When Contact Becomes Harm: Customer Treatment by Design in AI Collections

Automation changes the economics of customer contact. Once a campaign can place thousands of calls at low marginal cost, the operational temptation is predictable: contact more people, more often, for longer.

That is not a collections strategy. It is a capacity strategy.

Recovery depends on whether the customer can act, understands the available path, trusts the sender, and receives a workable next step. Repetition may increase response at first. Beyond a point, it teaches customers to ignore the number, distrust the lender, or wait for a human channel.

The risk is higher with a natural-sounding AI voice. A rigid IVR is easy to identify and abandon. A generative agent can sustain pressure, vary its language, and appear socially responsive. Without explicit controls, it can industrialise poor conduct while producing polished transcripts.

A default is a financial state, not a character judgement. The collection system should protect the lender's claim without stripping the customer of dignity or agency.

Contactability is not collectability

Three conditions are often confused:

1. Contactable: a valid channel reaches the customer.
2. Engageable: the customer is willing and able to understand the interaction.
3. Resolvable: a payment, arrangement, dispute, hardship referral, or other valid next action is possible.

An automated dialler optimises the first condition. A responsible collections system must optimise the third.

This distinction changes campaign design. A customer who answered twice, explained a temporary income shock, and requested a callback next week should not remain in the same contact pool as a never-contacted account. A wrong-party contact should not become another attempt. A customer who repeatedly opens a payment journey but fails at the final step may need payment support, not a more forceful reminder.

The objective is not maximum contact. It is minimum sufficient contact for the right resolution.

What regulators are signalling across markets

The rules are jurisdiction-specific, but the direction is consistent: the licensed or regulated institution remains responsible for customer outcomes even when communication is automated or outsourced.

India: responsibility does not leave the regulated entity

The Reserve Bank of India's 2022 recovery-agent circular applies across banks, many NBFCs, housing finance companies, co-operative banks, financial institutions, and asset reconstruction companies. It prohibits intimidation, humiliation, privacy intrusion, threatening or anonymous calls, persistent calling, false representations, and recovery calls before 8 a.m. or after 7 p.m.

RBI's NBFC outsourcing directions go further into operating design. They place responsibility for service providers on the NBFC, require need-to-know access to customer information, and expect recovery agents to be trained in care, sensitivity, calling hours, privacy, and correct product communication.

TRAI's 2025 amendments to India's commercial communications framework distinguish promotional, service, and transactional communications. They require senders to notify the originating access provider in advance when using auto-diallers or robocalls and to state the intended objective. The exact classification and consent requirement for a collections communication should be confirmed with counsel and the relevant telecom provider; "the customer owes us" is not a safe substitute for channel governance.

United Kingdom: outcomes, understanding, and vulnerability

The Financial Conduct Authority has told firms across regulated markets to avoid excessive debt communication, use supportive language, signpost free debt advice, and make it easy for advisers to act for customers. Under the Consumer Duty, communications should be understandable, tested, and adapted to customer characteristics, including vulnerability.

That is a higher bar than proving a script contained the legally required sentence. It asks whether the customer could understand and act on the communication in practice.

United States: frequency and artificial voice are regulated questions

The Consumer Financial Protection Bureau's Regulation F creates call-frequency presumptions for debt collectors covered by the Fair Debt Collection Practices Act. The often-cited threshold is more than seven calls about a particular debt within seven consecutive days, or a call within seven days after a telephone conversation about that debt. The rule's scope and its application to first-party creditors are legal questions, not universal operating assumptions.

Separately, the Federal Communications Commission has confirmed that AI-generated human voices fall within the Telephone Consumer Protection Act's treatment of "artificial or prerecorded voice" calls. Prior consent requirements and exceptions depend on the call and relationship. A model sounding human does not move the call outside robocall law.

Seven controls for humane automation

1. A contact budget, not only a channel limit

Set a portfolio-level and customer-level contact budget across voice, SMS, WhatsApp, email, field, and partner activity. A limit per channel can still produce excessive total pressure when channels operate independently.

The orchestration layer should see all attempts, successful contacts, customer requests, promises, disputes, and field actions before choosing the next intervention.

2. A clear opening

The customer should quickly understand who is calling, on whose behalf, why the interaction is occurring, and whether the voice is automated where disclosure is required or appropriate. Do not make a customer spend a minute proving identity before knowing whether the call is legitimate.

