Skip to article

What Roadmap Brief Can an AI Agent Build From Discovery Calls?

Build an AI agent in Gumloop that reads your discovery call notes and briefs the roadmap discussion, counting the calls behind each theme and listing them.

By Jacqueline V. TwillieNovember 9, 20266 min readGuide
Quick answer

The roadmap brief comes from an AI agent in Gumloop that reads your notes from customer discovery calls. It quotes customers only where your notes do, counts the calls behind each theme and lists them, sets contradictions side by side, and names the planned topics nobody mentioned.

No theme goes into the roadmap brief without the calls behind it. If a theme came up in some of your calls, the brief gives that count and lists those calls by label.

The brief has four parts: what customers asked for, in their own words; how many calls raised each theme; what contradicts what; and what nobody mentioned. The AI agent builds all four from the notes you paste. It's told to quote only what your notes quote, never to give a percentage, and to leave names out unless you choose to include them.

What do you paste into the instructions for the discovery brief?

The planned topics come from your call guide, and they're the only way the brief can say what nobody mentioned. The number of calls is the total every count is measured against.

Copy and fill in
You are my discovery call analyst. I'm the [MY ROLE] at [COMPANY NAME], and I'm preparing our roadmap discussion.
I'll paste my notes from [NUMBER OF CALLS] customer discovery calls, each under a label like Call 1. Refer to calls only by those labels. Use customer names in the brief: [YES OR NO]. If the answer is no, never repeat a name from the notes.
Text counts as the customer's own words only when my notes put it inside quotation marks. Quote those words exactly, with the call label after them. Report everything else as "per my notes".
The topics we planned to cover are: [PLANNED TOPICS]. Our current assumptions are: [OUR ASSUMPTIONS].
Write the brief in four parts.
1. What customers asked for. Group the requests into themes and give each theme's quotes.
2. How many calls raised each theme. Give the count out of the total and list those calls by label.
3. What contradicts what. List any two calls that disagree and any call that disagrees with an assumption above. Show both sides with their labels and don't settle it.
4. What nobody mentioned. List each planned topic that none of the notes touch.
Count a call once per theme, however often it came up. Never estimate, round or give a percentage, and say so when the notes are too thin to support a line. If the labels you can see don't match the number of calls I gave, tell me before anything else.

How do you build the discovery brief AI agent in Gumloop?

  1. 01

    Number the calls and put the customers' own words in quotation marks.

    Do this in your notes, before you open Gumloop. Call each one Call 1, Call 2 and so on, and take names out of the headings. Keep the key, who was Call 7, in a separate file the AI agent doesn't see. Put the customer's words inside quotation marks wherever you wrote them down exactly. Anything outside them is reported as your summary. Write down your planned topics and your team's assumptions too. If your company restricts customer data, read Can I Use an AI Agent at Work When My Company Restricts the Data I Can Connect? first.

  2. 02

    Create the discovery brief AI agent, with no connectors and the web abilities off.

    Open Gumloop, create a new AI agent and describe the job in a sentence. Add no connectors. The notes go into the task itself. In the Abilities section on the tab named Agent, switch off Web Search, Web Fetch and Search Past Tasks, and turn the browser off in the Browser section if it's on. Leave Ask Question on, so the AI agent can pause and ask you when the call count doesn't match. Checked on 2026-11-11 against Gumloop's AI agent builder page, abilities table, and Gumloop's AI agent browser page.

  3. 03

    Save the block, and let your corrections stick.

    On the tab named Agent, paste the block, fill the brackets and press Save (Ctrl/Cmd + S). Edit Own Instructions sits under the instructions box and is on by default. For this job it stays on. You'll run it by hand with your notes beside you, and a correction like "merge cost and packaging into one theme" is one you'll want kept. Before a new set of calls, read the instructions and cut anything that belonged to the last set. Copy your filled instructions into a document first. Checked on 2026-11-09 against Gumloop's AI agent builder page, self-improving instructions section.

  4. 04

    Paste every call into one task and watch the context meter.

    Open a new task and paste the notes for every call under their labels. The meter at the bottom right of the message box shows how much of the AI agent's context is in use. When it fills, Gumloop's page lists auto summarization, which compresses older messages, as one way out. Summarizing starts as a task nears its context budget, which can happen before the meter is full. A count read from compressed notes is one to check by hand: on a long set, start with the calls you pasted first. Skip the Triggers tab. You'll run this when the calls are in. Checked on 2026-11-11 against Gumloop's AI agent builder page, message input, context usage and task summarization sections.

  5. 05

    Open the notes behind three of its counts.

    Pick the three themes you remember best. For each, open the calls it lists and find the supporting line. Then search the notes of one call it didn't list. If the theme's there, the count is short. Search your notes for each quotation, word for word, and for every planned topic it says nobody mentioned.

I am an ambassador for Gumloop. Code VENTURENYC takes 20% off your first month.

What should the first three briefs show you about the counts?

A count that looks low is often one theme that's split across two sets of wording. Ask for them merged, then read the instructions afterward, since the merge may have been written into them.

An empty contradictions part means the calls agree or your assumptions line was thin. Add a real assumption and run it again, and you'll see whether any call contradicts it. If every contradiction sides with your assumptions, read How Do I Stop an AI Agent From Telling Me What I Want to Hear?

After the meeting, note which theme the room argued about and whether its call list held up.

Frequently asked questions

What if my notes are summaries and not exact words?

Then the brief carries no quotes and reports everything as "per my notes". Put quotation marks around lines you wrote down word for word, starting with your next call.

Can I leave customer names in?

Yes. Set the names blank to yes and they'll appear as your notes wrote them. At no, the brief uses call labels only.

What does the brief do with a request only one call made?

The brief lists it as a theme with a count of one and names that call, and leaves ranking the themes to your meeting.

More on Data for Corporate Executives