Sales Research Agent

What it does

A research canvas in Sales Hub where a sales leader asks an open business question and gets back a "blueprint" — a written analysis with charts, the reasoning behind it, and follow-on questions you can keep pulling on. It runs over Dynamics 365 Sales data by default and can be pointed at other Dataverse environments, a Fabric lakehouse, or files you upload.

Key facts

Reached from the Sales Hub site map under My Work > Sales Research Agent. It's a full-page canvas, not the Copilot side pane.
The 2025 release wave 2 plan lists the Sales Research Agent as generally available in March 2026, with sales operations insights and Fabric lakehouse integration listed as public preview February 2026 and GA April 2026. The responsible AI FAQ page is still titled as preview, so check the specific sub-capability you're promising.
Access is by security role. Only System Administrator and System Customizer have it by default; everyone else needs the Sales Research Agent Reader role, assigned directly or through a group team.
Supported data sources: Dynamics 365 Sales data (connected by default), other Dataverse environments, Microsoft Fabric lakehouse, and uploaded CSV, Excel and PDF files. Uploaded data can augment CRM data or be used on its own.
Business functions set the context for a question. Pipeline Exploration is selected by default, and the same data with a different business function produces a different blueprint.
Show work exposes the research plan, data sources, analysis steps and visualisations behind a blueprint. That transparency is the feature's main defence in a room full of sceptical sales leaders.
The AI cursor (Alt + C) lets you interrogate or restyle a chart in place — change the chart type, group differently, explain an acronym in the source data. Asking a new question from the cursor generates a new blueprint rather than editing the existing block.
Work is organised into named workspaces with journey lines recording the questions asked along the way, so a piece of analysis can be picked up again later.
It runs on Copilot Studio capacity and is explicitly subject to usage limits and throttling. Without credits, blueprints won't generate at all.
Bing search is a tenant-level consent set in the Power Platform admin center. Off, the agent works only from internal and uploaded data.
Feedback collection is configurable at environment level in the Power Platform admin center: off, thumbs up/down only, or detailed feedback with comments.
The canvas needs third-party cookies allowed in the browser. If the page won't load, that's the first thing to check.

When to use / skip

This is a tool for sales leadership and sales operations, not for sellers. The right user is the person who currently asks an analyst for a pipeline study and waits a fortnight. If your client has that person and that backlog, it's a strong demo and a genuinely useful capability.

Most projects don't need it on day one, and it is not a replacement for dashboards or Power BI. Dashboards answer the same question every week; this answers a one-off question well and then you throw the answer away. Where a client already has a mature Power BI estate and a sales ops team who know their numbers, the honest advice is that this competes with something that already works, and the case is weaker. Where reporting is thin and questions go unanswered, it's the fastest route to insight the platform offers.

Configuration decisions

Who gets the Sales Research Agent Reader role, and whether that's assigned via group teams rather than user by user.
Whether Bing search consent is granted, which decides if external context can enter the analysis at all.
Which data sources are in scope — CRM only, or another Dataverse environment, or a Fabric lakehouse where the warehouse already holds the numbers people trust.
Whether file upload is acceptable, and the client's position on staff uploading spreadsheets and PDFs into an AI canvas.
Which business functions and business context to define beyond Pipeline Exploration, so blueprints come out framed for the client's industry and role.
The feedback setting, particularly whether free-text comments are allowed to leave the environment.
Prepaid capacity versus pay-as-you-go, and what happens to users when throttling kicks in.

Gotchas

Nobody but administrators can see it until you assign the reader role. Expect "the menu item isn't there" as the first support ticket.
The blank page caused by blocked third-party cookies looks identical to a permissions problem or an outage. Check the browser before you raise anything with Microsoft.
Capacity throttling degrades the experience without an obvious cause — blueprints simply don't generate. Monitor Copilot Studio consumption before rollout, not after.
Blueprints are only as good as the data model behind them. Poorly named custom fields and inconsistent stage values produce confident, wrong analysis. The Show work panel is how you catch it, so teach people to open it.
Asking a follow-up through the AI cursor creates a new blueprint rather than updating the current one. Users expect an edit and get a fork.
Uploaded files are a data governance question in disguise. Get the client's position in writing before a sales ops analyst uploads next year's targets.
The GA and preview picture varies by sub-capability. Treat "the agent is GA" and "operations insights are GA" as separate claims.

Consultant notes

Demo it to the sales leadership team, not to sellers. Bring one real question from their world and let the blueprint answer it live — that lands far harder than a scripted walkthrough.
Open Show work in the demo. Sceptical leaders trust the tool once they can see which records it used.
Push back on any suggestion that this replaces Power BI or the forecast. It answers ad hoc questions; it isn't the reporting layer.
Do the business context and business function setup before the client's first serious session. Out of the box it answers as a generic pipeline analyst, and clients read that as the ceiling.
Confirm capacity and the reader role assignments before go-live, and tell whoever owns the environment that heavy use of this agent shows up on the Copilot Studio bill.

Worth another look once the operations insights and Fabric lakehouse capabilities have settled, since those change what the agent can actually reason over.