Motivational interviewing training: Practicing MI with an AI client




Most clinicians leave a motivational interviewing workshop able to explain the method, then stall when the next client says, “I know I should, but.” The skill lives in split-second choices about whether to ask another question or reflect what the person said, and those choices sharpen through practice that someone observes and scores.
Effective motivational interviewing training follows a sequence. It starts with a shared curriculum, moves to repeated practice and coded review of recorded sessions, and continues with coaching after the workshop ends. An AI client fits between supervised sessions, giving clinicians and coaches more reps for every coaching hour.
Motivational interviewing (MI) is a guiding, client-centered style of conversation about behavior change, in which the clinician draws out a person’s own reasons for change. The fourth edition of the Guilford Press text defines it as “a particular way of talking with people about change and growth to strengthen their own motivation and commitment.”
Most programs build the skill in roughly this order:
Short courses usually cover the first four steps well, and research on proficiency points to the last three as the steps that decide whether the skill lasts.
Nearly every course teaches the same core curriculum, whether it runs one hour or a full semester. The Motivational Interviewing Network of Trainers (MINT) lists open questions, affirmations, reflections and summaries as the core skills for person-centered engagement.
The spirit of MI is the clinician’s underlying intent, made up of partnership, acceptance, compassion and empowerment, an element MINT’s trainer manual also calls evocation. Four tasks (engaging, focusing, evoking and planning) then order the conversation. The fourth edition names these elements and frames the tasks as questions in its chapter titles.
OARS stands for open questions, affirmations, reflections and summaries, which MINT’s glossary calls “four basic client-centered communication skills.”
Reflections get the most attention in training, and MINT’s trainer manual recommends building reflective listening first, as “the most challenging of the four OARS skills.” Training separates simple reflections, which repeat or slightly rephrase the client, from complex reflections, which “make a guess, adding to and extending what was said.” A double-sided reflection holds both sides of the client’s ambivalence in one statement.
Change talk comes in seven kinds, summarized as DARN CAT (desire, ability, reasons, need, commitment, activation and taking steps). Clients also argue for the status quo and sometimes strain the relationship, so each kind of talk gets a different response:
Each response is chosen in the moment, so trainees need repetition on all three.
Clinicians build MI skill through practice that someone scores and discusses, such as role plays with scripted clients, real plays from their own cases and trainer demonstrations. A workshop on its own rarely gets them there. In a 2004 randomized trial of 140 licensed substance abuse professionals, only the groups that received feedback or coaching reached proficiency.
Even once the skill takes hold, it doesn’t stay on its own. A 2014 meta-analysis found that skills eroded within six months without feedback or coaching, and three to four sessions of feedback or coaching over that time held them steady.
Teams can keep skills steady with short weekly reps on one target skill and one coded recording a month. At each coaching session, the coach gives feedback on the two highest and two lowest ratings and sets the next practice target, as one coaching protocol does. The weekly reps are the hardest to staff, because they need a practice partner who is available every week.
Coaching hours are scarce, so the frequent, scored practice the research asks for rarely happens between workshops. An AI client supplies that practice by playing a client on demand and leaving a transcript a coach can score. It supports clinicians and coaches in training and has no role in patient care.
The form of the client counts too, because MI practice tests whether empathy lands. A Delphi study of telemedicine experts named facial expression, eye contact and tone of voice the most important cues for judging empathy. A face-to-face client adds presence, because the trainee is seen while reflecting and the client reacts to how something was said along with what was said.
An AI client can offer what a busy colleague or coach often cannot:
Sessions run whenever a trainee has time, including the weeks between coaching calls. One example comes from Tavus, a human computing company that builds the Personified Application Layer (PAL). A PAL is a real-time application you talk to face to face that sees, hears, remembers and responds.
How useful an AI client is for MI practice depends on a few setup choices:
In Tavus, a trainer writes the role into the PAL’s system prompt, then adds Objectives to set the goal and Guardrails to keep the client in role. A transcript callback hands the coach every turn after each session, and Memories let the client pick up where a trainee left off. For healthcare teams that handle patient data, the Enterprise plan adds HIPAA support and Business Associate Agreements, with a Zero Data Retention option.
PAL clients run on Tavus’s Conversational Video Interface (CVI), a pipeline for real-time, face-to-face video conversations. In CVI’s closed loop, Sparrow-2, Raven-2, the large language model (LLM) layer and Phoenix-4.5 work together, and each serves a skill MI practice depends on:
Every part runs during the live session, so the client speaks, listens and reacts in real time.
