TL;DR. “This meeting should have been an email” now has a third option: this meeting should have been an agent. But AI will not make that call for you. In a six-month field experiment across 66 firms and 7,137 knowledge workers, published as an NBER working paper by researchers from Microsoft Research and Harvard Business School, workers randomly given an AI assistant inside their email, meeting and writing tools cut their email time: in the second half of the trial, those who used it spent two fewer hours a week on email. Time in meetings did not change: the estimates rule out effects outside a range of 0.01 hours less to 0.21 hours more per week, against an average of 5.22 hours. In other words, the tool saved email time and left the calendar untouched. The redesign has to be deliberate. Communication research gives a clean rule for it. Media synchronicity theory splits communication into conveyance, moving information, and convergence, building shared meaning, and holds that conveyance works better on low-synchronicity media while convergence needs high synchronicity. Conveyance is the part agents can take. Convergence is the part worth keeping live. We turn that into a five-question protocol, classify 14 common meeting types, and show where the evidence says AI summaries are good enough and where they are not: on real meeting transcripts, even the best model tested in one benchmark produced at least one factual error in 10.9% of summaries of main topics and 19.8% of summaries of marginal topics.
This piece connects two clusters on CEOtudent. On the meetings side, it builds on the true cost of meetings, the evidence on meeting-free days and how AI meeting notes changed the meeting. On the agents side, it assumes you know how to brief an AI agent and where to put human checkpoints.
Why “should have been an email” is no longer the right question
The old complaint assumed two options: meet live, or write it down. Writing it down had a cost that kept many meetings alive. Someone had to gather the updates, chase the missing ones, merge them and send the result. A status meeting was often the cheapest way to force that gathering to happen.
Agents change that cost. An assistant with access to your tickets, documents and shared drive can compile a digest nobody had to write. A meeting recorder can produce a summary for everyone who skipped. So the question for each recurring meeting is now three-way: keep it live, turn it into a written decision process, or hand the information-moving part to an agent.
The evidence says that choice does not make itself. In the field experiment by Eleanor Dillon, Sonia Jaffe, Nicole Immorlica and Christopher Stanton, workers used the AI tool through their meeting software as well as their email, yet the authors found no shift in time spent in meetings, and they could rule out increases in the number of meetings larger than 4.6%. Two caveats matter. The paper is a working paper, not yet peer reviewed, and three of its four authors work at Microsoft Research, the maker of the tool studied. Even so, the direction is clear: individual AI access trimmed the work a person controls alone, such as email, and did not touch the work a group controls together, such as the calendar. Meetings are a team design decision. Someone has to make it.
What communication research says: conveyance and convergence
The most useful theory for this decision is more than 25 years old and was written for exactly this problem. In 1999, Alan Dennis and Joseph Valacich proposed media synchronicity theory as a rethink of media richness theory, and in 2008 Dennis, Robert Fuller and Valacich extended it in MIS Quarterly.
The core claim is that communication is made of two processes:
- Conveyance is the transmission of information and deliberation on what it means. It can be divergent: each person processes the information on their own schedule. The theory holds that conveyance works better with media of low synchronicity.
- Convergence is the development of shared meaning: people agreeing on what the information means and what to do about it. The theory holds that convergence works better with media of high synchronicity, where people work on the same thing at the same time with a shared focus.
The 2008 paper lists five media capabilities that shape this: symbol sets, parallelism, transmission velocity, rehearsability (how far a sender can refine a message before sending it) and reprocessability (how far a message can be re-examined later). Written and recorded media score high on the last two. That is why a written update is often better than a spoken one for conveyance: the writer can get it right, and the reader can reread it.
The older theory it builds on adds the second half of the rule. Richard Daft and Robert Lengel, writing in Management Science in 1986, defined media richness as “the ability of information to change understanding within a time interval” and ranked face-to-face conversation as the richest medium. Their argument was that the hard problem in organizations is often not a lack of data but a lack of clarity, which they called equivocality: situations with multiple and conflicting interpretations. Rich, synchronous media earn their cost when the situation is equivocal. For well-understood messages and standard data, leaner media are enough.
