Nobody schedules meeting prep. It happens in the four minutes between the previous call ending and this one starting, and it consists of the same three lookups every time: who is this, what does their company do, and what did we say last time. Done well it changes how a meeting goes. Done in four minutes it usually means opening a website homepage and skimming a headline. The gap between those two is a real cost, and it is almost entirely mechanical.
What Meeting Prep Actually Costs
Put a number on it in your own diary rather than trusting a benchmark. Count your external meetings in a week, multiply by the fifteen minutes proper prep would take, and compare that to the four minutes it actually gets. Most people find two things at once: the total is larger than expected, and the real problem is not the total but the distribution. Prep collapses precisely when the day is busy, which is the same day you have the most meetings.
The deeper cost is not time, it is the meetings where thin prep changed the outcome. Asking a question the person answered publicly last month. Not knowing their company just shipped the thing you are about to propose building. Repeating a pitch that a colleague already gave them in March. None of these are catastrophes and all of them are avoidable by fifteen minutes of reading that nobody has.
What makes this a good agent job is that the work is structured, repeatable and judgment-light. You are not asking anything to decide what the meeting is for. You are asking it to assemble what you would have looked up.
What Belongs in a Pre-Call Brief
The temptation is to include everything findable. Resist it. A brief that runs to three pages does not get read in the two minutes before a call, which means it may as well not exist.
| Section | What goes in it | Length |
|---|---|---|
| Who is in the room | Name, role, company, and one line on what that role usually cares about | A line each |
| What they do | The company in one honest sentence. Not the tagline, what it actually sells and to whom | One sentence |
| What changed | Recent and relevant only: a launch, a funding round, a leadership change, a public announcement. If nothing relevant changed, say so | Two or three bullets, or none |
| Last time | What was discussed, what was promised, what is outstanding. From your own notes and email, not the internet | Two or three lines |
| Worth asking | Two or three specific questions this brief makes possible. Generic questions are a sign the research found nothing | Three bullets maximum |
The "last time" section is the one that most justifies the whole exercise, and it is also the one a generic research tool cannot do, because it needs your records rather than the public web. It is the difference between a brief that sounds like a search result and one that sounds like a colleague caught you up in the corridor.
The "what changed" row deserves a specific instruction: permission to return nothing. An agent told to find recent news will find recent news, and if there is none it will pad with something tangential and six months old. "If nothing relevant has changed, write that nothing relevant has changed" is a one-line addition that measurably improves the output.
The Calendar Invite Is the Trigger
Everything the agent needs to start is already in the invite: the attendees, their email domains, the title, and often a description or an agenda. That makes the design simple.
Run the job the evening before rather than in the morning. Two reasons. The obvious one is that a brief waiting when you open your laptop is a brief you read, whereas one arriving at 8:55 for a nine o'clock call is not. The less obvious one is that tomorrow's calendar is mostly settled by the evening, while this morning's may still be moving.
Filtering matters more than it sounds. A rule as simple as "at least one attendee outside our email domain" removes internal standups, one-to-ones and recurring team meetings, which is most of the volume and none of the value. Add an explicit override, a keyword in the invite title, for the internal meetings that genuinely do need research, and the filter is done.
For the scheduling mechanics, writing a prompt for a recurring agent covers phrasing a nightly job so it stays useful over months rather than drifting. On the research side, building a research agent step by step goes deeper on gathering and citing sources, and what an AI research assistant does is the broader version of the same capability.
What to Hand the Agent
| Element | What to specify |
|---|---|
| Trigger | Each evening, read tomorrow's calendar |
| Filter | External attendees only, plus anything flagged. Skip recurring internal meetings entirely |
| Public research | Company and role level. What the company does, what changed recently and is relevant, what the attendee's role is |
| Internal research | Prior notes, emails and records about this company or person. This is the highest-value input |
| Sourcing rule | Every factual claim carries a link. Anything unsourced is marked as unverified |
| Blanks | Explicit permission to write "nothing relevant found" rather than padding |
| Format | One page, five sections, in that order. A hard length limit is a feature |
| Boundary | Professional and public context only. No personal profiling of attendees |
The sourcing rule is not bureaucratic. The whole risk of an automated brief is that you repeat something in a meeting that was never true, and the only practical defence is being able to check a claim in five seconds before you say it out loud. A brief where every fact is a link is a brief you can trust selectively, which is the right amount.
What We Saw Running This
We ran a nightly version against our own calendar for a few weeks. Three things stood out, and only one of them was the thing we expected.
The expected finding: the briefs were most useful for first meetings and least useful for ongoing relationships, where we already knew everything in them. That is fine, and it argues for keeping the filter loose rather than clever, since a redundant brief costs seconds to skip.
