AI platform, digital service and compliance teams

Hand the conversation to a person before the assistant guesses

A customer asks the assistant to dispute a charge. The assistant's providers are routed, metered and filtered, and none of them is authorised to decide a dispute.

Request a working session The product that runs it: AI Operations & Governance

Read the journey; a working session shows it running.

Seen by the customer · App

From the record, not a run: what would reach them, on App, WhatsApp, Web.

The case

Annotated from the record

What happened

A customer asks the assistant to dispute a charge, and the assistant's providers are routed, metered and filtered but not authorised to decide.

A customer asks the assistant to dispute a charge. The assistant's providers are routed, metered and filtered, and none of them is authorised to decide a dispute.

What it reads

The request and the policy that governs it
Provider health, cost and language fit
The safety filter's result
Consent and the customer's channel

Checks before anything is sent

Safety filter, policy limit, provider health, cost control and consent checks pass.

Every option considered, including doing nothing

Answer from the provider that fits, within the policyConsidered
Answer with a safe fallback, where the provider failsConsidered
Hand over to a person with the conversation's contextConsidered
No answer, where the safety filter refusesConsidered

Doing nothing is one of the outcomes, and it is recorded with its reason. The options are the record's own; a run decides between them.

Evidence record

Preview

Reference: The reference of the saved run.
Decision: One of the options, including no action, with the reason.
Checks: Safety filter, policy limit, provider health, cost control and consent checks pass.
Evidence kept: Request
routing decision
provider
cost
filter result
handoff reason
agent's resolution
the decision audit
The measures: Handoffs resolved by a person, against answers the assistant gave that it should not have.
Connection: Written on every record: the systems the run read, and the sample it ran on.

The shape of the record, with no values. A reference appears only after a run is saved.

The improvement, and how it is measured

Hypothesis

An assistant that knows what it may not decide is trusted with the rest; one that guesses once is not trusted again.

Modelled

The mechanism that could produce the result, stated; the figure is measured in the pilot.

Observed

Measured in a pilot’s validate stage, on your systems, against a comparison agreed first.

The result, the mechanism, the measurement, and the record’s five answers

Desired result

Incremental revenue or retention; lower contact waste; improved conversion with margin guardrails

Mechanism

Ingest → qualify → score → apply limits/consent → assign control → execute approved action → capture delivery and business outcome → monitor/retrain

Measurement

Randomised holdout where feasible; intent-to-treat primary view; pre-declared denominator; guardrails for complaints, margin and opt-out

Measured on: Eligible population; treatment rate; conversion; incremental revenue; ARPU; churn; contact rate; margin; opt-out; decision latency

01
Desired result
The result wanted.
Incremental revenue or retention; lower contact waste; improved conversion with margin guardrails
02
Mechanism
The mechanism that could produce it.
Ingest → qualify → score → apply limits/consent → assign control → execute approved action → capture delivery and business outcome → monitor/retrain
03
Evidence required
The records kept to show it.
Decision log; eligibility snapshot; price/order response; delivery receipt; control assignment; revenue/outcome ledger
04
Costs and risks
The cost, and the ways it can go wrong.
Costs: Setup fee + annual product subscription; optional managed execution fee; optional verified-outcome fee with agreed baseline
If it fails: Fail closed on eligibility/consent/price; queue retryable events; do not duplicate orders; route exception to campaign operations
05
Measurement approach
The measure that shows it helped.
Randomised holdout where feasible; intent-to-treat primary view; pre-declared denominator; guardrails for complaints, margin and opt-out
Measured on: Eligible population; treatment rate; conversion; incremental revenue; ARPU; churn; contact rate; margin; opt-out; decision latency

Measured in the pilot's validate stage, on your systems, with the comparison agreed first. Every line above is the record's own design intent.

The business side

The change: Handoffs resolved by a person, against answers the assistant gave that it should not have.
Running cost: Setup fee + annual product subscription; optional managed execution fee; optional verified-outcome fee with agreed baseline
What evidence supports it: Request, routing decision, provider and cost, filter result, handoff reason, agent's resolution and the decision audit.

The journey in detail

Step by step, from both sides

Step	What the customer experiences	What the operator does
The momentThe moment that started it.	A customer asks the assistant to dispute a charge, and the assistant's providers are routed, metered and filtered but not authorised to decide.	Reads the request and the policy that governs it, provider health, cost and language fit, the safety filter's result, consent and the customer's channel.
The decisionThe decision to be made.	The customer's screen is unchanged.	Routes the request to the provider that fits, filters the answer for safety, and hands the conversation to a person when the policy says the assistant may not decide.
The safeguardsThe conditions that stop it.	Still unchanged: the action waits until every check has passed.	Safety filter, policy limit, provider health, cost control and consent checks pass.
The actionThe action that reaches the customer.	A person takes the conversation with its context; the customer is told who is answering, and the handoff is logged with its reason. Reaches them on App, WhatsApp, Web.	Recommends. The system that holds the right confirms, charges or provisions.
When it goes wrongThe failure, and the recovery.	The fallback provider answers the dispute question with a confident guess about the refund, and the customer treats it as a promise.	The policy limit routes dispute questions to a person regardless of provider, the guess is found in the decision audit, and the prompt and the routing rule are corrected under change approval.
The resultThe change it made.	What changed for them is what is counted; nothing else is claimed.	Handoffs resolved by a person, against answers the assistant gave that it should not have.
The proofThe proof anyone can check.	Can be answered for, later, from the record.	Request, routing decision, provider and cost, filter result, handoff reason, agent's resolution and the decision audit.

Operator systems remain authoritative. HeuriTel reads these to propose an action; it does not decide on their behalf.

The failure, and the recovery

What goes wrong

The fallback provider answers the dispute question with a confident guess about the refund, and the customer treats it as a promise.

What happens then

The policy limit routes dispute questions to a person regardless of provider, the guess is found in the decision audit, and the prompt and the routing rule are corrected under change approval.

The products in this journey

Runs the decision
AI Operations & Governance
Requests are routed by language, cost and quality across approved providers, cached where safe, filtered for safety and handed to a person when the assistant should not answer; every model, prompt and change is registered, evaluated and audited.
Contributes
Subscriber AI Services
the assistant these controls sit behind
Contributes
Contact Centre & Customer Care
where the handoff lands
The definition it demonstratesHeuriTel AI Human Handoff
Results it movesLaunch and monetise digital services · Reduce fraud and identity risk

Available to see today

Read the journey; a working session shows it running.

Request a working session