September 25, 2026 · 10 min read

AI Notetaker for Consultants: How to Capture Client Sessions Without Awkward Bots or NDA Risk

Published September 25, 2026

Consultants have a very specific AI notetaker problem, and most tools on the market do not solve it. You are not a sales rep capturing 30-minute demos into a CRM. You are running an hour-long discovery call with a Fortune-500 CFO, a 90-minute strategy session with a founder's whole leadership team, or a two-day workshop where the deliverable depends on catching the loose sentence someone said in the third break. What you need is a record of what actually happened. What you do not need is a visible bot in the participant list that quietly makes your client stop telling you the truth.

This guide is for solo consultants, boutique firms, and fractional executives evaluating AI meeting notes for consulting work. It covers the three specific risks bot-based cloud notetakers create for your practice — client candor, NDA conflicts, and multi-client bleed — and the architectural alternative that removes all three at once.

The problem with a bot in the room

The core insight everyone underestimates: the moment a client knows they are being recorded by an AI service, the quality of the conversation drops. This is not paranoia; it is documented practice. AI implementation consultant Lilach Bullock has written about a discovery call where a financial-services prospect noticed Fireflies had auto-joined 12 minutes in. From that moment forward, the client "chose her words for the rest of the call the way you choose words when you know a lawyer might read them back to you one day." The contract closed. The call was worse. The tangents disappeared, and with them the disclosures that make discovery calls valuable in the first place.

Even Granola, a bot-free notetaker vendor, concedes the same dynamic in its own marketing: "When a visible recording bot joins a client call, clients become guarded. In a discovery session, executives speak in measured, cautious terms." That is the whole ballgame for consulting. If your notetaker degrades the raw material of the engagement, no summary quality can recover it.

Why NDAs and MSAs collide with cloud notetakers

The second problem is contractual, and it is not hypothetical. Nearly every consulting engagement runs under an NDA or a Master Services Agreement that restricts disclosure of the client's confidential information to "authorized third parties." When you route the audio of that meeting through Otter.ai, Fireflies, or Rev, you have introduced a third-party processor to the confidential relationship — usually one your client's procurement team has never approved.

Gouchev Law's analysis of AI vendor contracts flags this specifically: "a lot of consulting companies discover their AI-generated output comes with licensing restrictions their clients never agreed to accept." A vendor DPA that permits model training, a sub-processor list your client has never seen, and a retention window that outlives the engagement can each independently put you in breach of the master agreement you signed to win the work.

DataGrail's 2026 governance discussion put it directly: for organizations using AI notetakers, the question is "whether your Data Processing Agreements (DPAs), Business Associate Agreements (BAAs), or Master Service Agreements (MSAs) were written to cover AI processing." For most consultants, the honest answer is no.

The wiretap question consultants can no longer ignore

On August 13, 2026, Judge Eumi K. Lee of the Northern District of California ruled on Otter's motion to dismiss In re Otter.AI Privacy Litigation (5:25-cv-06911-EKL). The court, as summarized by Recording Law's analysis of the order, allowed the federal Wiretap Act claim, the California Invasion of Privacy Act claim, and both Illinois biometric-privacy claims to proceed to discovery — on the theory that Otter plausibly acted as a third-party eavesdropper rather than a mere tool of the meeting host.

The reasoning matters for every consultant using a cloud notetaker. As Sheppard Mullin's client alert explained, the court rejected Otter's "authorized participant" defense specifically because "Otter uses the notetaker for its own commercial purpose: training AI." That distinguishes it from earlier rulings where the recording tool acted as a pure extension of the customer. If your notetaker vendor is deriving separate commercial value from your client's audio, the vendor is arguably a separate ear on the call — and under California's Penal Code section 632, the statutory exposure is $5,000 per violation.

You are not the defendant in the Otter case. But you are the meeting host who invited the vendor into your client's conversation. Read our procurement checklist for AI notetakers after the Otter ruling for the questions to work through before the next engagement letter goes out.

Cloud notetakers vs on-device: the consultant matrix

Here is what actually varies between architectures, mapped to the consulting-specific criteria that matter.

Criterion	Cloud bot-based (Otter, Fireflies)	Cloud bot-free (Granola, Fathom)	On-device (Basil AI)
Visible in participant list	Yes — chills candor	No	No
Audio uploaded to vendor cloud	Yes, stored	Yes, transcribed then usually deleted	No — never leaves device
Default: trains on your client's audio	Often on by default (must opt out)	Varies by tier	N/A — vendor never receives it
Wiretap-suit exposure theory applies	Yes (see Otter ruling)	Partial (Granola sued Aug 2026)	No third-party ear on the call
Multi-client separation	Workspace feature; enforcement is policy	Workspace feature; enforcement is policy	Architectural — each device holds its own data
Works offline (site visits, workshops)	No	Limited	Yes
Subpoena surface at vendor	Full audio + transcripts	Transcripts (audio often deleted)	None

The multi-client bleed problem

Consultants running more than one engagement — especially where clients compete — face a problem that single-tenant SaaS was not designed for. The Fireflies 2026 consultant guide is candid about this: "Multi-client workspace separation and a no-training data policy are non-negotiable for any firm handling more than one engagement." If your tool doesn't offer workspace separation, you are relying on manual file organization, "which creates confidentiality risk as your client base grows."

