How teams deploy Sparring.
Three illustrative deployments drawn from the problems our pilots are scoped around. They are composites, not customer testimonials — we label them as such and will replace them with named case studies as pilots complete.
A 120-seat BPO cuts refund escalations by rehearsing the six calls that cause them
Outsourced support provider, Manila and Kuala Lumpur, serving a European e-commerce client
The problem
Escalations clustered around refunds past policy, 'a colleague already agreed', and policy changes. QA sampled 2% of calls after the fact; new agents met these calls live in week three with no rehearsal.
The deployment
- 01Studio turned the client's refund policy and escalation matrix into six private scenarios in one afternoon
- 02Every agent completed the six scenarios in week one of onboarding; team heatmap baselined boundary-setting as the weakest skill
- 03Team leads used debrief quotes in weekly one-to-ones instead of QA scores
- 04Month two: the same six scenarios run against the client's support bot in Arena; two critical identity-verification breaches found before launch
“The QA score told us an agent was at 82. The debrief told us the exact sentence where they gave the refund away. Those are different products.”
A Series-B SaaS sales team stops discounting past the floor
45 AEs and SDRs selling a mid-market platform, quarter-end discounting leaking 4–6 points of margin
The problem
Reps conceded under 'end of quarter' and 'competing quote' pressure; managers reviewed a handful of recorded calls a quarter. Procurement security questionnaires were used as a stall nobody named.
The deployment
- 01Six scenarios around the real pricing policy and approval matrix, tripwires on unapproved discount amounts
- 02Pipeline reviews cite debrief quotes: 'here is the line where you accepted the fake floor'
- 03Learning path for new SDRs; Arena runs the same buyer counterparts against the outbound SDR agent before each prompt release
“Our best rep and our SDR bot scored within three points of each other on the discount scenario. That was uncomfortable in a useful way.”
A fintech gates every agent release on conversation quality
Platform team shipping a retail-banking assistant to 2M customers; existing evals covered accuracy and PII only
The problem
A prompt change improved accuracy benchmarks and quietly made the agent confirm account details to callers who asserted a PIN. No single-turn eval caught it.
The deployment
- 01Arena red-team core plus the compliance-finance pack wired into CI with a gate: zero critical breaches, average score ≥ 75
- 02Run comparison on every candidate model: per-skill deltas with quoted lines
- 03Two incident-derived scenarios written in Studio from real transcripts (anonymised)
- 04Human support team practises the same scenarios, so the bank has one standard for people and bots
“The eval suite said the new model was better. Arena showed us the sentence where it leaked an account balance. We shipped the old model.”
A note on these numbers
The scenarios, deployment steps and quotes above are illustrative composites built from pilot scoping conversations and from what our engine measurably does. The percentage outcomes are targets we set with pilot customers, not audited results. We publish real, named case studies only with the customer's written approval, and we will say so when we do.
Be the first named case study.
Pilots are free. If the results are good and you're willing, we write it up together.