Company Context
You’re the PM for MercuryPay, a Series D B2B fintech that provides payment processing, invoicing, and dispute management for 120,000 SMB merchants across the US and Canada. MercuryPay processes $38B in annual TPV and makes money via a blended 0.55% take rate plus subscription tiers for advanced features.
Customer support is a strategic differentiator: merchants frequently cite “fast, accurate help during payment issues” as a reason they stay with MercuryPay instead of switching to Stripe, Adyen, or Square. Support handles ~1.8M tickets/month across email, chat, and phone, with 24/7 coverage.
MercuryPay’s leadership wants to invest in generative AI, but the CFO is skeptical after several industry incidents involving hallucinated advice and compliance breaches.
User / Market Scenario
Primary personas
| Persona | Share of tickets | What they’re trying to do (JTBD) | Sensitivity | Current pain points |
|---|
| Owner-Operator (1–5 employees) | 46% | “Keep cash flow stable; resolve payment holds fast.” | High urgency, low tolerance for complexity | Wait times during weekends; repeats info across channels |
| Ops Manager (6–50 employees) | 38% | “Maintain payment uptime; reduce time spent on disputes/refunds.” | Medium urgency, wants auditability | Inconsistent answers across agents; long back-and-forth |
| Finance Lead (50–500 employees) | 16% | “Ensure compliance; get reliable documentation.” | Very risk-averse | Needs precise policy citations; escalations take days |
Competitive landscape
- Stripe is marketing “AI-assisted support” and claims a 15–20% reduction in time-to-resolution for common issues.
- Square has introduced in-product guided flows that deflect simple tickets (password resets, basic refunds).
- Several fintechs have paused AI rollouts due to incorrect chargeback guidance and PII leakage.
Problem / Opportunity
MercuryPay’s support costs have grown faster than revenue:
- Support operating cost: $9.2M/month (agents + BPO + tooling)
- Average handle time (AHT): 11.5 minutes (chat), 18.2 minutes (phone)
- First contact resolution (FCR): 54%
- Merchant churn: 1.9% monthly overall; merchants who open 3+ tickets/month churn at 3.1%
- CSAT: 4.1/5 overall; drops to 3.6/5 for “payment holds” and “chargebacks”
The Head of Support proposes an AI Support Agent that can:
- Draft responses for human agents (agent-assist)
- Answer merchants directly in chat for low-risk topics (self-serve)
- Summarize case history and recommend next steps
Leadership asks you to assess the business impact and decide whether to fund an MVP.
What you have (research + early pilot signals)
A 2-week internal pilot with 60 agents using an LLM drafting tool (no auto-send) showed:
- Draft acceptance rate: 62%
- AHT improvement on eligible tickets: -14%
- Agents report higher confidence on “how-to” questions, lower confidence on disputes/chargebacks
Ticket mix analysis:
- 35% “how-to / navigation” (low risk)
- 28% refunds & cancellations (medium risk)
- 22% payment holds / risk reviews (high risk)
- 15% chargebacks & disputes (very high risk, regulated + policy-heavy)
Your Task (deliverables)
- Define the product goal and vision for the AI initiative (what problem you’re solving and for whom).
- Assess business impact: build a structured approach to quantify value (cost savings, revenue retention, and strategic benefits) and identify key assumptions.
- Prioritize the first MVP scope: decide which capabilities (agent-assist vs self-serve vs summarization) and which ticket categories to start with.
- Design a measurement plan: propose success metrics, guardrails, and an experiment or rollout approach that isolates incremental impact.
- Call out risks and trade-offs: compliance, hallucinations, brand trust, operational readiness, and how you would mitigate.
Constraints
- Timeline: 8 weeks to ship an MVP to production for at least one channel (chat or email).
- Budget: incremental $250k/month for model + vendor + infra costs.
- Compliance: must meet PCI expectations; cannot expose full PAN; all outputs must be logged for audit.
- Legal requirement: AI cannot provide definitive advice on chargeback outcomes; must use approved policy language.
- Technical: MercuryPay has a data warehouse and ticketing system (Zendesk), but limited ML platform maturity; only 2 ML engineers and 3 backend engineers available.
Your answer should focus on how you would assess and communicate business impact to get an informed go/no-go decision, not just what the AI could do.
