Skip to main content
Back to blog

Practice Studio — AI roleplay training for human agents

QA on live calls tells you who needs training, but the fix is practice before the next live conversation. Devotel Orbit's Practice Studio lets supervisors author scenarios and agents rehearse against an AI that plays the customer — with an automatic score on every run.

Orbit Editorial Team

Most contact centers find every training need through QA on live calls — a scored conversation that tells a supervisor something went badly on a real customer. That says who needs work but not whether they have gotten better. The missing step in between is practice: a place to make mistakes without a customer on the other end.

Devotel Orbit's Practice Studio is that place. It lives under Quality → Practice in the dashboard, and it trains the human agent — not the AI — by having an AI play the customer.

What a Practice Studio is

The name is literal. A scenario is a rehearsed conversation: a persona for the AI customer ("angry about a duplicate charge, impatient but reasonable once heard"), an objective ("de-escalate, confirm the duplicate, offer a refund"), and a rubric — one criterion per line — that grades the run. The agent talks back and forth with the simulated customer; the scorer grades the conversation against each rubric item and returns a 0–100 score plus short written coaching feedback.

That inversion — a person being assessed while a model acts as the customer — is what distinguishes it from the AI-agent QA content elsewhere on this blog. Pipelines for evaluating the AI voice agent itself are a different topic; this page is about training people.

The surface is role-aware. Owners, admins, and supervisors get a "New scenario" button and see the whole team's session history. Other roles browse the active scenarios, start runs, and see their own history — while the API auto-scopes the session list, so an agent never sees anyone's runs but their own.

How supervisor authoring works

Scenario creation happens through POST /api/v1/quality/practice/scenarios, available to owner, admin, and supervisor roles. The dashboard form asks for:

  • Title — for example, "Angry customer disputing a charge."
  • Channel — chat, voice, or email (pick the medium the trainee works in).
  • Difficulty — intro, intermediate, or advanced, so the library can form a progression.
  • Customer persona — the behavioral spec the AI follows: mood, situation, what the simulated customer wants.
  • Objective — what a good agent outcome looks like, optional but useful.
  • Scoring criteria — the rubric, one line per criterion. The roleplay is graded against every line.

Once a scenario is marked active it shows up in every teammate's Scenarios tab. Reviewers can also filter to active-only; agents see GET /api/v1/quality/practice/scenarios?active=true filtered automatically.

Running a session and getting scored

An agent starts a run with POST /api/v1/quality/practice/sessions carrying a scenario_id. The conversation happens in-app; each turn is exchanged with the AI persona, and the run completes when the trainee ends it. On completion the scorer produces:

  • an overall score from 0 to 100,
  • short feedback naming which rubric criteria passed and where the conversation went off-script, and
  • the turn count, status (in_progress, completed, or abandoned), and timestamps.

All of that lands in the My sessions tab, and a KPI strip at the top of the page summarizes the whole history: runnable scenarios, total runs, average score across completed runs, and best score. Reviewers see the same average and best across the team's runs, so a supervisor can spot the agent whose average lags the scenario's difficulty and coach that specific rubric item.

What it does not do

The roleplay is an in-app simulation, end to end. No PSTN or SMS traffic is generated — the "voice" and "email" channel labels pick the medium the trainee rehearses in, not a real dial or send. Practice runs never touch a real customer, and the session list stays inside the tenant. This matters for compliance: a scenario that rehearses a regulated disclosure (identity verification, recording notice, refund terms) still exercises the agent's knowledge of the disclosure without contacting anyone.

It also does not replace live-call QA. Evaluations still sample real conversations; the leaderboard still gamifies the team's live metrics. Practice Studio sits between the two: evaluations surface a weakness, and practice closes it before the leaderboard reflects it.

Where it sits against the other quality surfaces

Three sibling surfaces live under /quality, and they deliberately split the loop:

  • [Evaluations](/quality/evaluations) — the record of LLM-judged scores on real conversations. Tells a supervisor which conversations failed which rubric.
  • Practice — the remediation loop for human agents. Rehearsal happens here, off live traffic.
  • [Leaderboard](/quality/leaderboard) — points and rankings fed by QA scorecards, handled volume, and CSAT across human and AI agents alike.

A working coaching routine runs through all three: the evaluation flags the gap, the supervisor authors (or assigns) a practice scenario matched to that gap, the agent runs it until the rubric passes, and the leaderboard reflects the recovered live-call quality. Practice Studio is the practice leg of that loop; it exists because remediation-by-live-call is an expensive way to train.

Example: the insurance claim intake scenario

A home-insurance team found their intake adjusters skipping the incident-date question on calls that later escalated. The supervisor authors:

  • Title: "Claim intake — storm damage, upset policyholder"
  • Channel: voice · Difficulty: intermediate
  • Persona: "Homeowner after storm damage. Upset, talks fast, gets vague on dates unless prompted. Calms down once the agent shows they have a full record."
  • Objective: "Capture a complete claim record — incident date, damage description, policy number — before routing to an adjuster."
  • Rubric:

- "Captured the incident date before proceeding" - "Verified the policy number" - "Got a full damage description" - "Closed with next steps and a reference"

Agents run the scenario; the scorer checks each line. A run that goes "I have the storm damage noted" without an incident date fails the first criterion explicitly, and the feedback says so. The agent re-runs until intake is complete on the first shot; the supervisor watches the team-average on that rubric item climb in the sessions list.

Frequently asked questions

Who can create practice scenarios?

Owners, admins, and supervisors author and manage scenarios; the API returns 403 to other roles. Agents and other roles can browse the active library and run sessions against it.

Does practice traffic ever reach a real customer?

No. The roleplay is an in-app simulation only — no PSTN, no SMS, no outbound traffic of any kind. The channel label on a scenario (chat, voice, email) describes the medium the trainee rehearses, not a live send.

How is a scored session graded?

The auto-scorer grades the full conversation against each criterion in the scenario's rubric, returns a 0–100 overall score, and writes short feedback naming which criteria passed. In-progress and abandoned runs show no score; the average-score card only counts completed, graded runs.

Can an agent see other teammates' sessions?

No for non-reviewer roles — the session list auto-scopes to the caller. Reviewers (owner, admin, supervisor) see the team's runs, including each agent's name on each row.

What happens when a scenario is retired?

Setting is_active false removes it from agents' runnable picker. Reviewers still see it in the management list (filtering to active-only shows what trainees see), so a scenario can be restored without rewriting it.

Getting started

Open Quality → Practice in the dashboard. A supervisor authors the first scenario; the KPI strip fills in as the team runs. Pair it with a weekly evaluations review, and the loop — flag weakness, author a scenario, rehearse, watch the score climb — runs entirely in one Quality area.

Practice Studio — AI roleplay training for human agents — Orbit by Devotel