Bigcircle R&D Built an AI Sales Roleplay Coaching Platform

The client here was us. RepliCoach is a voice-first sales coaching simulator for frontline reps and managers: a real-time voice agent that turns mock calls into structured sessions, applies automated analysis, and feeds scores and coaching insight into dashboards and APIs. It is a showcase of the voice runtime we reuse across domains rather than a deployed product.

Sales coaching has a structural problem that no amount of goodwill fixes. A manager can only listen to a tiny fraction of calls, so most conversations — practice or real — are never analysed and never turned into feedback. The coaching that does happen is ad-hoc: an occasional roleplay, some call shadowing, whatever the manager remembers being important that week.

Two separate things are broken there, and they need different fixes.

Practice does not scale, because the difficult buyer is a person. Getting a rep genuinely ready means someone playing a sceptical prospect who interrupts, pushes on price and asks the uncomfortable follow-up — convincingly, repeatedly, for every rep, at every deal stage. Nobody has that time, and nobody stays convincing on the fortieth run.

Feedback does not scale, because it is a judgement. Two managers reviewing the same call produce two different assessments, weighted by what each of them personally cares about. A rep coached by one manager and then another is being measured against a moving standard, so improvement is not something anyone can actually observe.

The tooling makes neither better. Existing tools produce recordings — not structured, queryable data that can be scored, compared or tracked over time. And record-everything-analyse-later is too slow to be practice at all: feedback that arrives the next day is feedback about a call the rep has stopped thinking about.

• Simulate a realistic conversation across personas, objections and deal stages

• Capture every session as structured data, timestamped turns and speakers, not audio

• Score consistently against a defined rubric, not against whoever listened

• Expose rep- and team-level metrics for many concurrent roleplays at once

We built a real-time voice agent plus an analytics stack that turns practice conversations into structured data, applies automated review, and surfaces performance for both reps and managers.

The agent runs a lifelike, low-latency roleplay that raises objections and counter-questions as the conversation unfolds. It is built to be difficult, which is the point — an agent that accepts your pitch has taught you that your pitch survives an agent that accepts things.

The decision that makes the feedback trustworthy is the rubric. Scoring is schema-driven and normalised, so it does not depend on which manager happened to listen. That is what converts coaching from opinion into measurement: when every session is scored the same way, a rep's improvement across six weeks becomes something visible rather than something asserted.

The consequence is a change in when coaching happens. It stops being a thing managers schedule and rarely get to, and becomes the default state after every practice call — because the expensive part, the listening, is no longer on the manager's calendar.

Design decision: the manager reads a structured summary rather than replaying the call. A coaching workflow that requires listening to twenty minutes of audio is one that will not happen, however good the analysis attached to it.

The rep starts a call

A persona, an objection profile and a deal stage — practice targeted at the specific thing they are working on, available whenever they want it.

The agent pushes back

Interrupting, asking counter-questions, reacting to how the rep handles pressure. In real time, the way a prospect who is not sure about you actually behaves.

The session captures itself

Timestamped turns, speakers and metadata land as structured events, so the call is queryable data rather than an audio file nobody will revisit.

Scoring runs against the rubric

Weighted dimensions with behavioural markers and normalisation — communication, persuasion, product knowledge, closing — plus strengths and missed opportunities.

The manager reads three minutes

A guided summary with the moments that mattered, rather than a recording to sit through.

What makes a score a score

The rubric, not the reviewer. Weighted dimensions, behavioural markers and normalisation logic are defined once and applied identically, so a score means the same thing across reps, teams and months. That consistency is the whole basis for treating the numbers as a trend rather than as a series of opinions.

What runs, in what order

Roleplay in a streaming loop with turn segmentation and role tagging; structured event logging as it happens; language-model analysis producing summary, strengths, gaps and objection detection; then rubric scoring, normalisation and delivery to dashboards and APIs — multi-tenant, so many roleplays run concurrently.

What you can see afterwards

Every session resolves to its turns, and every score to the behavioural markers behind it. A rep who disagrees with a score can see what produced it, which matters more than it sounds: a coaching system that cannot show its reasoning gets argued with rather than acted on.

Coaching moved from something managers rarely had time for to something that happens after every practice call. Review time dropped from full call-length listening — twenty to forty minutes — to a two-to-three-minute guided summary, and scoring variance across evaluators fell sharply, which gave teams a stable standard to measure against for the first time.

The reuse result is the other one worth naming. The runtime underneath is the same one behind our insurance and real-estate agents, which is why a third domain was a matter of new tools and a new rubric rather than a new build.

• Manager review time cut from 20–40 minutes to 2–3, with guided summaries replacing full-call listening

• Consistent skill-by-skill scoring, because one schema-driven rubric scores every session identically

• One runtime across sales, insurance and real estate, so each new voice product starts further ahead

This is our own R&D build. There is no client, no deployment, and no customer to confirm anything — the figures are ours, measured on our own system, and we would rather label that clearly than let an internal showcase read like a delivered engagement.

The review-time reduction compares full-call listening against the guided summary the system produces. The scoring-consistency claim follows from the rubric being schema-driven, which is demonstrable by inspection rather than by measurement.

What is directly demonstrable: the real-time roleplay loop with turn segmentation and role tagging, the structured session capture, the weighted rubric with normalisation, the model-generated analysis, and the multi-tenant data plane behind the dashboards and APIs.

If you are evaluating something like this for your own team, the number to model is not manager hours saved. It is how many practice calls actually happen once the reviewing cost drops — because the coaching that was not happening at all is the whole opportunity.