Getting Started

How it fits into your agents

Second Opinion for Sales is not a standalone app. It installs as an MCP tool on an existing Copilot Studio agent. From there, two patterns work:

• Tool on your agent. Add it directly to a sales agent your team already uses. That agent gains the behavioral read as one of its tools.
• Child agent for a Master Sales Agent. Wrap it in a small dedicated agent, and connect that agent as a child of your master sales agent, the one with access to Salesforce, Dataverse, or your CRM. The master agent pulls the conversation from your CRM, sends one speaker's words to the child, and weaves the read into its own deal guidance. Copy-paste description and instructions for the child agent are provided below.

Before you start

• An active Second Opinion for Sales subscription from Microsoft AppSource, with a license seat assigned to each user who will connect.
• A work or school Microsoft 365 account. Personal Microsoft accounts are not supported.
• An existing agent in Copilot Studio to install the tool on, or make a new one for the child-agent pattern.

Administrator step, one time per organization

Before anyone in your organization connects, an administrator makes Second Opinion for Sales known to your Microsoft 365 tenant. Until this is done, sign-in fails with error AADSTS650052.

az ad sp create --id 86171dc8-e8f7-499f-ad82-595d75162961

https://login.microsoftonline.com/common/adminconsent?client_id=86171dc8-e8f7-499f-ad82-595d75162961

Connect in Copilot Studio

1. 1
   Copy the MCP endpoint.

   MCP endpoint

   https://mcp.pansophyai.com/mcp

   Copy
2. 2
   Open your agent at copilotstudio.microsoft.com, the one you want to add Second Opinion for Sales to.
3. 3
   Add the MCP tool. Click Add a tool, then New tool, and choose Model Context Protocol. Paste the endpoint from step 1 as the server URL.
4. 4
   Set authentication. This is the step that makes or breaks the connection.

   For authentication, choose OAuth 2.0, then Dynamic discovery. Nothing else.

   Do not choose "No authentication." The tool may appear to save, but every call will be rejected, and the half-made connection has to be deleted before you can retry cleanly. If you already created one with No authentication, delete that tool and start again from step 3.

   Leave the other fields alone; they fill in from the server. Click Create.

   The tool is created without asking you for a client ID, client secret, or Authorization, Token, and Scope URLs. If it asks for those, authentication is not set to Dynamic discovery; go back and fix it before continuing.
5. 5
   Add to your agent and sign in once. Add the tool to your agent and choose Create new connection. Sign in with the Microsoft 365 account that holds a license seat, and approve the consent prompt.

   If sign-in fails with AADSTS650052, the administrator step above has not been completed in your tenant. If it fails on the license, confirm the account has a seat in the Microsoft 365 admin center under Billing, Licenses, Second Opinion for Sales.
6. 6
   Connect for the first time in the test chat. Open the test chat and send any message that uses the tool. The first time, the chat will not run the tool; it tells you a connection is needed. Click Open connection manager, find Second Opinion for Sales in the list, click Connect, sign in, and allow. Then return to the test chat and try again.

   The connection shows as Connected in the connection manager, and the test chat stops asking.

   Every user gets this one-time connect prompt the first time they use the agent. It is normal; they connect once with their own Microsoft 365 account and it holds from then on.
7. 7
   Run it from a chat, then publish. Test from an actual chat with your agent, not the tool setup page. Paste a block of sales conversation and ask for a second opinion on the seller or the prospect. Then save and publish; each user signs in with their own Microsoft 365 account.

   If the setup page says "No tools available," that is cosmetic. The tool runs from a real agent chat.

Set it up as a child agent

To make the read available to a master sales agent, wrap the tool in a small dedicated agent and connect it as a child. The description and instructions below are ready to paste; they carry the guardrails with them, so the read stays safe no matter which parent agent calls it.

1. 1
   Create the wrapper agent. In Copilot Studio, create a new agent named Second Opinion for Sales Agent in the same environment as your master sales agent. Add the MCP tool to it using the connect steps above.
2. 2
   Paste the description.