Put the two together and you get the basis of the protocol below: conveyance can go asynchronous, and increasingly to agents. Convergence on equivocal questions is what live time is for.
A 2021 study in Information and Management by Willem Standaert, Steve Muylle and Amit Basu adds a practical detail. Analysing 612 meetings described by organizers at one international company, it grouped 19 meeting objectives into four families: exchanging information, communicating sentiments, making decisions and building relationships. The number of media capabilities organizers needed rose steadily from information exchange to relationship building. For routine information exchange, only two capabilities mattered much: hearing voices and sharing screens. All modes in that study were synchronous, so it cannot tell us whether a written digest would do as well, but it confirms the gradient the theory predicts.
What synchronous time costs, according to the evidence
Our meeting cost analysis covered the headline numbers on time and money. The studies below add something different: evidence on what live meetings do to the people in them, and why the design of a meeting matters more than its existence.
Table 1. Selected evidence on the costs and benefits of synchronous meetings
| Source | Design | Finding | What it means for the protocol |
|---|---|---|---|
| Luong and Rogelberg, Group Dynamics, 2005 | Daily diaries: 37 employees, 5 days | Participants averaged three meetings and 157.94 minutes of meetings a day; the number of meetings was positively related to daily fatigue and subjective workload | Every live meeting has a cost beyond its duration |
| Shockley and colleagues, Journal of Applied Psychology, 2021 | Four-week field experiment: 1,408 daily observations from 103 employees | Having the camera on in virtual meetings increased daily fatigue, which reduced voice and engagement in meetings; the effect was stronger for women and newer employees | If a meeting must be live, it does not have to be on camera |
| Fauville and colleagues, Computers in Human Behavior Reports, 2023 | Online convenience sample of 9,787 people | More frequent and longer video calls with fewer breaks were associated with more fatigue | Shorter, rarer live sessions are cheaper per decision |
| Perlow, Hadley and Eun, Harvard Business Review, 2017 | Survey of nearly 200 senior executives | Only 17% said their meetings were generally productive uses of group and individual time; 54% described the worst case, poorly run meetings that also crowd out solo work | Most meeting systems fail on design, not on existence |
| Kauffeld and Lehmann-Willenbrock, Small Group Research, 2012 | 92 videotaped team meetings, coded interaction | Constructive interaction such as problem solving and action planning was linked to meeting satisfaction and team productivity, and predicted organizational success 2.5 years later | Keep live the meetings where real problem solving happens, and protect that interaction |
| Mullen, Johnson and Salas, Basic and Applied Social Psychology, 1991 | Meta-analysis of brainstorming studies | Brainstorming groups were significantly less productive than the same number of people working alone, in both quantity and quality; the loss was larger in larger groups | Generate ideas separately, converge together |
Diehl and Stroebe, in a set of four experiments published in the Journal of Personality and Social Psychology in 1987, traced most of that brainstorming loss to production blocking: in a live group, only one person can speak at a time, so ideas wait and get lost. That finding is the cleanest example of the conveyance and convergence split in action. Producing ideas is a divergent, parallel task. Choosing among them is a convergent one. A single live brainstorm tries to do both at once and does the first part badly.
What agents can and cannot do with meeting work
Before handing meeting work to an agent, look at what the evidence says about its quality. The picture is consistent: AI is fast and good enough for most conveyance, and unreliable in exactly the places that need a human check.
