A compliance lead at a regional bank hears the contact centre will pilot an AI voice agent next month. She has no policy to show. An AI customer communication policy is a short document, built in five steps, that ends in a nine-clause outline.

List every place AI talks to customers

Start by building an inventory. Your policy can only cover what you can see. Ask each team for every AI touchpoint, then record the channel, the direction and what the AI is allowed to do.

  • Channels: voice, chat, SMS and email.
  • Direction: inbound and outbound.
  • Listening AI: tools that score, transcribe or summarise calls. They never speak to a customer, but they still handle customer data.

Next, sort each touchpoint into one of three risk tiers:

  1. Inform: the AI shares facts, such as a branch hour or a delivery update.
  2. Advise: the AI recommends or explains, such as a product fit or a repayment option.
  3. Decide or move money: the AI approves, changes an account or takes payment.

The tier sets how strict every later rule is.

Regulated buyers, banks first, name risk as the main blocker. The usual fear is an agent that goes rogue or makes a public mistake. That is why this document comes first.

Begin with outbound calls in the inform tier. Banks also want a short, low-risk pilot on a small sample, so keep the first one small.

Policy outline, part 1: write the purpose, the scope (your inventory) and the definitions (the three tiers) in one page.

Write the disclosure and consent rules

Tell customers they are talking to AI, and get consent before an AI voice calls them. Laws in several regions already point this way, so write the rule once and set it to the strictest standard.

  • EU: Article 50 of the AI Act has applied since 2 August 2026. Any AI that talks to people, including automated phone systems, must say it is AI from the first interaction.
  • US, federal: AI-voice calls already need consent under the TCPA, following the FCC's 2024 ruling. A proposed "say it is AI at the start of the call" rule was still not final in mid-2026.
  • US, Colorado: The Chatbot Safety Act takes effect on 1 January 2027. It requires public-facing conversational AI to disclose that it is AI.

Your policy should say where the words go. On voice, put the disclosure in the first seconds, before any pitch or question. In chat, show it in the first message and keep a visible label in the window. For SMS and email, use the opening line.

Add two neighbours of disclosure, and confirm both with your legal team. First, consider a recording notice, spoken right after the AI disclosure. Second, set do-not-contact handling: check the list before every dial, honour an opt-out inside the same call, and re-check any copied campaign list.

Policy outline, part 2: write the disclosure and consent clause. It holds the exact wording per channel, who records consent, where the record is kept, and a rule that no campaign launches without it.

Set data, model and vendor rules

Write down what the AI may see, who may touch it, and what you will ask every vendor. Buyers in regulated firms ask where recordings live, which AI models touch them and who owns them. Some also ask for a SOC 2 report.

Data rules. Cover these points:

  • The customer fields the AI may read, and the fields it may never read.
  • Where recordings and transcripts are stored, and for how long.
  • Redaction of personal data before any text goes to a model.
  • Who owns the data and the outputs.

Residency. Name the rules that apply to you. EU GDPR transfer rules limit where personal data can go. In Nigeria, the central bank told banks and payment firms to keep payment transaction data on servers in the country by 1 January 2027. Ask your legal team which rule binds each data set.

Model governance. Use the regulators as your benchmark. Nigeria's March 2026 AML guidelines accept AI but expect transparency and governance. TechCabal reports that AI models need independent validation at least once a year. Require a validation date for every model.

Third-party clause. Send every vendor the same questionnaire. Ask about hosting location, subprocessors, audit rights and how fast they must report an incident.

Policy outline, part 3: write the data handling, model governance and third-party clauses.

Design the human handover and error rules

Write the route to a person into the policy as a requirement. Gartner surveyed 3,566 customers, and 87% said companies using GenAI for service must offer a way to reach a human. So handover is a control you must build.

List the triggers that send an interaction to a person:

  • A disputed amount.
  • A complaint.
  • A vulnerable customer.
  • A direct request for a person.
  • Silence, or a caller the AI cannot understand.

Then settle the design questions. Is the handover polite? Who owns the queue? Does anyone chase a handover nobody picked up? Give the queue one named owner and a time limit.

