AI Conversational Agent — Pilot RFI

How to respond

1. Complete the matrix in Section 6, marking each row Native / Supported, Partial or via workaround, Not supported or Not applicable in the Response column, and a short note in Detail.
2. Provide the additional materials requested in Section 7.
3. Please let us know if you have any questions.

1. Project Context

PayJoy is evaluating AI conversational agent platforms to handle inbound customer conversations on WhatsApp (and potentially other channels) at scale, with a controlled hand-off path to human agents for cases the AI cannot or should not resolve on its own.

The pilot is inbound-first: customers reply to messages PayJoy already sends them. The AI agent picks up those replies, resolves the common ones end-to-end, and escalates the rest to specialized human teams (collections negotiation, customer support, fraud, legal) with full conversation context.

2. Current Messaging Stack

The WhatsApp number we use today is provisioned through Infobip. Any solution we adopt must reuse this number — happy to explore alternatives if feasible.

Reference Architecture

The pilot introduces an AI conversational agent to handle inbound WhatsApp messages from customers that today go unanswered. The design reuses the existing stack and splits responsibilities by message direction:

• Outbound (current state, unchanged): Databricks feeds segments and triggers; Customer.io orchestrates what is sent, when, and to whom; Infobip (WhatsApp Business API, existing number) delivers the message to the customer.
• Inbound (new in the pilot): when the customer replies, Infobip fires a webhook to the AI conversational agent. The agent detects intent, fetches customer context from Databricks (balance, due date, DPD), and responds — including a payment link — back through Infobip.
• Coexistence: Customer.io still owns outbound and the AI owns inbound; they sync (pause journey / notify outcome) to avoid double-messaging.
• Human hand-off: on low confidence, out of scope, negotiation, complaints, fraud, or legal, the agent escalates to the human team with a summary and conversation context; the conversation can return to the AI.

Key principles:

• Reuse the existing Infobip WhatsApp number (no re-provisioning).
• Any response with numeric data (amounts, dates) must be backed by a tool-call against Databricks — not generated freely by the LLM.
• Databricks is the source of truth for customer context.
• Pilot scope: inbound only, Mexico, default segment Collections DPD 0–30.

3. Problem Statement

Today our customer messaging is essentially one-way (outbound). Customer.io triggers millions of messages per month through Infobip, but when a customer replies, no one is on the other side.

Inbound replies on WhatsApp and SMS are not systematically monitored across countries. The current outcomes are: the customer calls our voice support line (expensive, long queue); the customer abandons the conversation and the underlying intent (pay, get info, raise a complaint) goes unresolved; operationally critical signals (e.g. a customer saying “I already paid”, “wrong number”, “the device is broken”) never reach the right team.

We have no agent pool large enough to staff inbound WhatsApp on the volumes we generate, and we do not want to build one. We want an AI agent to handle the long tail of low-complexity intents and route the rest to the humans who can solve them.

4. MVP Scope — What “Pilot Success” Looks Like

The MVP is intentionally narrow. We want to validate in one country, one customer segment, before scaling.

In scope for the pilot

1. Channel: WhatsApp inbound on the existing Infobip number.
2. Country: Mexico.
3. Customer segment: default — collections DPD 0–30.
4. Intents the AI must resolve end-to-end:
   • “How much do I owe?” / current balance
   • “When is my next payment due?”
   • “How / where can I pay?” (generate payment link or instructions)
   • “I already paid — why am I getting reminders?”
   • “My device is locked — help me unlock it after I pay” (top-level feasibility)
   • “Update my contact info” (acknowledge + route to system of record)
   • Basic FAQs (interest, fees, contract terms)
5. Hand-off to human when:
   • Intent is out of scope (negotiation, complaint, fraud, legal threat).
   • Customer explicitly asks for a human.
   • AI confidence is below threshold.
   • Hand-off must include a conversation summary and customer context for the agent picking up.
6. Basic analytics: volume, resolution rate, hand-off rate, CSAT, response time.

Out of scope for the pilot (but in scope for v2)

• Outbound conversations initiated by the AI (we want Customer.io to keep owning outbound for now).
• Voice channel (WhatsApp voice notes or phone IVR).
• Full Databricks personalization (we can start with a smaller context payload via API).
• Multi-country rollout.

Stretch goals (nice-to-have, won’t block pilot)

• Native Databricks / Snowflake connector for live customer context.
• Bidirectional sync with Customer.io (e.g. pause an outbound journey when the AI takes over the conversation).
• Multi-topic on the same number (collections + customer support + general inquiries).
• Maintenance / device-status questions answered from internal product data.

5. Sizing

The numbers below frame expected pilot and steady-state volumes. Inbound figures are estimated using industry-typical reply rates for WhatsApp Business in financial services (8–12%); we will refine with actual Infobip inbound data before contract.

Outbound message volume (current state, via Customer.io → Infobip)

Pilot sizing (proposed)

Steady-state sizing (post-pilot, if we expand globally)

Steady-state volume tiers (1M+ active users)

Tiers assume ~1 outbound message per active user per month (to be refined with actual Customer.io delivery data) and the same 8–12% inbound reply rate used above.

6. Technical Evaluation Matrix

Vendors should respond per row with one of: Native / Supported · Partial or via workaround · Not supported · Not applicable. Add a short comment in the “Detail” column.

7. Information We Need From the Vendor

Beyond the matrix in Section 6, please include in your response:

1. Pricing for the pilot volume in Section 5, plus a forecast for steady-state.
2. Reference architecture diagram for the Infobip + Customer.io + Databricks scenario.
3. Time-to-live estimate for the MVP scope in Section 4.
4. Two reference customers we can talk to, ideally fintech in LATAM with Infobip, or with other Meta Business Partners (e.g. Twilio, Sinch).
5. Sample conversation transcripts (anonymized) from a similar deployment.
6. Access to a live demo or production-like environment where PayJoy can interact directly with an AI solution deployed with existing clients.
7. Roadmap items relevant to our stretch goals (Section 4).