Table 2. The AI and meetings evidence in numbers (published figures and CEOtudent calculations)
| Study | What was measured | Published result | Our reading |
|---|---|---|---|
| Dillon and colleagues, NBER Working Paper 33795, 2025 (revised November 2025) | Six-month randomized field experiment, 66 firms, 7,137 workers | Users spent two fewer hours a week on email; meeting time effects outside -0.01 to +0.21 hours a week ruled out, against a mean of 5.22 hours | The data rule out AI access cutting meeting time by more than 0.6 minutes a week on average, or 0.19% of meeting time. Meeting reduction has to be designed |
| Cambon and colleagues, Microsoft technical report, December 2023 | Lab task: summarize a 35-minute recorded meeting you missed; 33 of 57 participants had Copilot | Copilot summaries included 11.06 of 15 rubric details, against 12.40 without; Microsoft’s companion article reports 11 minutes 13 seconds against 42 minutes 34 seconds | About 26.4% of the time, or 3.8 times faster, for 10.8% fewer key details. Fast conveyance, with a measurable fidelity cost. Not peer reviewed; authors are Microsoft researchers |
| Tang and colleagues, TofuEval, NAACL 2024 | Human-annotated factual errors in topic summaries of real city council meeting transcripts (MeetingBank) | Share of summaries with at least one factual inconsistency: best model 10.9% on main topics and 19.8% on marginal topics; average of five models 30.4% and 43.6% | Even the best model tested got roughly 1 in 9 main-topic summaries and 1 in 5 side-topic summaries wrong somewhere. Side topics, where decisions often hide, are riskier |
| Ramprasad, Ferracane and Lipton, ACL 2024 | Errors in LLM summaries of everyday dialogues | About 23% of GPT-4 summaries contained inconsistencies; roughly 38% of LLM errors were circumstantial inferences, plausible assumptions not supported by the conversation | The typical error is not a wild invention but a reasonable-sounding guess about who agreed to what |
| Dell’Acqua and colleagues, Harvard Business School working paper, 2023; peer-reviewed version in Organization Science, 2026 | Field experiment with 758 BCG consultants | Inside the frontier of AI capability: 12.2% more tasks completed, 25.1% faster. On a task outside it, the working paper reports AI users 19 percentage points less likely to be correct | Agents help most on well-defined work and mislead on ambiguous work: the same line as conveyance versus convergence |
Two notes on the table. The TofuEval models are from 2023 and newer ones are likely better; the benchmark is still the most direct public evidence on meeting-transcript summaries we could find. And every AI finding here comes from studies run or funded by, or built on products of, companies with a stake in the result, or from academic benchmarks on older models. Treat them as the current best estimate, not the final word.
The practical rule that follows is simple. An agent can own the conveyance. A person must own the check on anything that becomes a commitment. A digest of project status can go out unreviewed; readers will flag errors. A summary that records who agreed to deliver what by when needs a named human to read it before it goes out, because the most common error type is exactly a plausible inference about agreement.
The CEOtudent Sync-or-Agent Protocol
Five questions, asked in order, for every recurring meeting on your calendar and every new one someone proposes.
Table 3. The Sync-or-Agent Protocol (CEOtudent editorial framework)
| # | Question | If yes | If no | Evidence basis |
|---|---|---|---|---|
| 1 | Is the main job of this meeting to move information from some people to others? | Conveyance: go to question 2 | Convergence: go to question 4 | Media synchronicity theory: conveyance suits low-synchronicity media |
| 2 | Does the information already exist somewhere an agent can read it, such as tickets, documents, dashboards or a shared drive? | An agent compiles it into a written digest | People write a short structured update and an agent merges the updates | Written media are rehearsable and reprocessable; nobody should read aloud what is already written |
| 3 | Would an error in the digest be expensive, because it records commitments, numbers or anything sent outside the team? | A named person checks the digest before it goes out | Send it; readers flag errors | Summary error rates of 10.9% to 43.6% in TofuEval; plausible inference is the most common error type |
| 4 | Is there real disagreement, ambiguity, or something personal at stake? | Keep it live, and shrink it with an agent-prepared pre-read and an agent-drafted record | Decide in writing: a proposal, a deadline and a named approver | Equivocality is what rich, synchronous media are for; constructive live problem solving predicts team and organizational outcomes |
| 5 | Does someone need to own the outcome visibly in front of others? | Name the decision owner in the invite and in the record | The written decision is enough | Agents can prepare and record, but cannot be accountable; see our piece on keeping accountability when agents do the work |
The protocol produces four outcomes, not two:
- Agent digest. Pure conveyance from existing sources. No meeting.