Speed matters because complaint handling is regulated. Regulators in several regions fine slow handling. In one example, Nigeria's central bank fined institutions N430M in 2025 for slow complaint handling.

Add a critical-error rule. One breach, such as a missing disclosure or a wrong figure, fails the whole interaction. A checklist score can hide this, because a call can follow every procedure and still leave the customer angry. Name your zero-score steps in advance.

Policy outline, part 4: write the escalation and critical-error clause. It holds the triggers, queue owner, time limit and zero-score list.

Monitor and review the policy

Score every AI call and every human call against a weighted scorecard, not a sample. A policy nobody checks is a wish. Set alerts on keywords and low scores. Keep an audit trail of each flag and who closed it. Book an independent model validation at least once a year, and review the whole policy on the same cycle.

Name your measures in clause 8. Gartner analysed 432 AI customer-service use cases. Only one quarter produced a return, another quarter lost money, and 42% could not be measured. A policy that lists its measures, such as disclosure rate, handover time and critical-error count, avoids that last group.

Insight7 can carry the monitoring load. It scores every call against a weighted scorecard the team owns, and scores roll up to the person, the project and the organisation. Its compliance monitoring scans 100% of calls for mandatory disclosures and builds an audit trail. Alerts on keywords and low scores can go out by email or Slack. Its voice campaign setup is in beta and includes consent marking, calling hours, retry rules and a do-not-call list.

The finished outline: nine clauses

Put the steps into one document:

  1. Purpose and scope: why the policy exists and which AI touchpoints it covers.
  2. Definitions and risk tiers: inform, advise, and decide or move money.
  3. Roles and owners: one named owner per clause, plus who signs off.
  4. Disclosure and consent: wording per channel, consent records, do-not-contact rules.
  5. Data and privacy: what the AI may see, storage, retention, redaction, residency.
  6. Model and vendor governance: validation dates, the vendor questionnaire, audit rights.
  7. Handover and escalation: triggers, queue owner, time limit, zero-score list.
  8. Monitoring, testing and incident response: measures, alerts, who acts and by when.
  9. Training, review cycle and version control: staff briefings, yearly review, change log.

Related reading: How to Get Internal Approval for AI Voice Agents From Risk, Compliance and IT Teams, How to Write AI Voice Agent Guardrails That Keep Conversations On Script in Regulated Industries and AI Voice Agent Business Case Template for Presenting to Leadership.

Questions people also ask

Does an AI customer communication policy have to say the AI is AI?

Yes. Write disclosure into the policy even where no law binds you yet. Rules are spreading region by region, and the EU already applies Article 50. Keep a record of each disclosure so auditors can see it. A full audit trail of flags also protects you.

What should an AI policy for a bank include?

Use the nine clauses in the outline above. If time is short, draft four first: scope with risk tiers, disclosure and consent, data rules, and handover. They cover what a regulator or auditor will ask for first.

Who should own the AI policy in a bank?

Give one accountable executive the whole document, usually the head of compliance or risk. Then name an owner per clause: contact centre, security, legal and QA each hold their own. A policy owned by a committee gets no updates.

How often should an AI customer communication policy be reviewed?

Review it at least once a year, on the same cycle as model validation. Also reopen it after a new channel, model or vendor, a regulation change, or a serious incident.

Do customers have to be able to reach a human?

Rules differ by region, so check your regulator. In the policy, make the request work in one step during the same interaction, with no need to call back or start over. Track how long customers wait for a person.

How do you monitor AI conversations for compliance?

Score every conversation against a weighted scorecard instead of a sample. Flag missing disclosures, banned wording and data breaches. Then have a reviewer confirm or override each flag, and keep the record of who closed it.

Can a vendor's AI be used with customer call recordings under data rules?

Yes, if your policy and contract allow it. Ask the vendor where recordings live, which models touch them, who owns them and whether it holds a SOC 2 report. Check residency rules for each data set. Redact personal data first.

Enforce a short policy

A short enforced policy beats a long one nobody reads. This week, list every place AI talks to or listens to your customers, and mark each low, medium or high risk. That page becomes your scope.

Book a demo of Insight7