Privacy still matters. The opening must avoid exposing debt information to an unverified third party.

3. A usable stop or change-channel path

An opt-out does not erase a debt, and some required servicing communications may continue under applicable law. But a customer should be able to stop an automated conversation, request a human, register a preferred time, or move to a usable digital channel without fighting the agent.

The system should distinguish refusal to use this channel from refusal to resolve the account.

4. Vulnerability and hardship exits

Words indicating bereavement, illness, job loss, coercion, fraud, self-harm, disability, or inability to understand should trigger a safe handoff, not a more persuasive prompt. The model should not diagnose vulnerability. It should recognise that ordinary treatment may no longer be appropriate.

The exit must be operationally real: a trained queue, retained context, a callback service level, and suppression of conflicting activity.

5. Minimal action packets for people and machines

Many collections operations share broad account dumps with field agencies or call centres because extracting a smaller set is inconvenient. The result is not necessarily better action. High attrition, uneven training, and time pressure mean the recipient may still lack the one fact that matters now while possessing many facts they do not need.

A better design creates a time-bound action packet:

• verified identity and authorised contact details;
• current amount and payment status;
• permitted action and prohibited action;
• relevant customer preference or vulnerability flag;
• approved offer or resolution range;
• last material interaction and promised next step;
• expiry time and required disposition code.

This follows the need-to-know principle and improves actionability. More data is not the same as more context.

6. Language that preserves agency

Compassion is not vague softness. It is accurate language that separates obligation from shame and gives the customer a concrete path.

Avoid false urgency, moral judgement, manufactured social pressure, and consequences the agent cannot substantiate. State the amount, due status, available option, consequence of inaction where lawful, and route to assistance. Acknowledge what the customer has already said.

7. Outcome monitoring by customer group

Average containment or payment rate can hide harm. Monitor opt-out, abandonment, repeat contact, complaint, broken promise, wrong-party contact, and human-transfer outcomes by language, product, delinquency stage, age band where lawfully available, vulnerability status, and channel history.

If one group receives worse comprehension or resolution outcomes, the institution should investigate before increasing volume.

Why reckless IVR underperforms

In a retail lending collections program, call-flow analysis revealed that important offer messages appeared late in the interaction. A meaningful share of customers ended the call before the offer was played. The campaign technically delivered the script; customers did not receive its value.

That failure is common because IVR design is often measured from the institution's perspective. Compliance text, authentication, reminders, and offer information are arranged in internal priority order. The customer experiences time, uncertainty, and friction.

The practical correction was not simply a warmer voice. It was to simplify the message, surface the relevant action earlier, improve the payment path, and use observed response to decide the next contact. Combined with better base selection and controlled suppression of field activity, the redesign contributed to a doubling of digital self-pay within six months.

The analogy is a branch queue. Repainting the branch does not help if the customer must visit four counters and repeat the same information. Generative speech is the repainting unless the underlying journey changes.

The board should ask for conduct-adjusted value

A business case should not present recovery uplift and contact cost in one table while complaints, opt-outs, repeat attempts, and downstream remediation sit in another.

Use a conduct-adjusted value equation:

Incremental recoveries

minus voice, model, telephony, and operations cost

minus human transfer and exception handling

minus complaint, remediation, and repeat-contact cost

minus adverse movement caused by inappropriate treatment

The final term is difficult to estimate. That does not justify treating it as zero.

Good treatment and business interest are not opposing objectives. Respectful, well-timed, actionable contact reduces wasted attempts, protects trust, improves the quality of promises, and reserves human effort for cases where judgement matters.

Primary sources

1. Responsibilities of regulated entities employing recovery agents — Reserve Bank of India, 2022
2. Directions on outsourcing of financial services by NBFCs — Reserve Bank of India
3. Telecom Commercial Communications Customer Preference Regulations amendments — TRAI, 2025
4. Joint regulatory expectations on debt collection — Financial Conduct Authority, 2024
5. Delivering good outcomes for vulnerable customers — Financial Conduct Authority, 2025
6. Debt Collection Practices, Regulation F — Consumer Financial Protection Bureau
7. Declaratory ruling on AI-generated voices and the TCPA — Federal Communications Commission, 2024

Regulatory references are an operating interpretation, not legal advice. Requirements should be confirmed for the institution, activity, and jurisdiction at deployment.