- Agent-merged written update. Pure conveyance, but the information lives in people’s heads. People write, the agent merges, nobody meets.
- Written decision. Convergence without real disagreement: a proposal circulates with a deadline and a named approver, and silence by the deadline counts as consent.
- Short live meeting with an agent before and after. Convergence on something ambiguous, contested or personal. The agent prepares the pre-read and drafts the record. The humans spend the live time only on the part that needs them.
How 14 common meetings classify
We ran the protocol against 14 meeting types that appear on most knowledge workers’ calendars. The classification is our editorial judgment, not a study result; your context can move a meeting from one column to another, and the protocol questions are there to let you check.
Table 4. Common meeting types through the Sync-or-Agent Protocol (CEOtudent editorial framework)
| Meeting type | Dominant process | Recommended format | What the agent does | What stays human |
|---|---|---|---|---|
| Status update or round-robin report | Conveyance | Agent digest | Pulls status from tickets and documents, flags blockers | Resolving any blocker it flags |
| Daily stand-up | Conveyance, with occasional convergence | Agent-merged written check-in; live 10 minutes only for flagged blockers | Collects check-ins, groups blockers | The blocker conversation |
| Weekly metrics review | Conveyance, then interpretation | Agent digest plus short live discussion of anomalies | Drafts the metric narrative and highlights changes | Deciding what the anomalies mean and what to do |
| Announcement or all-hands | Conveyance with trust at stake | Recorded or written briefing plus live questions | Drafts the written version and collects questions in advance | Answering questions, especially hard ones |
| Document review or approval | Mostly conveyance | Asynchronous comments; live only for unresolved disagreements | Consolidates comments and lists open conflicts | Settling the conflicts |
| Scheduling and coordination | Conveyance | Agent | Proposes times and circulates the plan | Nothing, unless priorities clash |
| Brainstorm | Divergent, then convergent | Ideas generated separately and in writing, then a short live selection session | Clusters and de-duplicates the ideas | Choosing among them |
| Decision meeting with clear options | Convergence | Written decision if no real disagreement; otherwise short live meeting | Drafts the options memo and the decision record | The decision and its owner |
| Project kick-off | Convergence on goals and roles | Live, once | Drafts the brief and the role map beforehand | Agreeing on goals and who owns what |
| Retrospective | Convergence on meaning | Written input first, live discussion second | Collects and groups the input | Interpreting it and committing to changes |
| One-to-one with a manager | Convergence and relationship | Live | Prepares an agenda from open items, with consent | The conversation |
| Performance or difficult feedback | Convergence, high personal stakes | Live, without recording or agent in the room | Nothing during; at most, helps the manager prepare | All of it |
| Conflict resolution or negotiation | Convergence under equivocality | Live | Background preparation only | All of it |
| Incident response | Urgent convergence | Live | Keeps the timeline and action log | Every decision |
Two patterns stand out. First, the meetings that are easiest to hand to an agent are also among the most common: in a 2014 study by Joseph Allen and colleagues in Management Research Review, which built a 16-category taxonomy of meeting purposes from 491 working adults, the two most common purposes were discussing ongoing projects (11.6%) and routinely discussing the state of the business (10.8%). Together that is 22.4% of meetings described, about 6.8 times the share of brainstorming (3.3%), the least common category. Treating those two categories as update meetings is our reading, not the authors’ label, but the direction is hard to miss: a large share of meetings exist mainly to move information.
