From the first message to the closed case — and exactly where the person fits in.
A customer sees one conversation that gets resolved. Behind it, flow runs five stages, moves the underlying process forward, and decides at each one whether a human needs to be involved. This page walks through them with a bank as the example; the mechanics are the same for an insurer, a telco, a clinic — and for a supplier or an employee on their own desk.
Channel & match
WhatsApp, call, email, SMS or web chat. Matched to a record by phone, email or ID — or started as “unknown” and identified.
Identity
The identity agent picks the verification level the request needs and runs it with the components you've switched on.
Resolution
A service or transaction agent answers from policy and executes in your systems, within its policy card.
Handover
Above a limit, above a risk score, on request, or when unsure: a person takes over with everything prepared.
Close & learn
Summary and actions on the record, tasks created, metrics logged, corrections captured.
One queue, whatever the channel.
Messages and calls land in the communication center and are matched to the customer, supplier or employee they belong to. When the sender is unknown, the agent opens an identification flow instead of guessing. Nothing waits for a person just to be routed.
- Same conversation continues across channels — a call can end in WhatsApp
- Routing by language, skill, department, segment and SLA
- Business hours are a policy: after hours, agents keep resolving and queue only what needs a person
Verification is chosen by the request, not by the channel.
Each request type declares the level it needs. The identity agent enforces it using whichever components the deployment has: OTP, document + liveness, voice biometrics, device and behavioral signals, or your existing identity provider. A question about opening hours needs nothing; a transfer to a new payee needs more than a balance check.
| Level | What it proves | Typical methods | Unlocks (bank example) |
|---|---|---|---|
| L0 · none | Nothing — public information only | — | Hours, branches, product info, how-to |
| L1 · possession | Sender controls the phone on record | OTP to the number on record · WhatsApp number match | Case status, appointment changes, generic reminders |
| L2 · possession + device | Known device or app session, plus possession | OTP + device fingerprint · app deep-link confirmation | Balances, recent transactions, card freeze, transfers to known payees within limits |
| L3 · document + liveness | The person is who the record says | ID document capture + selfie liveness · partner IDV component | New payees, limit changes, address changes, loan requests |
| L4 · strong / human | Regulatory-grade or in-person assurance | Voice biometrics + banker callback · branch visit · qualified e-signature | Account closure, beneficiary changes, high-value operations |
Levels and mappings are configured per tenant and per request type; the table shows a typical retail-banking configuration.
Agents answer from your records and act through your systems.
A service agent reads the record and the knowledge base and answers with its source. A transaction agent calls the component — core banking, ERP, field service — and reports the actual result. Neither improvises: what they may do is on the policy card, and every tool call is logged.
- Read tools: accounts, transactions, cases, orders, policies, appointments
- Write tools: transfers, refunds, bookings, plan changes, case updates — each with limits
- Interactive elements: buttons, lists, confirmations, secure links, payment links
- Confirmation step before any write action, phrased for the channel
- Side questions handled without losing the main thread
When a person steps in, the work is already 90% done.
Handover is a first-class event, not a failure. The agent tells the customer what's happening, packages everything it knows and did, drafts the reply and the action, and places it on the right person's desk. The person reads, judges, clicks.
What triggers a handover
- Policy limit — amount, frequency, segment or action type above the agent's card
- Risk score — the risk agent flags the request or the session
- Customer asks — “I want to talk to a person” always works
- Agent uncertainty — no source, conflicting records, or low confidence
- Sentiment or vulnerability — distress, complaint language, hardship
- Regulated topics — advice, disputes, closures, anything your compliance team lists
What the person receives
- Verified identity and level, risk score, segment
- A summary of the conversation and the customer's intent
- Everything the agent already did, with links to the records
- A drafted reply in the right tone and language
- The proposed action, ready to approve, edit or decline with a reason
Every conversation ends on the record — and improves the next one.
The summary, the actions, the documents and the outcome are written to the customer, supplier or employee record. Tasks that came out of it are assigned. SLA and quality metrics are logged. If a person edited the agent's draft or rejected its action, the difference becomes a proposed policy change and a test case.
Every request type is assigned a mode. The mode decides how much of the work a person does.
| Mode | Who resolves | What the person does | Examples |
|---|---|---|---|
| Autonomous | Agent, end to end | Nothing — reviews samples via the supervisor agent | Balance, transactions, order status, appointment booking, document intake, invoice status |
| Approve | Agent prepares, person approves | Reads summary + draft, clicks approve / edit / decline | Limit changes, refunds above threshold, claim payouts, plan downgrades with retention offer |
| Assist | Person leads, agent co-pilots | Runs the conversation with live drafts, retrieval and actions offered inline | Complaints, hardship, complex product questions, negotiations with suppliers |
| Human only | Person | Everything, with the record and history on screen | Bereavement, legal disputes, regulated advice, anything the compliance list reserves |
Alerts, reminders and requests run through the same five stages.
A fraud alert, an overdue installment, an expiring document, a renewal, an appointment reminder — flow starts the conversation with an approved template, the agent handles the reply, and the record is updated. Outbound is subject to consent, quiet hours and channel rules per market.
- Approved WhatsApp templates with variables from the record
- Consent and preference checks before every send
- Replies handled by the same agents with the same policy cards
- Campaign-scale sends with per-person personalization and throttling
Walk through it with your own request types.
Bring the ten requests your contact center handles most. We'll map each to a mode, a verification level and a policy card — and show the handover on screen.