Verified self-service on WhatsApp. Bankers approve what policy reserves for them.
Card questions, transfers, disputes, limits, loans, alerts. Agents handle them under policy cards written by your risk team, execute through the core-banking component, and hand to a banker — with the draft ready — when the amount, the risk score or the regulation says so.
Three conversations a retail bank runs every minute.
Each one uses the same five stages: channel, identity, resolution, handover, close. The difference is the identity level and the policy card.
A reference configuration for retail banking.
Every deployment adjusts these. The point is that each request has a mode, an identity level and a policy card — written down, versioned and auditable.
| Request | Mode | Identity | Agent may | Person when |
|---|---|---|---|---|
| Balance & recent transactions | Autonomous | L2 | Read accounts, explain merchants and holds | Never (sampled by supervisor) |
| Card freeze / unfreeze / replace | Autonomous | L2 | Execute in card system; issue virtual card | Repeated replacements in 30 days |
| Transfer to known payee | Autonomous | L2 | Up to daily limit per segment | Above limit or risk score ≥ 40 |
| Transfer to new payee | Approve | L3 | Verify, prepare, small amounts | Above small-amount band |
| Dispute a transaction | Autonomous | L2 | Open dispute, collect evidence, provisional credit per policy | Merchant response conflicts |
| Credit limit change | Approve | L2–L3 | Pull history, score, prepare case | Always signs off above band A |
| Loan / mortgage inquiry | Assist | L2 | Retrieve products, pre-qualify, book banker | Leads the advice conversation |
| Hardship / arrears | Assist | L2 | Recognize, prepare options within policy | Leads; approves arrangements |
| Address / contact change | Approve | L3 | Verify document, prepare change | Approves; cooling-off applies |
| Account closure, beneficiary change | Human only | L4 | Book the appointment, prepare the file | Always |
Built to pass the review of the bank's second and third lines.
We designed the agent model around what risk, compliance and internal audit ask for: who allowed this, what did the agent see, what did it do, and can we replay it.
Policy cards as controlled documents
Owned by the business, versioned, approved through your change process. The audit log references the version that allowed each action.
Complete, replayable log
Every conversation turn, tool call, input, output, identity event and approval. Exportable to your SIEM and records systems with retention you set.
Segregation of duties
Build permission, policy ownership, approval rights and administration are separate roles. Customer-built capabilities go through the same review gate as ours.
Simulation before change
Replay last month's conversations against a new policy or model version; see containment, violations and the conversations to inspect — before anything goes live.
Human-in-the-loop by design
Handover isn't an exception path. Thresholds are the product's core mechanic, and the person always gets the prepared case, never a cold transfer.
Deployment where you need it
Private cloud region or dedicated infrastructure inside your boundary, including model inference. Data residency by tenant.
Risk components are agent tools — and they get a vote on every action.
Your fraud engine, device intelligence and behavioral signals plug in as components. The risk agent scores requests and sessions; the policy card says what a score means for each action. Alerts go out as conversations; confirmations, blocks, replacements and cases happen in the thread.
- Session risk: device, SIM change, location, velocity, social-engineering patterns in the conversation itself
- Step-up verification triggered by score, not by a fixed script
- Outbound alerts with structured yes / no, then block, freeze, replace, case — automatically
- Analyst handover with the full session and the agent's reasoning attached
Early, polite, in the channel people actually answer.
The collections agent contacts customers at the right stage with options the policy allows — pay now, new date, spread installments — executes the arrangement, and recognizes hardship language to hand over to a person with the file prepared. Everything is logged for conduct review.
- Stage-based journeys with consent, quiet hours and contact-frequency rules
- Payment links and direct-debit changes executed through payment components
- Arrangements within policy without a person; above it, prepared for approval
- Vulnerability and hardship detection routed to trained staff
Phased, with controls signed off before each step.
Most banks start with read-only self-service and inbound service, add transaction agents once the control framework is approved, and open outbound journeys last. Partners run the program; we support the first phases together.
Foundation
Channels, identity components, customer record sync, communication center for staff. Agents in shadow mode: they draft, people send.
Read-only self-service
Balances, transactions, card status, case status, FAQs from policy — autonomous with sampling. Handover with drafts live.
Transactions under policy
Card controls, transfers to known payees, disputes, limit requests in approve mode. Policy cards approved by risk; simulation on history.
Outbound & expansion
Fraud alerts, collections, renewals, onboarding. New request types added by the bank's own team with the build tooling.
Solution partner, technology partners, and flow.
A bank rarely buys a platform alone. The solution partner owns the core-banking integration, the control framework and the program. Identity, fraud and telephony vendors deliver as components. flow provides the platform, the agent model, reference policy cards and the build tooling the bank's own team uses afterwards.
Solution partner
Core-banking and card-system integration, control framework, program management, first-line support.
Identity & fraud vendors
IDV, device intelligence, behavioral biometrics and fraud scoring packaged as components.
Telephony & BSPs
WhatsApp Business solution providers and carriers delivering channels and numbers.
flow
Platform, agents, policy model, build tooling, reference configurations, joint delivery on the first phases.
Bring your top ten request types.
We'll map each to a mode, an identity level and a policy card, connect a sandbox core, and show the agent, the handover and the log — in a session with your risk team in the room.