The first unexpected finding was that the "last time" section carried most of the value, and it was the section we had built last. Public research about a company is genuinely available in a minute of searching; what nobody has time to do is find the thread from four months ago and remember what was promised. Once we wired prior notes and email in, the briefs stopped reading like research and started reading like handover.
The second unexpected finding was about padding. Before we added explicit permission to return nothing, roughly a third of the "what changed" sections contained something technically true, mildly related, and useless: a year-old blog post, a job listing, a mention in a roundup. It read as thorough. It was noise, and worse, it made the genuinely important updates harder to spot because everything had three bullets. Adding one line to the instruction fixed it, and the briefs got shorter and better in the same week.
The honest limit: a brief does not make you prepared, it makes preparation cheap. The two minutes of reading is still yours, and the meetings where it mattered most were the ones where we actually read it rather than skimming on the way in.
What Goes Wrong
The brief gets long. This is the default failure. Every additional useful-looking section reduces the chance the whole thing is read. Enforce one page as a rule, not as guidance.
Confident invention. An agent asked what a company does will always produce a sentence, including for companies about which little is published. Requiring a source per claim turns that from a silent risk into a visible blank.
Researching people rather than roles. The line to hold is professional and public. Role, company, published work. Assembling a personal picture of someone from scattered traces is a different activity with different obligations, it makes people uncomfortable when they find out, and it does not actually help the meeting.
Briefing meetings nobody needs briefed. A brief for the Monday standup trains you to stop opening briefs. Filter hard.
Stale internal context. If the "last time" section pulls from notes that were never written, it will confidently report no prior contact for an account a colleague has been working for months. Worth knowing which of your own records the agent can actually see, and saying so in the brief.
This fits alongside the rest of the pre- and post-meeting stack. Board meeting prep is the higher-stakes version, where the artefact is a pack rather than a page, and a competitor tracking agent covers the standing version of the "what changed" research rather than the per-meeting one. For teams doing this at scale across a pipeline, the best AI agents for sales teams compares the platforms built specifically around that workflow, and the cheapest AI agent platforms in 2026 covers what you can do without a budget conversation.
Frequently Asked Questions
What is an AI meeting prep agent?
A scheduled job that reads your upcoming calendar, works out which meetings are external and worth researching, and writes a short brief for each one before it starts. The brief covers who you are meeting, what their company does and what has changed recently, what your own records say about the last conversation, and a few specific questions worth asking. It runs before the day starts rather than when you remember.
How is this different from a meeting notetaker?
Opposite ends of the meeting. A notetaker records and summarises what was said after the fact, and there are good tools for that. A prep agent works before the meeting, assembling what you should already know walking in. Teams often run both, and the prep agent gets more value out of the notetaker's archive because last month's notes are one of its best inputs.
What should a pre-call brief contain?
Five things and no more: who is in the room and what they do, what the company does in one honest sentence, what has changed recently that is genuinely relevant, what happened last time you spoke, and two or three specific questions. Anything longer stops being read. The failure mode of automated briefs is not being too short, it is arriving as three pages nobody opens before dialling in.
Can the agent research the people I am meeting?
It can gather what is publicly published about them in a professional capacity, which is what a careful colleague would look up anyway: role, company, public writing or talks. Keep it to that. Building a personal profile from scattered sources is a different activity, it is not what makes a meeting go well, and depending on where you and they are it may carry obligations under data protection law. Role and context are useful; a dossier is not.
Which meetings should it skip?
Internal recurring meetings, one-to-ones, anything with no external attendee, and anything you have had forty times. Those need no research and a brief for them is noise that teaches you to ignore the briefs that matter. A simple rule that works: research meetings with at least one attendee from outside your email domain, plus anything explicitly flagged.
How much does a meeting prep agent cost to run?
On Gravity the free tier covers one agent at $0 a month, which is enough for one person's calendar. Paid plans start at $20 a month with $20 of usage included, and extra usage can be bought beyond the plan for a whole team's worth of daily briefs.
Will the brief be wrong sometimes?
Yes, and designing for that is the difference between a useful brief and a liability. Every claim should carry its source so you can check anything you plan to repeat out loud, and the agent should say it does not know rather than filling a gap. A brief that admits three blanks is more useful than one that reads complete and invented two facts.
Sources
- Gravity internal trial: a nightly calendar-triggered prep job run against the team's own calendar over several weeks in August 2026. Source for the observations in the "what we saw" section, including the padding behaviour before an explicit permission-to-return-nothing instruction was added. Qualitative, from our own use, not a controlled study.
- Gravity pricing: free tier with one agent at $0 per month; paid plans from $20 per month including $20 of usage, with extra usage available beyond the plan. Checked 26 August 2026.
- Related reading: building a research agent step by step, board meeting prep, competitor tracking agent, what an AI agent is.