Case Management Agent

What it does

Case Management Agent is an autonomous agent that creates and updates cases without a rep doing the data entry. It reads conversations (chat, voice, digital messaging) and emails, predicts field values, and writes them to the case. It also covers resolution and closure tasks, configured separately.

Key facts

GA. Case lifecycle automation went to public preview on 10 April 2025 and GA on 31 October 2025 under the 2025 wave 2 plan. Enhanced admin options reached GA on 5 January 2026, and simulated AI case enrichment on 6 March 2026. Shadow mode is preview from March 2026 with GA planned for May 2026.
Customer Service only. It is not part of standalone Contact Center.
Consumption-billed. You need a Power Platform pay-as-you-go plan against an Azure subscription plus the Customer Service consumption-based billing setup. No PAYG, no agent.
Two Power Platform admin center prerequisites people miss: AI form fill assistance must be enabled, and Copilot data residency must be configured.
Predictable field types are lookup (up to 50 options per field), Boolean, integer, choice, currency, single and multiple lines of text, and email.
Case creation from chat requires authenticated chat. Anonymous chat will not produce cases. Voice requires transcription enabled on the channel. Email-sourced case creation runs on standard automatic record creation and update rules.
Cases are created when a rep accepts the conversation and again when they end it. The agent needs enough context to populate at least issue description or contact, or it won't create anything.
Update rules are configured under Customer support > Case settings > Case Management Agent > Manage > Case creation and update. Each rule has optional conditions, a list of Fields for AI prediction, and an Allow AI agent to override human edits option.
Rep-facing access is controlled in experience profiles, under Copilot AI features — autonomous case creation and update "From conversations", plus form fill assistance during case creation from conversation and from email. Out-of-the-box profiles have this on.
Simulation lets you test predictions before production: up to 100 case records from the org, or an Excel upload capped at 1 MB and 100 records. Output is a downloadable report with a prediction match percentage.
Shadow mode runs predictions against live cases without writing anything, but still consumes AI credits.
Required privileges include read on the msdyn_entityattributepredictionrules table, plus CSR Manager or Customer Service Representative.

When to use / skip

Worth it where reps handle high volumes of short conversations and spend meaningful time on case admin — telco, retail, utilities support desks. The value is in seconds saved per interaction multiplied by a large number, so the business case needs volume to work.

Skip it on low-volume, high-complexity desks. If a case takes forty minutes and involves three specialists, the two minutes of data entry is not the problem, and you'll spend more on tuning lookup descriptions than you save.

Also skip if the client's channel setup isn't there. Authenticated chat is a real prerequisite with real identity work behind it, and clients running anonymous web chat often assume this feature will just work for them. It won't.

Configuration decisions

Which fields go into Fields for AI prediction per rule. Start narrow — category, priority, product. Every added field is another thing to tune and another thing to explain when it's wrong.
Whether the agent may override human edits. This is the single most consequential toggle in the feature and it should default to off.
Which channels are in scope for autonomous creation: conversations only, email only, or both, and whether reps get AI-assisted creation (review before save) instead of full autonomy.
Rule conditions. An unconditional rule applies everywhere, which is how a general enquiry chat ends up creating a warranty case.
Whether lookup records get descriptions written for them, and who writes them. This is the actual tuning work in this feature and it's a content job, not a config job.
Hierarchical lookups: whether parent lookups need adding to the Quick Find Active view on child tables.

Gotchas

Lookup prediction accuracy depends on descriptions on the lookup records themselves. Out of the box, most clients have none, predictions are poor, and everyone concludes the agent doesn't work. Two or three sentences per record, saying when to use it and when not to, is the fix.
Hierarchical lookups need the parent lookup column in the Quick Find Active view specifically. Adding it to a custom view does nothing and there's no error.
Lookups are capped at 50 options per field for prediction. A product lookup with 400 rows is out of scope, and that's usually the field the client most wants predicted.
"Allow AI agent to override human edits" quietly undoes rep corrections. Turn it on without telling anyone and reps will report cases changing under them.
Shadow mode is safe for your data but not for the bill — it consumes credits at the same rate as live predictions.
The historical/consumption costs are not on a Customer Service licence. If the client believes their Enterprise seats cover it, that conversation needs to happen before the design is signed off, not at the first Azure invoice.
Turn on Dataverse auditing before go-live or you'll have no way to answer "why did this case change?" for AI-driven updates.

Consultant notes

Run simulations before anything else. Start with 20–30 records, look at the prediction match percentage per field, and use it to decide which fields are worth automating. It's the only evidence-based way to scope this.
Tell the client that lookup description writing is a deliverable with an owner and an effort estimate. It is the difference between a working agent and a demo.
Demo AI-assisted case creation (rep reviews before save) first, then autonomous. Clients almost always want to start at the assisted end, and it's a better landing.
Get the Azure subscription and PAYG plan in place during mobilisation. Procurement on the Azure side has killed more go-live dates on this feature than anything technical.
Before go-live, confirm authenticated chat is genuinely working and that voice transcription is on. Both are easy to assume and easy to find missing on day one.

Worth revisiting once shadow mode reaches GA, or if the 50-option lookup ceiling moves.