A live agent talks to customers all day, and one prompt edit changes what it says to every one of them. You must ship fixes fast and prove nothing broke. This AI agent change management checklist ties versioning, approval, test gates and rollback together.
The checklist at a glance
Print this, paste it into your release ticket, and tick every box before a change goes live. It covers voice and chat alike. It follows the spirit of the NIST AI Risk Management Framework, which treats managing AI risk as work that continues after deployment.
| Gate | ☐ Check | Pass rule | Owner |
|---|---|---|---|
| Version | ☐ New version ID for every change | Covers prompt, model, voice, knowledge source, flow, tools, language and calling rules. A live version is never edited. | Agent owner |
| Version | ☐ Change record filed | Who, what, why and the test link are filled in | Author |
| Version | ☐ Calls pinned to a version | QA can filter scores by version ID | QA lead |
| Approval | ☐ Risk tier set | Low, medium or high is written on the record | Author |
| Approval | ☐ Approver is not the author | High tier also carries compliance sign-off | Named approver |
| Test gate | ☐ Talk to the agent yourself | A browser chat or a test call to your own phone runs clean | Author |
| Test gate | ☐ Disclosure and consent check | The agent says it is AI from the first interaction. Consent is marked on the campaign, and AI-voice calls need consent under the TCPA in the US. | Compliance |
| Test gate | ☐ Fixed case replay | Known failures, interruptions, unknown questions, language switches and handover requests pass | QA lead |
| Test gate | ☐ Channel parity | Voice and chat give the same facts, disclosures and handover | Agent owner |
| Test gate | ☐ Small live canary | A small sample runs with stop thresholds set in advance | Operations |
| Test gate | ☐ Canary scored against the old version | Same scorecard, and the new score is at or above the old one before traffic widens | QA lead |
| Rollback | ☐ Triggers defined | Score floor, any critical error, dead air, complaint spike, disclosure miss | Agent owner |
| Rollback | ☐ Pull-the-plug right named | One person, plus a backup, can pause without asking | Operations lead |
| Rollback | ☐ Last good version restorable | Restore was rehearsed, and the contact and do-not-contact lists are checked after | Agent owner |
| Rollback | ☐ After-action note | Filed after the incident, with transcripts reviewed | Author |
Use the checklist on every release
Start with the version, because everything else hangs on it. Give each change a new ID and count far more than the prompt. A version is the model and voice, the knowledge source, the script and flow, the tools and CRM field mappings, the language, and the calling rules. A "small" edit to retries changes who gets called and when, so it needs a version too.
Use a simple pattern like major.minor.patch. Major means a new model, voice or language. Minor means new content or a new offer. Patch means a typo fix. Never edit a live version. Ship a new one, and snapshot the old one first.
Then set the tier and the approver. Keep author and approver apart. Name one approver per tier and a backup, so a holiday never freezes a release. Regulated buyers, banks first, say their main blocker is an agent that "goes rogue" or makes a public mistake. They want proof from real deployments. Signed approval records are that proof, so store them with each version.
Next, pass the gates, and write each pass rule before you run it:
- Talk to the agent. Listen for tone, wrong facts and a missing AI disclosure.
- Replay a fixed set of cases. Keep one list and rerun it on every change. Include hostile callers too. Score each case pass or fail.
- Run a small canary. Banks tell us they want a short, low-risk pilot on a small sample. An insurer wanted its proof of concept to run on live, existing interactions instead of a sandbox, because of a past vendor experience.
- Score it against the old version. Use the same scorecard for both. Widen traffic only when the new score matches or beats the old one.
Define pass metrics before launch. Gartner's analysis of 432 AI customer service use cases found that only a quarter showed positive returns, and 42% could not be measured. A gate with no written pass rule cannot fail.
Finally, agree on rollback before you ship, so nobody debates it during an incident. Run the same steps in the same order every time:
- Pause the campaign.
- Send callers to a human or voicemail.
- Restore the last good version.
- Check the contact list and the do-not-contact list.
- Notify the owners.