Workspace features help, but they enforce separation through vendor policy, not physical isolation. On a fully on-device tool, the isolation is architectural: transcripts and audio live in the local file system of the device you used for that meeting, not in a vendor-side searchable index. A former staff member who leaves the firm cannot search transcripts they were never on because there is no cloud index for them to log into.

The Bring-Your-Own-Notetaker problem, in reverse

Every consultant has now been on the receiving end of an uninvited client notetaker. Bullock's analysis of NDA risk is worth reading in full, but the practical rules for the consultant side are:

• Announce your notetaker verbally at the top of every call, even with clients you have worked with for years. If you are using a bot-free on-device tool, that sentence is short and honest: "I'm running a private notepad on my Mac that transcribes locally — nothing leaves this machine."
• Ask what the client is running, and note it in your own file. If a bot appears mid-call, stop and confirm what happens to the recording afterwards.
• Add a recording clause to your engagement letter. Specify whether either party's AI notetaker may be used, where processing happens, and the retention window. Your GC drafts this; do not freelance the language.
• Default to disclosure regardless of jurisdiction. California, Illinois, and Florida require all-party consent, and GDPR requires explicit consent for processing personal data, which includes meeting recordings. Announcing is cheaper than defending.

Discovery risk: your notes are now business records

Consultants underweight this until it happens. Once an AI-generated transcript exists, it is a business record. Duane Morris's February 2026 analysis warned that AI-transcribed conversations "may become discoverable in litigation" and that meeting notes "can create a permanent, searchable record that may later be preserved and produced in litigation."

If your client is sued and you are a witness or third-party subject to subpoena, the transcripts of your discovery sessions, workshops, and strategy calls are on the table. Every cloud vendor that holds copies expands the subpoena surface. On-device transcripts you delete when the engagement closes shrink it. This is not a promise that on-device makes discovery vanish — your firm's retention policy still governs — it is a promise that fewer copies exist in fewer places. Our deeper piece on AI transcripts as discoverable evidence walks through the Heppner ruling and the mechanics.

What "good" looks like: a consultant's evaluation checklist

Before the next client engagement, work through the following. Where your current tool fails a row, decide whether to change the tool or change the engagement letter — but do not leave the gap open.

Architecture

• Where is the audio processed? On-device, in the vendor cloud, or on a third-party inference API?
• Is any raw audio retained after transcription, and if so where and for how long?
• Is a visible bot added to the participant list, or is capture bot-free?

Contract

• Does the vendor DPA prohibit model training on customer content by default (not just as an opt-out)?
• Has your MSA with each client been checked against the vendor's sub-processor list?
• Is there a clean deletion path, and does the vendor confirm deletion within a stated window?

Consent and disclosure

• Do you have a standard opening sentence you deliver on every call?
• Does your engagement letter explicitly address AI transcription?
• For all-party consent states, do you get on-record verbal agreement from each participant?

How Basil AI solves this

Basil AI is built for exactly this problem. It runs on iOS and macOS, uses Apple's on-device Speech Recognition on the Apple Neural Engine documented in Apple's privacy architecture, and never uploads your client's audio to a vendor cloud — because Basil is not a cloud service. There is no server holding the recording of your discovery call, because there is no server in the loop.

Concretely, that means:

• No bot in the participant list. Basil captures device audio and microphone input locally. Your client sees you on the call, not a third-party bot.
• Nothing to subpoena at a vendor. Because the audio and transcript live only on your device, a subpoena to Basil's vendor entity would not produce your client's session content — it is not there.
• No training use of your client's audio, ever. This is architectural, not a policy toggle. The audio never reaches us, so we cannot train on it if we tried.
• Multi-client separation by physical isolation. Each device holds its own data. There is no shared cloud index for a former team member to search.
• 8-hour continuous recording for full-day workshops, discovery sprints, and offsites — offline, so it works on a client site with locked-down Wi-Fi.
• Apple Notes integration so your action items and structured summaries land in the workflow you already use, without a separate cloud dashboard.

The important honest note: Basil AI is not a compliance product and we do not claim to be "compliant" with any specific regime. Compliance is a determination you and your client's counsel make. What we can say factually is that on-device processing means there is no vendor server holding your client's recording. That architectural fact is what the Otter, Granola, and Fireflies plaintiffs are asking courts to weigh. It is easier to defend an architecture where the vendor never had the audio than one where the vendor promised to be careful with it.

If you also serve regulated clients, our companion pieces on AI meeting notes for solo attorneys and privacy-aware AI notes for asset managers go deeper on the industry-specific angles.

The bottom line for consultants

Your business runs on trust and on the loose sentence in the third break. The tool that best serves it is the one that lets the client forget it is there. A bot-based cloud notetaker fails on both counts: the bot is visible, the vendor is present, and the audio outlives the engagement. An on-device notetaker fails on neither: no bot, no vendor, no persistent third-party copy. That is not a marketing claim — it is what the architecture forces to be true.

If you take one action from this article, make it this: on your next discovery call, listen for the moment your client relaxes. If a bot in the participant list stops that moment from ever arriving, you are paying a price your notes cannot pay back.