A contact center AI risk assessment checks four things before an AI agent talks to customers: the data it touches, how it behaves, which laws apply, and how you watch it after launch. This checklist gives pass/fail items for each.
Checklist at a glance: four groups, one score
Each group has an owner and a sign-off rule. Mark every item pass or fail. Tag each item "critical" or "standard".
| Group | Items | Owner | Sign-off rule |
|---|---|---|---|
| Data | 8 | Security | No critical fail |
| Behaviour | 8 | Ops | No critical fail, pilot passed |
| Compliance | 7 | Legal | Legal confirms local rules |
| Monitoring | 6 | QA | Owner and cadence named |
Scoring rule: any critical fail blocks launch. Every other fail needs a named owner and a fix date.
Risk and measurement belong in the same document. A Gartner analysis of 432 AI customer-service use cases found only a quarter produce a return. Another quarter lose money, and 42% have unclear value. If you cannot measure your agent, you cannot show it is safe or worth running. So the monitoring group is a launch gate, not an afterthought.
Data: know where every recording goes
Start here, because every other group depends on it. In our advice to buyers, the first questions are about where recordings live, which AI models touch them, and who owns them. Some buyers also require hosting in their own country. Security owns this group. Use "C" for critical and "S" for standard.
- Storage location (C). Name every place recordings and transcripts sit, including backups. A good answer is a data flow map, from phone line to model to CRM.
- Hosting region and residency (C). Confirm the region matches your rules. Examples: the EU's safeguards for moving personal data to non-EU countries, and the CBN's June 2026 order to keep payment transaction data on local servers from 1 January 2027. Get the region in writing.
- Models that touch data (C). List each model and sub-processor. A good answer says whether your data trains any of them.
- Data ownership (S). The contract says you own recordings and transcripts, and what happens on exit.
- Retention and deletion (S). Set a retention period per data type, with a tested delete on request.
- PII redaction before the model (C). Card numbers and IDs are masked before the AI sees them. Test with a sample call.
- Access roles and logs (S). Only named roles open recordings, and every access is logged.
- SOC 2 or equivalent, plus API scope (S). Ask for the current report. Check that CRM write access is limited to the fields the agent needs.
Context data is the item people skip. Teams in regulated finance often ask whether account data given to an AI agent, such as funding status, needs compliance sign-off first. Settle it before the build.
Behaviour: test what the agent says and does
Behaviour is where a public mistake starts. Ops owns this group. Run the items in test calls first, then in a small pilot. Banks tell us they want a short, low-risk pilot on a small sample before they commit.
- Invented promises (C). The agent never quotes prices, terms or outcomes that are not in its approved material. Ask it leading questions and check each answer against your policy.
- Unknown questions (C). When it does not know, it says so and offers a human or a callback. It does not guess.
- Handover to a human (C). A caller can reach a person on request. The human sees the transcript so far, and the caller does not repeat themselves.
- Sensitive and disputed issues (C). Complaints, fraud claims and distress go to a person at once. Write the trigger words and topics down.
- Scope limits (S). The agent declines topics outside its job and ignores attempts to talk it out of its rules.
- Tone and voice (S). Listen to ten test calls. Buyers fear AI that sounds scripted or robotic, so check pacing, accent and wording for your market.
- Browser test before phones (S). Talk to the agent in a browser first, with no phone number needed. Insight7 campaigns offer this in beta.
- Pilot passed (C). Run a small live sample. Set pass marks in advance for escalations, complaints and wrong answers.
Compliance: match every rule to a control
Legal owns this group, and rules differ by region and sector. Legal should confirm which ones apply to each call route.
- AI disclosure at the start (C). EU AI Act Article 50 has required automated phone systems to say they are AI since 2 August 2026.
- State and regional disclosure laws (S). Colorado's Chatbot Safety Act requires public-facing AI to disclose itself from 1 January 2027. List every state or country you call.
- Consent for AI-voice calls (C). The FCC's 2024 ruling brought AI voices under US consent rules. Store proof of consent per contact.
- Do-not-contact lists (C). Check every list before each dial, including numbers added since the campaign was copied.
- Calling hours (S). Set legal hours by the contact's time zone, not yours.
- Retry caps (S). Limit attempts per contact and per day.
- Local rules confirmed (C). Legal signs that sector rules, recording consent and complaint duties are covered for every region.
Monitoring: score every call, not a sample
Monitoring keeps the other three groups true after launch. QA owns it. Use "C" for critical and "S" for standard.
- Score every AI and human call (C). Use one weighted scorecard for both, so you can compare them. Manual review covers about 2% of calls and leaves the rest unseen.
- Alerts and weekly digest (S). Low scores and risky keywords alert the owner, and the weakest calls arrive each week.
- Review and drift loop (C). Read flagged transcripts. Re-score a fixed set of calls after every prompt or model change.
- Owner and cadence (S). One named person reviews on a set schedule.
- Pause rule and incident log (C). Define who can pause the agent and what triggers it. Log every incident.
- Yearly re-validation (S). Have the model independently checked at least once a year. The CBN's March 2026 AML guidelines set that bar for banks.
Insight7 can score every call against a weighted scorecard your team owns. Scores roll up to the person, project and organisation. Alerts can watch keywords and low scores and notify by email or Slack. In beta, the campaign dashboard shows live calls and transcripts. You can try the free Call Quality Monitor on a few recordings first.
Gartner found customer-service AI spending rose 38% while overall service budgets rose 2%. Your leaders will ask for proof. Scored calls, an incident log and a dated review give you that report.
Related reading: 25 Security Questions to Ask AI Call Center Software Vendors Before You Sign, How AI Voice Agents Redact Sensitive Data From Call Recordings and Transcripts and How to Write AI Voice Agent Guardrails That Keep Conversations On Script in Regulated Industries.
Questions people also ask
What is a contact center AI risk assessment?
It is a structured review that scores a customer-facing AI agent on data, behaviour, compliance and monitoring. Every item passes or fails. Any critical fail blocks launch.
How do you do an AI agent risk assessment step by step?
Map the data flow first. Then test behaviour with test calls and a small pilot. Match each legal rule to a control, and name a monitoring owner. Tag items critical or standard, and give every fail a fix date.
What are the biggest risks of AI agents in customer service?
The main risks are leaked customer data, invented promises, and a broken handover to a human. Missing AI disclosure and silent drift after a prompt change are also common. Regulated buyers name an agent that makes a public mistake as their main blocker.
Does an AI voice agent have to say it is AI?
Often, yes. EU AI Act Article 50 has required automated phone systems to say they are AI since 2 August 2026. Colorado adds a similar duty from 1 January 2027. Disclose at the start of the call.
What are the AI voice agent compliance requirements for outbound calls?
Get consent before any AI-voice call. The FCC's 2024 ruling brought AI voices under the US TCPA. Also set legal calling hours, honour do-not-contact lists and cap retries.
How often should you review a customer-facing AI agent?
Score every call and read flagged transcripts weekly. Re-score a fixed set of calls after each prompt or model change. Get an independent model check at least yearly, the bar the CBN sets for banks.
Who should sign off a contact center AI risk assessment?
Each owner signs their own part: security for data, ops for behaviour, legal for compliance and QA for monitoring. One senior sponsor holds the final launch decision.
Review every call after launch
A risk assessment is a living score, not a one-time form. Every AI call after launch needs a review too. If your AI agent made a public mistake tomorrow, which call would show you first?

