Sales Opportunity Agent

What it does

An agent that researches a defined segment of open opportunities on a schedule and writes back an importance rating, a risk rating, a written deal summary and up to three recommended actions. Sellers read it from a My top opportunities from AI agent view and a Summary banner on the opportunity form. Both the risk and importance models are configurable, including criteria you invent yourself.

Key facts

You point each agent instance at a product line and an opportunity segment — for example Hot opportunities with estimated revenue over a threshold. Only opportunities matching the selection criteria get researched, and only high and medium importance ones appear in the seller view.
Up to 10 active agent instances per organisation, each with its own profile, selection criteria and knowledge-source filters. They all draw from the same capacity pool and share one Copilot Studio knowledge base.
Research refreshes every three days by default; the admin can set High (daily) or Low (every seven days). Edits to Account, Contact, Est. close date, Actual close date or Est. revenue trigger an immediate re-research.
Eleven built-in risk criteria — no recent engagement, unresponsive champion, not multithreaded, missing stakeholder roles, close date deviation, stalled in stage, objections raised, missing BANT info, aging opportunity, competitor mention, negative news. Thresholds, roles and per-stage values are all editable, and each can be switched out of the assessment.
Six built-in importance criteria — existing customer, high account win history, Top 500 account, strategic references, large deal size, executive visibility. Same deal: editable thresholds, toggle per criterion.
Custom criteria on both sides are defined by name, description and mapped opportunity (or related-table) fields, tuned with natural-language instructions.
It depends on predictive opportunity scoring. If the model isn't configured in the environment, starting the agent configures it for you.
Email insight gathering runs one of two ways: Microsoft 365 Services consent, where the agent reads the opportunity owner's mailbox directly every six hours without storing it — and that consent can't be revoked — or server-side synchronisation, where it only sees mail already synced into Dataverse.
Email insights need a real external email on the opportunity whose sender address maps to an associated contact. No mapped contact, no insight.
Insights are generated in the agent's configured language; section headings render in the user's language. A German-language agent with English users produces exactly that mix.
Runs on Copilot Studio capacity under consumption-based billing, and needs the tenant-level Power Platform prerequisites done once: Copilot Studio capacity, cross-region data movement accepted, AI prompts on, AI insight cards on, AI Agents on.
Data policies must allow the Copilot Studio, Direct Line and knowledge-source connectors, plus SharePoint/OneDrive or document connectors if internal documents are in scope.
Sellers need read on Connector at user level to give the Microsoft 365 consent. Salesperson and Sales Manager already have it; custom roles won't.
Requires the modern look enabled on the Sales Hub app.
Configured agents export as a solution, so sandbox-to-production promotion works the normal way.

When to use / skip

This earns its keep where sellers carry more open opportunities than they can properly track and deals go quiet without anyone noticing. The risk criteria are the real product — stalled in stage, no recent engagement and missing BANT are the things that actually kill pipeline, and having them evaluated nightly against every deal in a segment is genuinely more than a seller does by hand.

Skip it where the opportunity data is thin. The agent reasons over estimated revenue, close dates, stage, stakeholder connections and email traffic. If your client's sellers treat the opportunity as a placeholder and do the real work in Outlook and spreadsheets, the agent will produce confident nonsense and sellers will stop reading it within a fortnight. Fix the CRM discipline first. Also skip if nobody will own the criteria configuration — the defaults are generic and a 30-day stalled-in-stage threshold is wrong for most sales cycles.

Configuration decisions

Which product line and opportunity segment the first instance covers, and how small to keep the pilot segment while you tune credit consumption.
Whether to run one agent or several — enterprise versus mid-market, different product lines — remembering selection criteria must be distinct per instance.
Microsoft 365 Services consent versus server-side synchronisation. This is a data-protection decision as much as a technical one, and the consent route is one-way.
Refresh frequency against capacity: daily on a large segment is a very different bill from every seven days.
Which risk criteria stay on, and what the thresholds and stage-specific values are. Stalled-in-stage and close-date deviation need the client's real cycle times, not the defaults.
Which stakeholder roles map to your organisation's actual role names, rather than leaving the built-in Champion and executive assumptions in place.
Objection keywords and competitor names, which are only as good as the list someone maintains.
Which importance criteria reflect how the business genuinely prioritises, and whether custom criteria are needed for strategic accounts or key product lines.
Agent language, if the client is multinational.
Which knowledge sources are in scope, and whether public web data is acceptable at all.

Gotchas

The Microsoft 365 email consent cannot be revoked. Get that in front of whoever owns data protection before a seller clicks the banner, not after.
The agent only researches opportunities matching selection criteria, so "the insights aren't showing" is almost always a segment mismatch rather than a fault. Check the criteria first.
Sellers see only high and medium importance opportunities in the agent view. Low-importance deals are researched-but-invisible, which reads as missing data.
Connected skills is still marked preview, and the Recommended Actions section needs the separate Recommended Actions Agent enabled — so a demo can show no recommended actions at all and look broken.
The prerequisites page still labels this agent as preview in its related links while the agent's own articles no longer carry a preview banner. Treat the specific capability you're promising as the thing to verify, not the agent as a whole.
Predictive opportunity scoring gets configured silently when you start the agent. If the client has opinions about scoring models, they'll want to know that happened.
Ten instances sharing one capacity pool means one badly scoped instance on a daily refresh can starve the others.
Email insights depend on sender-address-to-contact mapping. Sellers who email from a shared mailbox, or contacts stored without their real address, produce nothing and no error.
Changing Est. revenue or close date triggers an immediate re-research, so a seller doing routine housekeeping on a hundred records can generate a burst of consumption.

Consultant notes

Scope the criteria configuration as a workshop, not a setting. The defaults produce plausible-looking risk flags that don't match the client's sales process, and the first thing a sales director says in UAT is "that deal isn't at risk". Bring their real cycle times and stage definitions and set the thresholds in the room with them.
Raise the irrevocable email consent early and in writing. It's the one decision on this feature that can't be walked back, and it tends to reach the data-protection lead late, after sellers have already been told to click the banner.
Pilot on a segment small enough that the Copilot Studio bill is boring. Consumption is driven by segment size times refresh frequency, and the temptation to turn it on for all opportunities at daily refresh is exactly how clients get a surprise invoice in month two.
The agent is a mirror on CRM data quality, and that surfaces awkwardly post-go-live. Where stage values are inconsistent or stakeholder connections aren't maintained, the agent's output is visibly wrong and sellers blame the tool. Audit the data before enabling, and use the first month's risk flags as a data-quality report rather than a seller performance one.
Sequencing matters: the tenant-level Power Platform prerequisites, capacity purchase and data-policy connector allowances all sit outside the Sales project and involve people who aren't in your standups. Start them the week you agree the feature, not the week you configure it.

Worth revisiting once connected skills leaves preview and the criteria set settles — both sides of the model are new and Microsoft is still adding to them.