Recommended Actions Agent

What it does

An aggregation and ranking layer over your other sales agents. Instead of each agent surfacing its own insights in its own interface, the Recommended Actions Agent collects actions from all of them, scores and ranks them globally, and presents the top item per record in a Suggested actions carousel on the Contacts, Opportunities, Leads and Accounts list pages. It reached general availability in September 2026, having been public preview, and now covers leads as well as opportunities.

Key facts

Sources today are the Sales Opportunity Agent, the Sales Qualification Agent and the Data Enrichment Agent, plus custom agents you onboard yourself.
The three built-in sources behave differently. Sales Opportunity Agent and Sales Qualification Agent actions are prioritised and charged. Data Enrichment Agent actions appear in the carousel but bypass the prioritisation engine, are stacked at the end, and are not charged.
Billing is consumption-based on the number of actions the prioritisation engine processes — not per seat, and not per action displayed. Actions shown without being prioritised don't count.
Ranking uses a four-dimension scoring model Microsoft calls UICE: Urgency, Impact, Confidence and Effort, each scored 0–10, combined into a weighted priority score. Agents also map their actions to prioritisation principles (Risk Reduction, Deal Velocity, Revenue Generation), which nudge the score.
Scores are derived from entity signals (deal value, stage, competitors), agent-provided signals passed through msdyn_prioritizationdata, and derived insights such as sentiment or deal health.
Ranking is global across all agents, then floor and ceiling rules filter or boost, and only the top-ranked actions reach the carousel. One card per record at a time; resolve it and the next one appears.
Sellers need Read Only Grid selected on the list page for the carousel to appear at all.
Admin configuration is a per-agent toggle: Create and manage agents > the agent > Connected skills > Recommended Actions. Turning it off leaves the agent running and generating insights in its own interface, just absent from the carousel.
Custom agents are onboarded by registering a custom source (name, unique source ID, description) and then calling the msdyn_PushActionDataToRecommendedActionAgent API in the required format.
Custom security roles need nine explicit privileges across msdyn_rawactioncatalogue, msdyn_prioritizedactioncatalogue, msdyn_recommendedactionsourceagentconfig, msdyn_salesagentprofile, msdyn_dataqualitysuggestion and msdyn_dataqualitysource, at a mix of User (Basic) and Organization (Global) levels. Built-in sales roles already have them.
Configuration is solution-portable: export from sandbox, import into production.
Sellers can dismiss a card as Not relevant, Mark as done, jump to the record form, or thumbs-up/thumbs-down it with free-text feedback. Each card also exposes a per-dimension breakdown of why it was prioritised.

When to use / skip

Turn it on when the client is running two or more sales agents and sellers have started ignoring all of them — the problem it solves is agent sprawl, where each insight surface competes for attention and none of them wins. It's also the only sane route if you're building custom agents and want their output in front of sellers without shipping your own UI. Skip it while the client runs a single agent: you add a consumption line and a ranking model you have to explain, in exchange for a carousel showing what the agent already showed. And skip it if nobody has agreed a budget for consumption, because the cost scales with how many actions your agents generate, not with how many sellers look at them.

Configuration decisions

Which agents get the Connected skills toggle turned on. This is your only real cost control — each enabled prioritised source adds processed actions, and therefore spend.
Whether Data Enrichment stays in the carousel. It's free and it's stacked at the end, but it's also more cards competing for the same attention.
Whether custom agents are worth the API work. The push API and custom source registration are a development task, not configuration.
The security role position: built-in roles are covered, so this is only a decision in orgs running bespoke roles — where it's nine privileges at specific levels and worth scripting rather than clicking.
Which record types you actually want this on. The carousel covers contacts, opportunities, leads and accounts; there's no value in enabling source agents for entities the client doesn't work that way.

Gotchas

The carousel needs Read Only Grid. Sellers on a different view see nothing and conclude the feature doesn't work — this is the first support call you'll get.
A missing carousel is more often an unconfigured source agent than a broken Recommended Actions setup. The agent has nothing to aggregate until a source agent is configured for that entity.
Charging follows processed actions, so a noisy source agent is an expensive one. Cost lands on how much your agents produce, which is not a number most clients can forecast at design time.
Data Enrichment actions being free and unranked surprises people who assume everything in the carousel went through the same prioritisation. They didn't — they're a separate pane.
Only the top card per record shows. An important-but-second-place action is invisible until the first is resolved, which makes the carousel a poor place to look for a complete picture of a record's risks.
The scoring model isn't tunable in the way "custom prioritization models" implies — you influence it through agent-side signals and principle mappings, not by editing weights.
One Microsoft page links onward to "Recommended Actions overview" pointing at sales-close-agent.md. Expect some churn in the naming and linking; the capability has been renamed before.

Consultant notes

Frame this to the client as attention management rather than a new capability, because that's what it is. The question in design workshops should be "which agents earn a slot in front of sellers", not "shall we switch it on" — and the answer has a direct cost consequence.
Get consumption monitoring and an owner in place before go-live. Billing tracks actions processed, so spend grows with agent activity and deal volume, invisibly, without anyone changing configuration. That surprises finance a month in, not on day one.
Pilot with one prioritised source agent and Data Enrichment, then add the second. Turning everything on at once gives you no way to attribute either the consumption or the sellers' verdict on card quality, and the thumbs-down feedback is much more useful when you know which agent produced the card.
Push back on the "one card per record" behaviour early if the client's sales process expects a full risk list — this is a nudge mechanism, not a work queue, and managers who read it as coverage will draw the wrong conclusions in pipeline reviews.
Sequence it after the source agents are trusted. Ranking bad insights more prominently is worse than leaving them in the agent's own interface where sellers already ignore them. See [[feature-sales-qualification-agent]] and [[feature-sales-opportunity-agent]].

Reached GA September 2026 after a short preview and the naming has already shifted once — recheck the supported source agents, the charging model and the page naming after the next release wave.