Voice AI Collections Conduct Assurance Director

The mandate

The lender’s voice agents handle payment reminders and limited settlement journeys in several Indian languages, yet quality teams cannot trace a customer outcome through speech recognition, dialogue policy, retrieved account facts, generated wording and tool execution. The bounded problem is to determine whether conduct controls operate across language, vulnerability and contested identity before expanding account populations.

The named deliverable is a Voice AI Collections Conduct Assurance Casebook containing a journey-and-action inventory, consent and identity gates, language evaluation design, vulnerable-customer scenarios, negotiation limits, factual-grounding tests, human-transfer rules, call-evidence schema, control findings and a reusable release standard. It must separate model, telephony, account-data, policy and agent-operating failures.

Work starts 4 January 2027. Milestone one on 29 January is the agreed conduct taxonomy, sampled lineage baseline and urgent containment notice; milestone two on 12 March provides multilingual evaluations, call reviews, action tests and root-cause findings; milestone three on 28 May delivers the accepted casebook, witnessed assurance rerun, remediation priorities and operating transfer.

Acceptance requires conduct risk to select sampled calls and reproduce the chain from consent and account state through transcript, model and policy versions, tool action, escalation and final customer disposition. Independent fluent reviewers must agree material findings within the signed threshold, and every severe exception needs containment and ownership; aggregate promise-to-pay or call-completion improvement cannot establish acceptance.

The client will provide approved call recordings and transcripts, consent evidence, account-state snapshots, model and prompt versions, policy rules, tool logs, customer outcomes and secure review facilities. Collections, legal, language and engineering owners will resolve policy interpretation; the engagement excludes customer contact, debt decisions, legal opinion, system remediation and operation of controls after transfer.

Why this is external work

Collections leaders sponsor automation, while model teams see technical components and internal quality samples the existing script framework. No internal group independently connects generated interaction to vulnerable-customer treatment and executed account action across languages. External assurance supplies specialist challenge and an evidence trail without taking the lender’s continuing conduct accountability.

What you will own

• Map every voice journey from dial and consent through identity, account grounding, generated wording, negotiation, tool action, transfer and call disposition.
• Design multilingual samples across language, dialect, code switching, speech condition, delinquency stage, vulnerability and contested-account context.
• Test recognition and dialogue failures separately so transcript error, policy breach and unsupported generation receive the right control response.
• Verify that payment, promise, settlement and callback actions respect authenticated state, delegation, confirmation, idempotency and human-approval boundaries.
• Review fluent-language evidence for pressure, ambiguity, disclosure, interruption, dignity, refusal, vulnerability recognition and effective escalation.
• Reconcile automated quality scores with independent call review, complaint, reversal and later customer outcomes to expose false assurance.
• Transfer the casebook, sample logic, severe-exception route, release gate and repeatable evidence queries to conduct and quality owners.

Candidate qualifications

• Led conduct, model-risk or product assurance for conversational AI used in lending, payments, insurance or another high-consequence customer process.
• Evaluated voice systems across multiple Indian languages using fluent reviewers and customer outcomes rather than translated scripts alone.
• Traced a harmful interaction through recognition, retrieval, prompt, model, policy and tool layers to a defensible root cause.
• Tested transactional agent controls including identity, permission, confirmation, negotiation delegation, idempotency and human transfer.
• Presented material customer-conduct findings to a risk committee and can evidence a deployment population that was stopped or narrowed.
• Delivered assurance methods that internal conduct and quality teams reproduced after independent engagement closure.

Non-negotiables

• The named director must attend Mumbai discovery and acceptance, both operations-centre reviews and the independent language workshop.
• No recording, transcript, account data, model output or customer identifier may leave the lender’s approved review environment.
• Will issue a population stop where severe conduct evidence is unresolved, irrespective of collections performance or rollout commitments.
• Must disclose relationships with collections agencies, model providers, telephony platforms and the lender’s assurance suppliers.

1. 49 words maximum. Describe one harmful voice-agent outcome you traced across recognition, dialogue and tool execution, including the release decision.
2. 49 words maximum. How would you sample multilingual collections calls so fluent quality evidence captures vulnerability and power imbalance?
3. 49 words maximum. Confirm five-month capacity, required India travel and every collections or voice-platform relationship relevant to independence.