AI Meeting Notes for Small Teams: What You Already Pay For and What Actually Gets Followed Up

Six weeks into using an AI notetaker, a four-person agency had 94 transcripts, 94 tidy summaries with bulleted action items, and a CRM in exactly the state it was in before. Nobody had read a summary past the second line. The notes were accurate. AI meeting notes had simply solved a problem that was never the bottleneck, because writing down what to do has never been the hard part of a meeting.

That gap is the whole subject of this guide. For a small team the tooling is now close to free — you very likely already pay for a notetaker without having switched it on — so the interesting decisions are not about transcription accuracy. They are about which meetings get recorded, where the resulting actions land, and who is accountable when nothing happens. Pricing comes first, because it is the part most teams get wrong in the expensive direction, and then the part that earns the money.

Check what you are already paying for

Three of the biggest vendors folded meeting AI into subscriptions you are probably already on, and the cost of not noticing is a duplicate $20-a-seat bill.

Google Workspace includes Gemini in Meet — the thing that takes notes and writes the summary — from Business Standard upward, at $14 per user per month on an annual commitment or $16.80 on the flexible monthly plan. The trap is the tier below it. Business Starter gets Gemini in Gmail but not Gemini in Meet, so a Starter team that wants meeting notes is choosing between an upgrade and a standalone tool, and the upgrade usually wins because it carries 2 TB of pooled storage with it.

Microsoft splits the difference differently. Copilot Chat comes with eligible Microsoft 365 plans at no extra cost, but Copilot inside the apps — including meeting recap in Teams — needs the paid add-on, which runs $18 per user per month on an annual commitment or $25.20 month to month. The $18 figure is a promotional rate running to 31 December 2026 and applies to the first year only; the standard price is $21. It attaches to Business Standard or Business Premium, so it is an addition to your existing bill rather than a replacement for it.

Zoom gives its free tier a taste and then meters it hard. ZoomMate Basic covers meeting summaries and in-meeting questions for three hosted meetings a month, AI note-taking three times a month, and 20 AI queries, on top of the 40-minute meeting cap the free plan has always had. Paid tiers come with 2,200 AI credits per user per month. For a team running fifteen client calls a week, the free allowance is a demo, not a plan.

Before you buy anything, open your billing page and find out which of these you already own. A team on Workspace Business Standard and Zoom Pro has two notetakers running and may need no third.

What a standalone notetaker adds

The bundled tools write a decent summary of a meeting held on their own platform. Standalone tools do three things they mostly do not: they work across Zoom, Meet and Teams from one place, they keep a searchable archive of every conversation your team has ever had, and they push structured fields into other systems. Only the third one reliably changes behaviour.

Prices below are from each vendor’s own pricing page on 21 October 2026, since this category reprices constantly and the comparison sites lag by months.

Otter meters by minutes. The free plan gives 300 minutes a month with a 30-minute ceiling per conversation, which rules out most real client calls. Pro lists at $16.99 a month, or $8.33 on annual billing, for 1,200 minutes and a 90-minute cap. Business lists at $30 a month, or $19.99 annual, for unlimited transcription and three concurrent meetings. The per-conversation ceiling matters more than the monthly pool: a 95-minute workshop on the Pro plan gets cut off five minutes from the end, which is where the decisions usually are.

Fathom meters by feature instead. Its free plan is unusually generous — unlimited recording and transcription with instant AI summaries — and the paid tiers buy judgement rather than volume. Premium at $20 a month, or $16 annual, adds AI-generated action items and a conversational assistant over your calls. Team is $19 per seat, or $15 annual, with a two-seat minimum. Business at $34, or $25 annual, is where CRM field sync, deal views and AI scorecards live. If the reason you want a notetaker is to stop updating HubSpot by hand, that is the tier you are actually buying, and the free plan will not get you there.

Fireflies does not publish per-seat pricing on its pricing page at all; it is billed per seat with volume discounts above 50 members and a sales conversation in between. That is a reasonable signal about who it is built for. Its CRM and app integrations sit on the paid tiers.

The short version for a team of two to ten: if you are already on Workspace Business Standard or paying for Copilot, start there and spend nothing. If you need one archive across three meeting platforms, Fathom’s free tier is the cheapest honest starting point. Move to a paid tier only when you can name the field you want written into your CRM.

The part that decides whether any of it was worth it

A summary that lands in a notes app is a record. A summary that lands where work already happens is a workflow. Most teams never make the second one, and then conclude that AI meeting notes do not do much.

Three patterns work, in ascending order of effort.

Native CRM sync. Pick a tool whose integration writes into the fields your pipeline reports on — next step, close date, pain points, competitors mentioned — not one that pastes a transcript into the notes box. This is almost always a higher-priced tier, and that is the honest cost of the feature. Once discovery calls are landing structured, the downstream work gets easier too: the same extracted fields are what make automated proposal generation produce something you would actually send.

Webhook into your task system. Every serious notetaker can fire a webhook or a Zapier event when a meeting ends. Route the action items into Asana, Linear, Todoist or whatever the team already opens every morning, with the owner and the due date set from the transcript. Actions that live in a notes tool do not get done, because nobody’s day starts there.

A digest instead of a stream. If your team has more than about ten meetings a week, per-meeting notifications become noise within a fortnight. A single Monday digest — every commitment made last week, grouped by person, with the ones still open flagged — gets read. The same logic drives the scoring step in AI lead qualification: aggregate first, then surface the exceptions.

Setting up AI meeting notes so follow-ups actually happen

Name the three fields before you choose a template. Not “summarise the meeting” — decide what a sales call must always produce (next step, budget signal, objection), what an internal planning call must produce (decisions made, owner, deadline), and what a support call must produce (issue, resolution, whether it needs a human). Everything else is colour.

Route by meeting type, not by meeting. Use calendar titles or attendee domains to classify. External attendee plus a title containing “discovery” means sales treatment and a CRM write. All-internal with more than four attendees means decisions-only treatment and no CRM write at all. Getting this wrong is how a salary discussion ends up summarised in a shared channel.

Put the human checkpoint at the write, not the read. Have the tool draft the CRM update and the follow-up email, and have the meeting owner approve both in one place within the hour. The approval step is what keeps a hallucinated commitment from becoming a contractual one, and it takes about ninety seconds. The same draft-then-approve shape works well for the post-meeting email itself, which is the pattern in AI-drafted follow-ups.

Read the first thirty transcripts yourself. Not the summaries — the transcripts, next to the summaries, checking what the model dropped. Every team finds the same class of error in its own vocabulary: product names mangled, a conditional “if we can get legal to agree” recorded as a commitment, a quiet participant’s objection missing entirely. Fix the template; do not switch tools. This is the same discipline that makes an AI phone answering setup work rather than embarrass you.

Decide the retention period on day one. Default settings keep everything forever, which is a choice nobody consciously made.

What this looks like with real numbers

Take the agency from the top of this piece: four people, around fifteen external calls a week, mostly on Meet with the occasional client who insists on Zoom. Two of the four own deals; the other two attend as specialists and never update the pipeline.

The instinct is to buy everyone a seat on a well-reviewed notetaker. Otter Business at $19.99 a seat on annual billing is $79.96 a month for four, and three-quarters of that spend goes to people who will never open the archive.

The cheaper arrangement starts with what is already on the invoice. They pay for Workspace Business Standard, so Gemini in Meet is covering every Meet call at no additional cost, and that handles the bulk of the summarising. The two deal owners then get Fathom Business at $25 a seat on annual billing — $50 a month — because that tier is where CRM field sync lives, and CRM field sync is the entire reason for the purchase. The specialists get nothing extra and read shared notes. Total new spend: $50 a month rather than $80, with the capability concentrated where it changes an outcome.

The routing rule that makes it work is one line: external attendee plus a calendar title containing “discovery”, “proposal” or a client name triggers the CRM write and a task assigned to the deal owner; everything else gets a summary in the shared drive and nothing more. Internal retros, one-to-ones and the Monday stand-up produce no notifications at all, which is why anyone still reads the ones they get.

Where it goes wrong

The bot changes the meeting. A visible notetaker joining a sensitive call makes people careful, and careful people say less. For negotiations, performance conversations and anything with a lawyer in the room, the right setting is off. Several tools now offer bot-free capture that records locally instead of joining as a participant, which removes the social friction but none of the legal obligation.

Consent is not a checkbox you ticked once. Recording rules vary by state and country, and all-party consent jurisdictions require everyone on the call to agree, not just the host. An announcement at the top of the call plus a line in the calendar invite covers the ordinary case. Cross-border calls and regulated industries deserve twenty minutes with someone qualified rather than a vendor’s compliance page.

Transcripts are discoverable. A searchable archive of every sales call is a genuine asset and a genuine liability, and which one it turns out to be depends entirely on a retention policy you set before you need it.

Seats get bought for the whole company. Only people who run external meetings or own follow-ups need a paid seat. The rest can read shared notes. A ten-person company rarely needs ten licences, and vendors are not going to point that out.

Everything gets recorded, including the meetings that should have been emails. A notetaker makes bad meetings cheaper to hold, which is the opposite of useful. The weekly digest is good at exposing this: a recurring call that produces no commitments three weeks running is a call to cancel.

What to measure

Transcription accuracy is the metric vendors compete on and the one that matters least. Track instead the share of meetings that produce a next step recorded in your task system or CRM within 24 hours — start by measuring it for a fortnight before you change anything, because the baseline is usually under a third and that number is the business case. Then watch the time from discovery call to proposal sent, which is where the compounding is, and the share of AI-drafted CRM updates that get edited before approval, which tells you whether your template is wrong.

If the first number does not move within a month, the tool is not the problem. The routing is, or nobody owns the follow-ups, and no amount of transcription quality will fix either. Start with what your existing subscription already includes, point its output at the system where work actually happens, and only pay for a second tool once you can name the field it is going to fill.