Picture a new care coordinator at a community behavioral health clinic, practicing with a client written as a man who drinks most evenings and is unsure about cutting back. The training lead set one Objective (the client voices a reason for change) and a Guardrail that keeps the client in role and away from health advice.
At her monthly coded session, the coach can compare two versions of the same moment. One is advice dressed as a reflection and the other is a complex reflection that drew out change talk.
Simulated practice produces reps, and a coach still decides what those reps should build. In a randomized study of simulated counseling practice (CARE), counselors who practiced alone saw their empathy decline, while those who also received feedback improved their questions and reflections. The authors concluded that training client-centered skills “requires more than simulated practice alone.”
AI clients can also be too agreeable. Peer supporters found one client’s “overly compliant and polished style” limited practice, and counselors in another study felt simulated patients mostly reflected motivated individuals. An AI client works best alongside trainers and coaches, with a coach picking each week’s target skill and reviewing the transcripts it produces.
Choosing MI training means picking a level and format, checking cost and CE credit, and deciding how skill will be measured.
MI courses sort into three levels. MINT’s introductory courses suit anyone learning MI, its intermediate courses suit those advancing existing skills, and advanced courses assume proficiency. Format then decides how much scored practice a trainee gets:
The cheapest formats describe no scored practice, and the formats with instructor feedback cap it at two or three sessions. Team delivery costs more, with APT USA charging $6,460 for up to 15 people and Prime for Life listing $8,000 for two days for up to 20, plus travel. Even a strong course leaves most of the weekly reps to the team after it ends, and an AI client can take on those reps.
Most MI courses for clinicians carry continuing education (CE) credit, usually one credit per hour, though each accrediting body sets its own rules. Psychologists, social workers, counselors, addiction professionals, nurses and health coaches answer to different bodies, and some require a passing quiz score, such as the 80% NAADAC asks for. MINT says its membership does not certify training competence, so teams should also look for post-workshop feedback and coding ability in a trainer.
MI skill is measured by coding a recorded conversation against a published instrument, most often the MITI 4.2.1 code or the Motivational Interviewing Competency Assessment (MICA). A MITI coder listens to a 20-minute segment once, rates four global scores from 1 to 5 and counts behaviors such as questions and simple and complex reflections. The manual suggests fair and good thresholds, such as 40% and 50% complex reflections, but notes that they rest on expert opinion and lack validity data.
Sessions with an AI client can be coded like any other recording, and services such as HETI code simulated client sessions. A PAL session produces what a coder needs, since Tavus delivers a timestamped transcript of every turn and can write the recording to a team’s own cloud storage.
A team plan works when it schedules the workshop, PAL practice, coded review and coaching as one program and tracks outcomes from the first week. A behavioral health or care team can follow these steps:
Track MITI scores and the percentage of complex reflections. Also track practice reps per clinician and coaching sessions attended, since studies with high coaching attrition showed skill erosion.
MI asks clinicians to stay with a person’s mixed feelings long enough for that person to find their own reason to change. Clinicians hold that presence under pressure only after practicing it many times, out loud, with someone across the screen reacting to what they say and how. More reps, scored and coached, mean the next client who says “I know I should, but” meets a clinician who listens first.
See it for yourself. Book a demo.
Starting a new skill with AI practice and moving to colleague role plays or real plays second is a reasonable order. One pharmacy-education study found students gained MI confidence from an AI-supported simulation used before a peer interview but not after, though it measured confidence, not coded skill.
Send one supervisor to supervisor-level training such as the ATTC’s MIA:STEP, built for people expected to give MI feedback and coaching. While that supervisor trains, buy coded feedback from an outside coding service and run a structured peer group.
Accrediting bodies’ published guidance does not appear to address whether AI-based practice counts toward CE hours, so confirm with your accreditor first. The documented route is an accredited activity that includes simulation. One example is a 0.50 continuing medical education (CME) credit obesity-care course that paired a micro-module with an AI-driven virtual-patient encounter.
Write the client role explicitly as low-readiness, because default AI clients tend to arrive cooperative. Skip decisional balance with this client, since a peer-reviewed article reports it is contraindicated for ambivalent people with low interest in changing. When little change talk appears, practice the pendulum approach, exploring sustain talk first and then swinging back.