Second, every meeting that stays fully live involves either personal stakes or unresolved ambiguity. That is not a coincidence. It is the protocol working as the theory predicts.
The hybrid pattern: agent before, humans during, agent after
Most meetings that survive the protocol can still be much shorter. The pattern that does this has three parts.
- Before: the agent writes the pre-read. It pulls the relevant documents, summarizes positions already stated in writing, and lists the specific questions the meeting must answer. Participants read it before they arrive. The live time starts at the disagreement, not at the background.
- During: humans converge. No status reports, no reading slides aloud. The meeting exists to resolve the listed questions. If a question gets resolved in the pre-read comments, it comes off the agenda.
- After: the agent drafts the record, a human signs it. The draft lists decisions, owners and dates. A named person checks it before it is sent, because a record of commitments is exactly where a plausible but wrong inference does the most damage.
If your organization records meetings, the norms we described in AI meeting notes changed the meeting apply: consent, retention and the effect of recording on candour.
A two-week calendar audit
You can run the protocol on your own calendar in an afternoon.
- List every meeting from the last two weeks. Include recurring ones you declined.
- Tag each one with its dominant process: conveyance, convergence, or mixed.
- Run the five questions and write the recommended format next to each.
- Estimate the hours that would move from live time to an agent digest, a written update or a written decision.
- Pick the two highest-hour recurring meetings and propose the new format to their owners for a four-week trial, with a review date set in advance.
- Measure what matters: decisions made on time, blockers resolved, and whether anyone missed information they needed. Hours saved are a by-product, not the goal.
The CEO half of this exercise is the redesign: meetings are a system someone owns, and the field evidence says AI will not redesign it for you. The student half is the review: after four weeks, look at what broke, what nobody missed, and what should come back.
When not to replace a meeting
The protocol is a default, not a law. Keep the live meeting, even for conveyance, in four situations:
- When trust is low or falling. A written digest cannot rebuild trust. Seeing and hearing people can.
- When the news is bad or personal. Layoffs, reorganizations, failures and feedback deserve a live human, and the recording question deserves a careful answer.
- When people are new. New team members need the relationship-building that Standaert and colleagues found requires the richest set of capabilities, and they are the group Shockley and colleagues found most fatigued by camera-on meetings, so make those sessions live but light.
- When nobody reads the written version. An agent digest only replaces a meeting if people read it. If they do not, fix the reading habit before you cancel the meeting. Our piece on async-first work architecture covers the norms that make written communication work.
Frequently asked questions
Can AI replace meetings?
It can replace the information-moving part of many meetings, such as status updates, coordination and document review. It does not replace the part where people build shared meaning, settle disagreements or make commitments. In the largest field experiment so far, access to an AI assistant reduced email time but did not change time spent in meetings, so the replacement has to be designed deliberately.
Which meetings should be replaced by AI agents first?
Status updates, round-robin reports, scheduling and coordination meetings, and most document reviews. They are mainly conveyance: moving information that often already exists in tickets and documents. In one taxonomy study, project and business-status discussions alone made up 22.4% of the meetings people described.
Are AI meeting summaries accurate enough to trust?
Accurate enough for low-stakes information, not for unreviewed records of commitments. In one benchmark on real meeting transcripts, the best model tested produced at least one factual error in 10.9% of main-topic summaries and 19.8% of side-topic summaries. Have a named person check any summary that records decisions, numbers or promises.
What is the difference between conveyance and convergence?
Conveyance is moving information and letting each person make sense of it on their own schedule. Convergence is reaching a shared understanding of what the information means and what to do. Media synchronicity theory holds that conveyance works better asynchronously and convergence works better in real time.
Should brainstorming be done live?
Idea generation works better separately. A 1991 meta-analysis found brainstorming groups were less productive than the same number of people working alone, in both quantity and quality. Generate ideas in writing first, then meet live to choose among them.
How do I propose cancelling a recurring meeting without offending its owner?
