An AI agent for preparing meeting briefs

What an AI agent for preparing meeting briefs does

Start with the output, because it is easy to picture: one page, waiting for you a few minutes before the call, that answers “who am I talking to, why, and what do I need to decide.” The agent does not think for you. It does the gathering you keep doing by hand—opening the invite, hunting for the last thread, checking what you promised last time—and lays it out in the same shape every time.

A useful brief is short and predictable. These six blocks cover almost every meeting:

• Who and where—attendee names, roles, company, and the meeting’s stated purpose from the invite.
• Last real contact—the most recent genuine interaction, with a link, not a guess.
• Open commitments—what you or they said you would do, and whether it happened.
• One decision—the single thing this meeting exists to settle.
• Questions to ask—two or three, drawn from the open threads.
• Flags—anything unusual: a stalled deal, an unanswered message, a renewal date approaching.

Why it matters: a fixed template is what makes the brief trustworthy. You learn where to look, so a missing line stands out instead of hiding. How to judge it: you should be able to read the page in about two minutes and walk in ready. The failure case is the brief that invents a “recent conversation” to fill the space—confident, plausible, and wrong. The trade-off: a rigid template occasionally misses the odd curveball meeting, and you accept that small rigidity in exchange for a page you can trust at a glance.

Name the sources, keep the CRM read-only

What to do: list the exact places the agent may read, and no others—your calendar, one notes document or folder, and your CRM connected read-only. “Named sources” is the whole game. “Pull whatever seems relevant” invites the agent to wander, and a brief built from the open web is a brief you cannot check.

Why it matters: bounded inputs are testable and safe. Read-only access to the CRM means the agent can see the deal history but cannot change a stage, overwrite a note, or touch a field—so the worst a mistake can do is show you the wrong line, never rewrite your records. This is the least-privilege habit: grant the narrowest access that still does the job. Our CRM and agent checklist walks through connecting read-only first and earning each write later.

How to judge it: you can name every source in one breath. The failure case: a vague scope pulls a stale deal or a private internal note into the wrong client’s brief. The trade-off: naming sources means maintaining the list—when you adopt a new tool, the agent will not see it until you add it. That small chore is the price of a brief that never surprises you.

Keep the brief draft-only

What to do: treat the brief as an input to your judgment, not an action the agent takes. It appears in your dashboard for you to read. The agent does not email it to attendees, does not post it to a shared channel, and does not write anything back to the CRM.

Why it matters: the whole value of a brief is that it prepares you. The moment it auto-shares, a wrong figure or an internal note can reach the client before you have read it yourself. Keeping a person in control of anything that leaves the building is the boundary that makes automation safe as the stakes rise. How to judge it: when a brief generates, nothing has changed outside your dashboard—no message sent, no record edited.

“Brief ready for your 10:00 with Devi Menon (Northwind). One decision flagged: renewal terms. Draft only—nothing sent.”

The failure case: a tempting “auto-share to attendees” shortcut leaks a candid internal note—“founder seemed frustrated last call”—straight to the client. The trade-off: draft-only means you still do the last mile of reading and deciding. That is not overhead; it is the part only you can do.

Make every line traceable

What to do: give every fact in the brief a source—a link to the email, the note, or the CRM record it came from. Anything the agent cannot find, it marks “not found,” rather than filling the gap with a confident guess.

Why it matters: an unverifiable brief is worse than no brief, because it reads as authoritative while quietly being wrong. When a line links back to its source, you can settle any doubt in one click, and you quickly learn which parts to trust. A brief that also signals how sure it is—“last contact: 3 weeks ago (from Notes)” versus “last contact: not found”—lets you calibrate instead of assume. How to judge it: pick any line at random and try to reach its source; if you cannot, the brief is not done.

The failure case: the agent states last quarter’s revenue figure that is actually two quarters stale, with no date attached, and you repeat it in the meeting. The trade-off: source tags add a little visual clutter to a page you wanted clean. Keep them small, but keep them—the clutter is what lets you rely on the page.

Schedule it, then check whether it earns its place

What to do: a brief that arrives after you have walked in is useless, so schedule it to land a little before you need it—about thirty minutes before each meeting, or as one morning batch covering the day’s calls. Setting this up is a plain-language choice of task, timing, and timezone; our guide to scheduling recurring agent work covers it without any cron syntax.

Why it matters: timing is half the value. How to judge it: for the first two weeks, check three things on every brief—did it arrive on time, was it accurate, and did you actually use it. Those questions turn a vague “this is handy” into evidence you can act on. The failure case: a fixed 6 a.m. morning batch misses the client call added to your calendar at 10 a.m., and you walk in cold. If same-day additions are common, trigger a brief when a meeting is created instead of relying on one fixed batch. The trade-off: a morning batch is simple but blind to late changes; per-meeting timing catches everything but runs more often. Pick the one that matches how your calendar actually behaves.