“Implementing their AI agents completely transformed our customer support pipeline. Hands down the best tech integration we’ve done this year.”
Sarah Jenkins
CEO at NexaFlow
Why sales roleplay is hard to run without software
Door-to-door sales training runs on the manager-plays-homeowner roleplay: one person acts out a reluctant homeowner in a conference room while the rest of the team watches. Knock is built around the run-up to summer, the window before the season starts. As a category, the conference-room format has structural limits. One rep practices at a time. Difficulty drifts with the manager's energy and mood. The feedback is spoken once and then gone, so nothing is recorded and nothing can be compared later. Software that replaces the manager-plays-homeowner session has to do three things at once. It has to be an opponent that speaks in real time and pushes back, because a pitch is a spoken performance and rehearsing it silently rehearses a different skill. It has to grade consistently, so two reps on the same difficulty are held to the same bar. And it has to keep a record, because coaching that cannot point at a specific moment is only an opinion.
What we built
- Live voice roleplayThe rep picks a scenario and has a real spoken conversation with an AI homeowner that listens, interrupts, raises objections, and either closes or turns them away based on how the pitch goes.
- Realistic AI homeownersHomeowners come with multiple personalities, male and female voices, and difficulty levels from receptive to highly resistant. Objections are industry-specific across pest control and home security, because those two conversations are not interchangeable — a rep is arguing against different concerns in each.
- Scoring and feedbackEvery session is graded on five dimensions: demeanor, objection handling, pacing, persistence, and close effectiveness. Each judgement cites specific moments from the conversation, so a rep sees the sentence where the pitch turned instead of a number they have to take on faith.
- Gamified daily practicePoints, team and company leaderboards, streaks, and rank progression give reps something to climb toward in the run-up to summer. Gamification carries the pre-season stretch, when a rep has no customer in front of them.
- Company knowledge and manager viewCompanies tailor scenarios to their own niche, products, and the real objections their reps hear. Managers get a team roster, activity, and score summaries. A candidate qualification mode points the same roleplay-and-score engine at screening new hires rather than training existing ones.
How systems like this are built
A spoken AI roleplay cannot be built the way a text chatbot is built. Text chat is turn-based: you finish typing, you press send, you wait. A doorstep conversation is full-duplex, because both sides can speak and either can start before the other finishes. In a real-time voice system the pipeline runs continuously through capture, speech-to-text, a model turn, synthesis and playback, and every hop spends part of a latency budget the ear reads as hesitation. The split between the model and deterministic code matters more than the model choice. A model plays the homeowner; a grading step turns the session into a critique. Everything else is ordinary software: which persona and difficulty loaded, which objection set applies to which vertical, and how points are awarded. Scoring weights buried inside a prompt are unauditable, so they belong in code where they can be read, versioned and tested.
The transcript is the substrate an AI roleplay platform sits on. Live display during the call and the cited moments in the feedback are both reads of one clean, timestamped record, which is why that record gets built first rather than bolted on after the scoring works. Building real-time voice is the same class of engineering problem across very different products. The constraint that shapes a training app, where audio has to arrive fast enough that a pause reads as thinking rather than as a broken connection, is the same constraint behind the enterprise AI voice agent handling 250,000 calls a day that we built for OraFacePass. What changes between the two is what happens after the words arrive: one books an appointment, the other grades a pitch. The capture, turn-taking and latency work underneath is recognizably the same discipline.
Why this is harder than it looks
Interruption is the hard part of an AI roleplay opponent, not a polish item. An AI homeowner who waits politely for the rep to finish trains the wrong instinct, because real homeowners talk over you. Letting the AI cut in means detecting a natural interrupt point, canceling audio that is already being synthesized or queued for playback, and recovering the thread without repeating a half-delivered line. The same problem runs mirrored when the rep interrupts the homeowner. Difficulty in a roleplay platform has to be a controllable parameter, not a mood described in a prompt. Two reps who select the same level should meet comparable resistance, and a persona that starts resistant should not soften halfway through because the conversation drifted. In practice that means persona state held outside the model, resistance expressed as structured configuration rather than adjectives, and objections stored as company-scoped content rather than hardcoded prompt text.