   Agent descriptionCopy

   Behavioral sales analyst powered by Second Opinion for Sales. Give it one person's words from a sales conversation (call transcript, email thread, discovery notes), at least 400 characters, and say whether that person is the seller or the prospect. It returns a behavioral read: a one-line key finding, a fuller second opinion, two to three concrete next steps, and a read confidence. Use it before drafting outreach, prepping a call, or deciding the next move in a live deal. It reads patterns in language; it does not predict whether a deal will close.

3. 3
   Paste the instructions.

   Agent instructionsCopy

   You are a sales coach built on Second Opinion for Sales. Your one capability is the second_opinion tool, which analyzes one person's words from a sales conversation and returns a behavioral read.
   
   Input rules
   - The tool takes the text of ONE speaker plus their role: "seller" or "prospect". Role sets the interpretation lens. If you cannot tell which it is, ask; never guess silently.
   - You need at least 400 characters of that person's own words (up to 50,000). If you have less, ask for more material instead of calling the tool. Longer samples produce a more reliable read.
   - For a multi-speaker transcript, split by speaker and call once per speaker.
   - Do not scrub names or personal details before sending. That is handled server-side.
   
   Delivering the read
   - Never read the result back like a form, and never unload everything at once. Lead with the key finding, the single highest-leverage move, the hard call if the profile is unfavorable, and an offer to draft the next message. Hold the fuller read and reasoning in reserve until asked.
   - Deliver it as your own advice, in the user's voice and register, woven together with what you already know about their deal.
   
   Two limits that always hold
   1. Never claim the read predicts whether the deal will close, or that it is an accurate, validated, or scientific measure of the person. Confidence in the read is fine; claims of predictive accuracy are not.
   2. Trait and personality language is for coaching the user only. It never appears in any message drafted for the prospect. Translate the insight into behavior and benefit instead: not "he reads as high in neuroticism," but a line that reassures him he will not be left exposed after signing.
   
   The read is words, never numbers. Do not invent scores, percentages, or rankings that the tool did not return.
   
   When another agent calls you
   - Return the read in a structured, labeled form the caller can use: key finding, recommended move, next steps, read confidence.
   - Pass the two limits along as explicit handling notes, so the calling agent does not put trait language in front of a prospect or present the read as a deal prediction.
   - If the caller sends mixed-speaker text or no role, ask it to split and label rather than proceeding on a guess.
   
   Scope
   - You analyze sales conversations. Decline requests to profile people outside a sales context, to act as a lie detector, or to produce psychological assessments. Offer the sales read instead.

4. 4
   Allow it to be called. In the wrapper agent's Settings, turn ON Let other agents connect to and use this one. Publish the agent.
5. 5
   Connect it to your master sales agent. In the master agent, add the wrapper as a connected agent. We recommend unchecking "Pass conversation history to this agent". This does not cut off the data: the master agent always sends an explicit task when it calls the child, and the instruction snippet in the next step makes it put the speaker's text and role inside that task. What unchecking prevents is the child receiving your entire conversation, both speakers, your coaching exchanges, and CRM context along for the ride. The tool should read one person's words, nothing more.

   On very long transcripts, if the child comes back saying it received too little text, re-send with the speaker's text pasted directly into your request to the master agent, or re-check the box for that connection.
6. 6
   Add this to your master agent's instructions. It tells the master how to call the child correctly and keeps the guardrails intact on its side of the connection. The child agent will ask for a split and a role if it receives mixed text, but handling it here saves that round trip.

   Master agent instruction snippetCopy

   When you need a behavioral read on someone in a deal, call the Second Opinion for Sales Agent.
   
   Sending a request
   - Send the words of ONE speaker only, at least 400 of their own characters, and state their role: "seller" or "prospect". Pull their words from the transcript, email thread, or CRM record and leave the other speakers out. Do not send the whole conversation.
   - Do not rely on conversation history to carry the text. Include the speaker's text and role explicitly in each request.
   
   Using the read
   - The read comes back as a key finding, a recommended move, next steps, and a read confidence. Weave it into your own guidance; do not paste it back as a form.
   - Never present the read as a prediction that the deal will close, or as a validated measure of the person.
   - Trait and personality language is for coaching the user only. It never appears in any message drafted for the prospect; translate it into behavior and benefit instead.
   - Do not write the read or any trait language into memory, CRM fields, or any other persistent store. The read is for the conversation at hand.

Troubleshooting sign-in