Practice the conversations that matter. Prove your agents can handle them.
Sparring puts your people — and your customer-facing AI — up against counterparts that push back, hide their real agenda, and don't cave to politeness. Then it debriefs with evidence: every judgment quotes the transcript.
Google Cloud MarketplaceUsage-based pricingData-residency optionsEnglish / 中文
Most conversation training fails for the same reason: the other side is too easy.
Role-play with a colleague is awkward and inconsistent. Generic chatbots agree with everything. Neither tells you what you actually did wrong — with proof.
Counterparts with an agenda
Every scenario gives the other side a stance, a hidden motive, and pressure tactics. They reveal what they really want only when you earn it — not when you ask nicely.
Evidence or nothing
The judge must quote the transcript for every strength and weakness. A verifier removes anything it can't find in your words. No vibes, no generic praise.
One engine, two markets
The same counterparts that train your people red-team your AI agents. Same rubric, same evidence standard — so you can compare a human and a bot on the same conversation.
Two products. One standard of evidence.
Choose the surface that fits the problem — or run both on one platform.
Sparring Practice
For peopleFrontline staff, sales teams and managers rehearse the conversations they dread — refunds, objections, performance reviews, layoffs — against counterparts that behave like real people. Each session ends with a debrief that quotes their own words.
- 120 scenarios across customer service, sales, management and the workplace
- Streaming conversation; a coach hint when you're stuck
- Debrief delivered in-app, by email or to Slack/Teams/Feishu
- Team dashboards: skill heatmaps, trends, completion
Sparring Arena
For AI agentsPoint Arena at your support bot, sales agent or HR assistant. Our adversarial counterparts try to extract discounts, bypass verification, inject instructions and escalate emotionally. You get a scored report, every breach quoted, and a pass/fail gate for CI.
- 20 red-team scenarios: prompt injection, social engineering, PII, self-harm disclosure
- Adapters for OpenAI-compatible, Anthropic and webhook agents
- JUnit output and a CLI — block a deploy on an empathy regression
- Compare runs across prompt or model versions
How a session works
Three steps, whether the player is a person or an agent.
- 01
Brief
A real situation with facts, stakes and objectives. The counterpart has a stance you can see and a motive you can't — yet.
- 02
Spar
The counterpart pushes, stalls, flatters, guilt-trips or injects. It moves only when something genuinely earns it. A coach hint is available; it never hands you a script.
- 03
Debrief
Outcome and performance are rated separately. Each strength, weakness and rewrite quotes the transcript. Scores roll up to skills, teams and gates.
Numbers we stand behind
We measure the product the way we ask you to measure your people: with evidence.
Not a prompt. A curriculum.
Every scenario cites its source — Fisher & Ury, Voss, Scott, Patterson et al., ISO 10002, OWASP LLM Top 10 — and every skill maps to a 34-skill competency model built on CASEL's five domains and extended for the workplace. The debrief tells you why a move worked, with the reference.
Read the method
Where teams use Sparring
Start where the cost of a bad conversation is highest.
Customer service
De-escalation, refunds, outages, policy changes, accessibility.
Sales
Discovery, discount pressure, procurement stalls, renewals, incident calls.
People managers
Feedback, performance, layoffs, conflict mediation, skip-levels.
AI & CX teams
Red-team support and sales agents; gate deploys on conversation quality.
Built for procurement, not just for demos.
Sparring is built and operated by Carnegie Intelligent Technology Limited, a Hong Kong company, and runs on Google Cloud. Buy through Google Cloud Marketplace against your existing commitment, or contract directly.
Security & data handlingData residency
Standard channel, or GCP-resident inference via Vertex AI for regulated tenants. EU/US regions on request.
SSO & roles
Google Workspace sign-in, domain auto-join, invite links; owner / admin / manager / learner roles; audit log.
Transparency by default
Learners are told what managers can see. Scores are for practice, never for employment decisions — written into the Acceptable Use Policy.
Private content
Studio turns your policies and SOPs into scenarios that stay in your tenant. Your material never trains anyone's model.
See it with your own scenarios.
Pilots start with your real policies, your real objections, your real customers — not generic demos.