

Look inside the company before posting. The person who runs your live role-play sessions, the enablement lead who writes the call scripts, the clinical educator who runs the simulation lab, already knows which conversations go badly and why. They usually lack only the model side, which is the teachable part. The internal candidate also arrives with the domain relationships a scenario needs to survive review.

Outside, go where practice design is discussed rather than where learning technology is sold. The instructional design and learning-and-development professional communities are large and real. Clinical simulation has its own long-standing conference and society structure. Sales enablement has an established professional community. Interactive fiction and narrative game writing communities are full of people who have shipped branching dialogue with state and have never been recruited for training work, which makes them cheap to reach and easy to interest.

What closes this candidate, and it is rarely the money, is evidence that the simulations will be used and revised. Ask an experienced simulation designer about their last role and you will hear about scenarios that shipped once and were never touched again while the product changed underneath them. Name the review cycle. Name who owns updating scenarios when the pricing model changes. Give them access to real recorded conversations to write from, since scenarios built from imagination read as imaginary to the people being trained.

What kills the offer is discovering that the role is really content operations: uploading, tagging, scheduling, chasing completion percentages. It also dies when a candidate learns the rubric will be set by whoever bought the platform, or that success is measured by completion rate. Say plainly what outcome the program is judged on, and if the honest answer is that nobody has decided yet, say that and let them help decide. Where the training targets service conversations, expect this person to work weekly with the customer operations lead who owns the real queue the scenarios are drawn from.

What Does an AI Simulation Learning Designer Cost?

No wage series covers this title, and no compensation survey found for this piece prices it, so any number here would be invented. Hire against a proxy instead, and tell the candidate it is one: your senior instructional designer band, adjusted for whether the person owns the assessment criteria, which is heavier than authoring content, and for whether they tune the model layer themselves rather than handing specifications to an engineer.

Expect real competition from the platform vendors, who hire the same people to build reference content, and from enablement, where a strong scenario writer can be priced against a sales enablement manager rather than against a course designer. Domain depth moves the band more than tooling familiarity does: someone who can write clinically accurate patient encounters or genuinely technical procurement objections is scarce in a way that platform experience is not.

One caution about levels. The title is new, so a candidate's current title tells you almost nothing about scope. Ask what they were allowed to decide. Somebody who set the criteria and could refuse to ship a scenario was doing a different job from somebody who filled a template, even where the two share a title.

Where Does the Work Have to Happen, and What Belongs With Counsel?

The work is remote-friendly in its core: writing, tuning, reviewing transcripts, and revising. What resists remote is the input. Good scenarios come from sitting near the actual conversations, and designers who never hear a live call or observe a real consultation write plausible fiction. Budget travel or scheduled observation time deliberately, and treat it as the load-bearing part of the process rather than as a perk.

On-premise requirements appear where the source material is sensitive: recorded patient encounters, financial detail, anything under a data residency rule. The constraint there is usually the tooling rather than the desk, since transcripts and model calls have to stay inside a boundary.

Two questions belong with counsel rather than with an article. Where a simulation result feeds a decision about a person, whether promotion, certification or continued employment, several jurisdictions impose notice, explanation and record-keeping duties, and those rules differ by jurisdiction and are still moving through 2026. Separately, recorded conversations used as scenario source material carry consent obligations of their own. Check both in your jurisdiction before either becomes a design assumption.

See how it works >