What it does

An autonomous agent that picks up leads matching criteria you define, researches the lead and its company, scores fit against a target customer profile, and drafts an outreach email. In its fuller mode it also sends that email from its own mailbox, follows up, judges purchase intent against BANT criteria, and hands promising leads to sellers while disqualifying the rest.

Key facts

• Two modes. Research-only does research, target customer profile fit, email drafting and hand-over. Research and engage adds sending outreach, follow-ups, intent detection and BANT evaluation.
• Upgrading Research-only to Research and engage is supported and preserves the research settings. Downgrading is not supported — a genuine one-way door.
• The 2025 release wave 2 plan lists lead research and engagement as public preview from 5 May 2025 and GA from 31 October 2025, with email validation, custom web sources, competitive insights and auto-linked CRM data reaching GA the same day. Custom research was listed as public preview from 1 November 2025.
• Prerequisites for both modes: an Azure app registration and a Dataverse application user holding the AISalesPerson security role. Research and engage additionally needs a shared mailbox and server-side synchronisation configured against it.
• On that shared mailbox, set the Appointments, Contacts and Tasks synchronisation method to None. The agent only needs mail sync.
• Configuration is a gated wizard: prerequisites must be ticked as done before the next section loads, and Start agent stays disabled until a simulation runs successfully against real leads meeting your criteria.
• Selection criteria include a look-back period. On first start the agent processes every qualifying lead created inside that window; afterwards it works in near real time.
• Re-run behaviour is specific: a full re-run happens when a lead's email goes from empty or invalid to valid, when the account name goes from empty to valid, or when the lead newly meets the selection criteria. A change from one email address to another triggers only email validation. Related-table criteria are checked hourly, and only for leads inside the look-back period.
• The agent uses Copilot Studio capacity — prepaid message packs or pay-as-you-go — for research, target customer profile matching and initial outreach generation. Viewing insights already generated costs nothing.
• Language is set on the agent, not per user. Research page headings appear in the seller's language while insights and outreach emails come out in the agent's language.
• Research and engage adds an AI disclaimer and Other terms field alongside the email signature, for telling recipients they're corresponding with an agent. Get legal to write it.
• Once configured, the agent can't be deleted without Microsoft support. Stopping it prevents new leads being taken on; leads already in flight continue to hand-over or escalation.
• Where the agent is enabled, it owns the lead summary banner outright and any configured lead summary fields are ignored.

When to use / skip

This is for organisations drowning in inbound — web forms, events, webinars — where leads sit untouched because there aren't enough SDRs. That's the case it was built for. Research-only is a reasonable first step almost anywhere: it does the tedious company research and drafts the email, and a human still decides.

Research and engage is a different proposition and most projects shouldn't start there. You're authorising software to email prospects in the client's name, unsupervised. If the client sells into a small named-account universe, or has a compliance function with views about outbound contact, don't. For high-volume, low-touch inbound with legal on board, it's the most valuable agent in the Sales product today.

Configuration decisions

• Research-only or Research and engage — and be clear it's easy to go up and impossible to come back down.
• The selection criteria and look-back period, which together decide the initial volume of leads and the ongoing capacity burn.
• The target customer profile, plus company and product information. Vague inputs here produce vague research and generic emails; this is where the quality actually comes from.
• BANT hand-over criteria and the assignment rules that route handed-over leads to sellers or teams.
• Knowledge sources for company insights, competitor insights and email context — public web, SharePoint, OneDrive, uploaded documents — and whether internal documents are in scope at all.
• The agent language, signature, AI disclaimer and other terms, signed off by legal before the agent sends anything.
• Prepaid capacity versus pay-as-you-go, and who watches the consumption.

Gotchas

• The agent pauses silently in three situations: an admin revoked a consent it needs, capacity ran out, or the org is on an older agent version. Nobody gets a broadcast; you find out because leads stopped moving.
• Sellers not seeing handed-over leads is nearly always one of two things — missing permissions on the agent's tables, or the agent-specific views never added to their app.
• Deleting a Dataverse field used in a custom selection or hand-over criterion breaks summary generation and hand-off entirely. The agent doesn't degrade gracefully; you have to remove the criterion.
• Custom security roles need a precise permission set, including user-level access to the Sales Copilot Insight and Sales Copilot Email Insight tables and global-level read on Sales Agent Handover. The out-of-the-box Salesperson and Sales Manager roles have it, so the problem only shows up in orgs with bespoke roles.
• Upgrading the agent version can make previously generated research pages inaccessible; the data itself lives in the lead agent result and account research result tables.
• AI agent user accounts appear in the environment whether or not you enable the feature. Explain that before a security reviewer finds it.
• Leads with no account information are skipped — nothing to research — and surface as an unhelpful "error in a supporting process" message.

Consultant notes

• Insist on the simulation as a real gate, not a formality. Read the generated emails with the client's marketing lead in the room before anyone presses Start agent.
• Start in Research-only, run it for a few weeks, then decide on engagement with evidence. It's also the only sequence that lets the client back out.
• Get the AI disclaimer drafted by the client's legal team early. It blocks go-live more often than any technical prerequisite.
• Set up capacity monitoring on day one and agree who acts when it runs low, because depletion looks exactly like the agent being broken.
• Tell the client plainly that the agent can't be deleted afterwards without contacting Microsoft support. That single fact changes how carefully people configure it.

---

Worth another look after the next release wave — the agent has changed version and behaviour repeatedly, and the upgrade path is the part that bites.