Scoring consistency is the constraint that decides whether a coaching platform is trusted. A grader that returns a different number for the same performance twice teaches reps nothing, and they stop believing the score. That pushes the design toward scoring each dimension separately against explicit criteria rather than asking for one overall impression, and toward citing the specific moment that drove each judgement so a rep can check the reasoning rather than take it on trust.
The stack
Knock's named stack is Flutter, which gives one codebase across iOS and Android, plus a real-time voice engine handling capture, synthesis and turn-taking. We are not naming the voice and model vendors for this build here. Across our work generally, Evolvv AI builds with Claude, Retell, n8n, HubSpot, Clay, Apify and Phantombuster, the same stack behind our AI agent development and workflow automation work. A platform of this kind also needs two components that rarely appear on a feature list: a timestamped transcript record that every other surface reads from, and a grader that returns structured output per dimension rather than a paragraph of prose. Structured output is what makes a score storable, comparable across sessions, and testable against a reference set. Prose is not. Neither component is visible to a rep, and both decide whether the product can be improved after launch.
Work with Evolvv AI
Evolvv AI is an AI agency based in Brooklyn, New York that builds custom AI voice agents, business process automation, and AI SaaS products. If you are weighing an AI product build, our AI agent development company page covers how we scope and deliver agent systems, and our AI automation agency page covers workflow automation. You can book a 30-minute call to talk through a build.
Frequently asked questions
How does AI sales roleplay scoring work?
Scoring in an AI sales roleplay platform works best as rubric-based evaluation rather than one overall verdict. In Knock, every practice session is graded on five dimensions: demeanor, objection handling, pacing, persistence, and close effectiveness. Each dimension gets its own judgement, and each judgement cites specific moments from the conversation. The citation requirement is the part that matters most. A bare number invites an argument the rep always wins, because there is nothing to point at, while a quote from their own pitch is checkable against what they actually said. It also constrains the model, since forcing it back to the conversation makes it harder to produce a plausible general critique that would fit any rep on any call. Reproducibility belongs in engineering rather than in the prompt. For any graded AI product that means fixed reference conversations, expected score bands, and a regression check re-run whenever a prompt or a model changes.
Why is a real-time voice AI app harder to build than a text app?
A real-time voice application is a different engineering problem from a text chatbot, not a heavier version of one. Text chat is turn-based: the user finishes, sends, and waits. Speech is full-duplex, because both parties can talk and either can start before the other has stopped. Building for that means running a continuous pipeline of audio capture, speech-to-text, a model turn, speech synthesis and playback. Every hop spends part of a latency budget, and a human ear hears the overrun as hesitation rather than as lag. Interruption adds a second layer. Supporting barge-in means canceling synthesis that is already in flight, discarding audio queued to play, and recovering the conversational thread so the speaker does not repeat a half-delivered line. Mobile adds a third layer: microphone permissions, audio session handling, an incoming phone call arriving mid-session, and what happens to the conversation when the app is backgrounded.
What is an AI sales roleplay platform?
An AI sales roleplay platform is software that lets a sales rep rehearse a live, spoken pitch against an AI character playing a prospect, then scores the conversation and gives feedback on it. Knock, built by Evolvv AI, is one: door-to-door reps practice against AI homeowners across pest control and home security, at difficulty levels from receptive to highly resistant, and are graded on demeanor, objection handling, pacing, persistence, and close effectiveness. The category sits between two familiar things. It is not a chatbot, because the conversation is spoken, interruptible and timed. It is not a learning management system, because the content is generated per session rather than authored once. Building one is an AI product build rather than an internal automation, which is the kind of work an AI agent development company takes on: accounts and roles, per-company configuration, onboarding for first-time users, and a support path when something fails on someone else's phone.