- Review the affected transcripts.
- File an after-action note.
A rehearsed restore is faster and less error-prone than an improvised one, so we suggest a drill every quarter. Plan the human route with care. Buyers ask whether the handover is polite, how sensitive or disputed issues are handled, and who chases a missed handover. Gartner found that 87% of 3,566 customers say a company using GenAI for service must offer a way to reach a human.
In Insight7, you can test a campaign's agent in the browser with no phone number. A campaign also has a test call to your own phone, consent marking, calling hours, retry rules and a do-not-call list. A campaign can be duplicated with its contacts, and anyone added to the do-not-contact list since the original is left out of the copy. A dashboard shows calls in progress and lets a reviewer read each transcript when it ends. Every call can be scored against a weighted scorecard your team owns, so v1.4 and v1.5 can be compared like for like. Alerts can watch low scores and keywords and notify by email or Slack. A carrier's temporary refusal does not use up a contact's retry attempts. Voice campaigns are in beta, so confirm the campaign settings and the alerts in the app before you rely on them.
Adapt it to your risk and region
Let the tier pick the approver and the evidence. Adjust the rows to your own org.
| Change | Tier | Approver | Evidence needed |
|---|---|---|---|
| Typo in script | Low | Peer reviewer | Diff and one test chat |
| New offer or price wording | Medium | Product owner | Fixed case replay passes |
| New language, voice or model | Medium-high | Agent owner and QA lead | Replay plus scored canary |
| Disclosure, consent, escalation or compliance text | High | Compliance officer | Full gate record |
Tag any change that touches these regional rules as high tier:
- EU: any AI that talks to people must say it is AI from the first interaction.
- US: AI-voice calls already need consent under the TCPA.
- Colorado: public-facing conversational AI must disclose it is AI from 1 January 2027. Check the state's own published act before you rely on this.
- Nigeria: the CBN's March 2026 AML guidelines accept AI but expect transparency, governance and yearly independent model validation.
Add rows for each channel and language you run. A fix tested only on voice can leave chat on the old wording, so keep the parity check for every channel.
Related reading: How to Add an AI Agent to WhatsApp for Customer Service Using the WhatsApp Business API, How to Update a Live AI Voice Agent Script Without Breaking Compliance and AI Voice Agent Use Cases for Banks With Example Scripts for Inbound and Outbound Calls.
Questions people also ask
What is change management for AI agents?
It is the routine that controls every edit to a live agent. It sets who may change it, how the change is tested, who approves it, and how to undo it. No edit reaches customers without a record.
How do you version control an AI agent?
Treat the whole setup as one release bundle with its own ID, and freeze it once it is live. Keep each changelog next to its bundle, and keep old bundles restorable. In Insight7, a campaign can be duplicated with its contacts, and the do-not-contact list is respected in the copy.
Who should approve changes to a live AI agent?
Someone other than the author, chosen by risk. A peer can approve a typo fix. Compliance should approve any change to disclosure, consent or escalation text. Store each sign-off with its version.
How do you test an AI agent before going live?
Test in layers: a hands-on conversation, a fixed replay of known cases, then a small live canary scored against the old version. Insight7 offers a browser test and a test call before launch, in beta for voice, then a scorecard after.
How do you roll back an AI agent after a bad release?
Pause first and diagnose later. Restore the last good version, then read the transcripts of every call the bad version handled. Rehearsing the restore makes it faster and less error-prone.
What should trigger an AI agent rollback?
A single critical error or a missed disclosure should trigger it at once. Softer signals, like a score drop or a complaint spike, need thresholds set in advance. QA leaders say one critical error should fail a call that is otherwise "perfect" on procedure.
How often should you review or re-validate a live AI agent?
Review after every release. We suggest re-validating each quarter, and sooner when the model provider changes or the knowledge source is updated. Some finance rules expect independent model validation at least yearly.
Ship only what you can undo
Before any release, you should be able to name its version, approver, test result and route back. Pick one live agent and try it. If your agent erred at 9am, how fast could you restore the last good version?

