Should AI attend my meetings?

AI notetakers arrived in most organisations without anyone deciding. They were switched on by default, or by one enthusiastic person, and within a year the complete, searchable, permanent record of every conversation became normal. Almost nobody has asked what that changes, which makes this a good example of drift rather than design.

The productivity case is real and largely uncontested: fewer people typing, better recall, accessibility benefits for anyone who processes text more easily than speech. This page is about the four things the case leaves out.

What changes when the record becomes complete

1 · Note-taking was thinking, not transcription. Deciding what matters enough to write down is an act of judgement performed in real time. It is a compression that requires understanding. Automate it and you keep the record while losing the processing. This is germane load being removed, and it feels exactly like convenience.

There is a further loss that only shows up later. The person who took the notes usually understood the meeting best, and it was frequently the most junior person in the room. That was not an accident of hierarchy, it was how they learned what mattered.

2 · Complete records change what people say. A permanent, searchable, shareable transcript is a different setting from a room where someone jots down conclusions. Half-formed ideas, tentative disagreement and thinking aloud all get more expensive. Nobody announces that they are self-censoring, so the effect is invisible in exactly the way that makes it hard to argue about.

3 · The summary becomes the meeting. Within a few months, most people are reading the summary rather than attending or listening back. That is a reasonable individual choice and it means an increasing share of organisational understanding is mediated by a system nobody is checking. See how do I know when AI is wrong.

4 · Attendance stops meaning anything. If the record is automatic and complete, the marginal value of being present falls, and meetings drift towards broadcast. The thing that made a meeting worth having, that people had to be there and respond to each other, quietly becomes optional.

How little evidence there is

Directly: very little. There is no study of AI notetakers and organisational understanding, and this page is reasoning from adjacent findings rather than reporting a result.

The adjacent findings are relevant though. Cognitive load research establishes that removing the effortful processing removes the learning while leaving performance intact. Bastani and colleagues demonstrated exactly that pattern in a field experiment: same model, opposite outcomes, decided by whether the interface made the person do the work. And the summarisation problem is the reading-comprehension problem, where a summary of a document is reliably not equivalent to having read it.

What is genuinely unknown: whether organisations using automatic transcription understand their own decisions worse over time. It is a measurable question and nobody has measured it.

Which meetings are for producing a record

The useful frame is which meetings are for producing a record and which are for producing understanding in the people present.

For the first, automate freely. A status update, a project review, anything where the output is a set of decisions someone else needs: record it, summarise it, skip it if you can.

For the second, the recording is working against the purpose. A difficult conversation, a genuine disagreement, an early-stage problem nobody has framed yet, or any session whose real function is that four people leave understanding something they did not understand before. Those are the meetings where a complete record makes people more careful and less useful.

Most organisations have applied one policy to both, which is how a tool that is clearly good for the first category quietly degrades the second.

A practical rule

• Default on for record meetings, default off for thinking meetings. State which kind it is in the invitation. That single line does most of the work.
• Say it is running, every time. Consent is the minimum, and in many jurisdictions it is also the law. Take your own advice on that.
• Keep someone taking notes by hand in the meetings that matter. Not for the record. For the understanding, and rotate who it is.
• Read the summary against the transcript occasionally. A summary nobody ever checks is an unverified input into a lot of decisions.
• Notice if attendance is falling. If people stopped coming because the summary is enough, ask whether the meeting was ever needed, and be prepared for the answer to be no.

Related SuperSkills research

On the mechanism, cognitive load and desirable difficulty. On the organisational pattern, drift versus design and capability debt. On summaries as inputs, how do I know when AI is wrong. On what juniors lose, the missed reps.

Key sources

• Bastani, H. et al. (2025). Generative AI can harm learning. PNAS.
• Bjork, R. A. and Bjork, E. L. Desirable difficulties in theory and practice.
• Lee, H.-P. et al. (2025). The Impact of Generative AI on Critical Thinking. CHI 2025.

About this research

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026) and founder of The SuperSkills Intelligence Company. There is no direct evidence on AI notetakers and organisational understanding; this page reasons from adjacent findings and says so. Recording consent and notification are legal questions in many jurisdictions and this is not legal advice. Reviewed quarterly.