Propose a time-limited trial with a review date, not a cancellation. Suggest the specific replacement format, such as an agent digest or a written decision with a deadline, and agree in advance on what would bring the meeting back.
Sources
Dillon, E. W., Jaffe, S., Immorlica, N., and Stanton, C. T. Shifting Work Patterns with Generative AI. NBER Working Paper 33795, May 2025, revised November 2025.
Dennis, A. R., Fuller, R. M., and Valacich, J. S. Media, tasks, and communication processes: a theory of media synchronicity. MIS Quarterly, 2008, volume 32, issue 3, pages 575 to 600.
Dennis, A. R., and Valacich, J. S. Rethinking media richness: towards a theory of media synchronicity. Proceedings of the 32nd Hawaii International Conference on System Sciences, 1999.
Daft, R. L., and Lengel, R. H. Organizational information requirements, media richness and structural design. Management Science, 1986, volume 32, issue 5, pages 554 to 571.
Standaert, W., Muylle, S., and Basu, A. How shall we meet? Understanding the importance of meeting mode capabilities for different meeting objectives. Information and Management, 2021, volume 58, issue 1, article 103393.
Luong, A., and Rogelberg, S. G. Meetings and more meetings: the relationship between meeting load and the daily well-being of employees. Group Dynamics: Theory, Research, and Practice, 2005, volume 9, issue 1, pages 58 to 67.
Shockley, K. M., and colleagues. The fatiguing effects of camera use in virtual meetings: a within-person field experiment. Journal of Applied Psychology, 2021, volume 106, issue 8, pages 1137 to 1155.
Fauville, G., and colleagues. Video-conferencing usage dynamics and nonverbal mechanisms exacerbate Zoom fatigue, particularly for women. Computers in Human Behavior Reports, 2023, volume 10, article 100271.
Perlow, L. A., Hadley, C. N., and Eun, E. Stop the meeting madness. Harvard Business Review, July to August 2017.
Kauffeld, S., and Lehmann-Willenbrock, N. Meetings matter: effects of team meetings on team and organizational success. Small Group Research, 2012, volume 43, issue 2, pages 130 to 158.
Mullen, B., Johnson, C., and Salas, E. Productivity loss in brainstorming groups: a meta-analytic integration. Basic and Applied Social Psychology, 1991, volume 12, issue 1, pages 3 to 23.
Diehl, M., and Stroebe, W. Productivity loss in brainstorming groups: toward the solution of a riddle. Journal of Personality and Social Psychology, 1987, volume 53, issue 3, pages 497 to 509.
Allen, J. A., Beck, T., Scott, C. W., and Rogelberg, S. G. Understanding workplace meetings: a qualitative taxonomy of meeting purposes. Management Research Review, 2014, volume 37, issue 9, pages 791 to 814.
Cambon, A., and colleagues. Early LLM-based Tools for Enterprise Information Workers Likely Provide Meaningful Boosts to Productivity. Microsoft technical report, December 2023; with Microsoft WorkLab, What Can Copilot’s Earliest Users Teach Us About Generative AI at Work, November 2023.
Tang, L., and colleagues. TofuEval: Evaluating Hallucinations of LLMs on Topic-Focused Dialogue Summarization. Proceedings of NAACL 2024.
Ramprasad, S., Ferracane, E., and Lipton, Z. C. Analyzing LLM Behavior in Dialogue Summarization: Unveiling Circumstantial Hallucination Trends. Proceedings of ACL 2024.
Dell’Acqua, F., and colleagues. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. Harvard Business School Working Paper 24-013, 2023; published in Organization Science, 2026, volume 37, issue 2.
The derived figures in Table 2 (the 0.6-minute and 0.19% bound, the 26.4% time share and 10.8% detail gap, and the one-in-nine and one-in-five error readings) and the 22.4% and 6.8-times figures in the classification section were computed by us from the published numbers. Tables 3 and 4 are CEOtudent editorial frameworks.
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:
