What agencies should check before relying on an AI note taker

Which tools are covered here?

What should an agency check before AI notes become standard?

Most agencies arrive at this question late. An account manager signs up for a free AI note taker, uses it on a few client calls, and only when it is already standard practice does someone ask what happens if a client objects, a bot gets blocked at the client's door, a wrong action item lands in a recap, or a departing employee takes the recordings with them.

Before an AI note taker becomes standard practice for client meetings, an agency lead should work through eight checks: client consent, platform admission, vendor data terms, workspace ownership, recap sharing defaults, accuracy review, coverage of phone and in-person meetings, and a written policy. Each check exists because something specific can go wrong, and each has a setting or document where you can find the answer.

Notta is used throughout this guide as the example of a tool that satisfies these checks; the other vendors illustrate where specific checks can fail. The rest of this guide walks through each check, names the settings and documents to open, and ends with a policy template you can copy.

Has the client agreed, and will their platform let your note taker in?

Consent comes first. Tell clients you use an AI note taker, in the engagement letter or the kickoff agenda, and give them a way to opt out. Recording consent laws vary by state, and a client who is surprised by a bot in the room is a client relationship problem even when the law is on your side.

The second half of this check is technical: even a consenting client may sit behind a platform that blocks your note taker. Microsoft Teams holds detected external bots in the lobby by default — "when detected, require approval before joining" — so a bot-based note taker can simply never appear unless the client's admin admits it. Google Meet changed how guests, including notetaking bots, are admitted in 2026, so behaviour your team learned last year may no longer apply.

If your clients are large enterprises, expect the bot question to come up regularly. The alternative is a bot-free note taker that captures audio on the device rather than joining as a participant — Notta, for example, offers both a bot and a desktop app with a Privacy Mode that records without a visible participant, which keeps the decision in your hands rather than the client's IT department's. We cover the trade-offs in meeting bot vs bot-free recorder for consultants, and state-by-state rules in our recording consent laws guide.

Do the vendor's data terms fit your client contracts?

Your client contracts probably promise confidentiality. Your note taker's privacy policy may not. Three terms matter most: whether the vendor trains AI models on your meeting data, how long recordings are retained, and which sub-processors see the data.

The vendors differ, and the differences are not always where you expect. Otter.ai's privacy policy lists training its proprietary AI on de-identified recordings and transcripts as a processing purpose (its FAQ separately states AI service providers do not train on Otter customer data). Fireflies states it does not train on customer data. Fathom's terms allow de-identified data to improve its models. Granola's free and Business plans may train on anonymised data unless you opt out. Notta excludes customer data from training and offers automatic deletion on its Enterprise plan. Retention controls are frequently gated to Enterprise tiers across the market, so check whether a non-Enterprise admin can actually shorten retention for the whole workspace.

If your client contracts include NDAs or data-processing commitments, read the vendor's sub-processor list and ask for a DPA before the first recorded call. Litigation risk is no longer theoretical: In re Otter.AI Privacy Litigation was filed in the N.D. Cal. in August 2025, as reported by law-firm summaries — cite the court record or a dated summary rather than press coverage.

For a capability-level comparison of tools, see AI tools for consultants.

Who owns the recordings when an account manager leaves?

The quietest risk in this whole checklist is account structure. If your account managers signed up with personal free accounts, every client recording they ever made lives in their personal workspace — and leaves with them when they resign.

The fix is an agency-owned workspace with admin roles, so recordings belong to the agency and access can be revoked or transferred centrally. Check the caps before you assume this is possible: Otter's Basic workspaces are limited to 5 members and Pro to 5 billed users (Business allows 25, Enterprise unlimited), which constrains how small teams can consolidate. Also test what actually happens when you remove a user — whether their recordings are kept, transferred or deleted — before you rely on it in an offboarding.

Notta's Business and Enterprise plans are built around shared workspaces with admin-managed members, which makes the agency-owned model the default rather than a workaround. For what team plans cost in practice, see AI meeting tool team cost.

Who receives the recap, and can it leak outside the meeting?

A recap is a document about your client, written by software, and its default audience is not always who you would choose. Google Meet's "Take notes for me" emails notes according to the organiser's setting — which can be "all invited guests, including those outside of my organization" — and saves them to the organiser's Drive. Fireflies can expose recaps through public meeting links. Otter supports public links and channels.

Before your first client call, open the sharing settings and find the default: does the tool auto-share recaps with all attendees, including external guests? Can a link be made public? Then set the defaults you want rather than inheriting the vendor's.

Sharing controls are one of the clearest differentiators in this market. Notta, for example, supports password-protected sharing alongside links and email, so a recap sent to a client can be protected even if the link is forwarded — a meaningful safeguard when the recap contains strategy, pricing or personnel details. Fathom's capture-mode indicator and banner (which can be disabled on paid tiers) address the related consent question of whether participants know they are being recorded.

How accurate are the action items, and who checks them?

Accuracy is the check teams skip, because the notes look finished. They are not. In a 2025 Oasis Group test reported by Kitces, adviser note-taking tools captured data points with 85–96% accuracy — but action items were right only 62–87% of the time. An action item that is invented, misassigned or subtly wrong is worse than a missing one, because it looks authoritative.

Hallucination is a documented failure mode, not a rare anomaly: research on OpenAI's Whisper found it regularly invents phrases, including during silence. The same failure surfaces in meeting tools built on similar speech-recognition foundations.

The practical fix is a reviewer rule: no AI recap goes to a client, or into the project record, without a human checking the action items against what was actually agreed. Name the reviewer per account. For a deeper method, see how to verify decisions in AI meeting summaries.

Does it cover phone calls and in-person meetings too?

Client conversations do not only happen on Zoom. If your note taker only works by sending a bot into scheduled video meetings, it will miss phone calls, in-person workshops and coffee-shop check-ins — often the conversations where the real decisions get made.

Check whether the tool offers a mobile app for phone calls, a desktop app for in-person capture, and file upload for existing recordings. Notta is one of the few that combines all three capture modes — bot, desktop app with Privacy Mode, and upload — so one tool and one set of governance controls covers every meeting type an agency actually runs. We cover the full recording workflow in how agencies record client meetings.

What does a one-page AI note-taker policy for an agency include?

Once the checks pass, write them down. A one-page policy keeps ten account managers from making ten different judgement calls. Copy this template:

• Consent: every client is told we use an AI note taker, in the engagement letter; anyone may ask us to turn it off for a call.
• Tool and plan: we use [tool] on [plan], in an agency-owned workspace; personal accounts are not permitted for client calls.
• Data terms: the vendor's policy on model training and retention is [X]; client contracts with NDAs take precedence.
• Sharing: recaps are shared only with named recipients; public links and password-protected links follow [rule].
• Review: the account lead checks action items before any recap is sent to a client or filed.
• Departure: when staff leave, their workspace access is revoked and recordings transfer to [owner].
• Escalation: if a client objects after a call, the account lead handles it and logs the outcome.

A tool that already satisfies several of these checks makes the policy shorter. Notta, for instance, covers the sharing rule with password-protected links, the data-terms line with no training and auto-deletion on Enterprise, and the capture rule with bot, desktop and upload modes — so the policy documents your choices rather than working around the tool's limits. For the IT-side view of the same problem, see how to stop AI note takers joining the wrong meetings.