Do All AI Meeting Notetakers Train on Your Data? What Financial Firms Need to Know

A single missed detail in a diligence call can cost a deal. That's the whole pitch for an AI meeting assistant: catch every word, surface every commitment, never lose context between the partner meeting and the term sheet. But for firms handling deal notes, that same recording creates a question deal teams didn't have to ask five years ago: where does this data go after the meeting ends, and could it end up shaping someone else's AI model?

The honest answer is that it depends entirely on which notetaker a firm uses, and the differences between vendors are larger than most buyers assume. Deal notes routinely contain material non-public information (MNPI): valuation discussions, pending transactions, management commentary that hasn't been disclosed. If a notetaker's default policy is to use customer transcripts to improve its underlying AI model, even in de-identified form, a firm's most sensitive conversations become training data, processed in ways a firm can't easily audit. This piece compares how the major AI notetakers actually handle training on deal content, what regulators and compliance experts are now saying about it, and what a financial-grade evaluation looks like.

What does it mean for an AI notetaker to "train" on your data?

Training on a call means the vendor feeds your transcript, summary, or audio into the process that improves their underlying AI model, so patterns from your specific conversation can influence how the model behaves for other customers. This is different from the AI simply generating a summary for you in the moment. Training is what happens after that summary exists, when the raw content or the output gets fed back into a model's learning pipeline, sometimes after stripping identifying details first.

Many AI meeting assistants train on customer data by default, framed as a way to improve product quality, and often with an anonymization step presented as the safeguard. That tradeoff is reasonable for some use cases. But it typically requires the customer to notice a setting buried in an admin panel and manually opt out, if an opt-out exists at all. For a deal team recording a partner call, that default is a liability few firms would accept if they understood the mechanics, anonymization included.

How the major AI notetakers compare on training policy

Training policies vary more than vendor marketing pages usually let on, and they change over time as companies respond to scrutiny. Here's where each stands, verified as of August 2026:

A few things worth noting about this table. Granola and Otter both frame their default training as a lower risk because the data is de-identified first, but anonymization has limits: specific figures, names, and deal-specific language can sometimes be reconstructable even after identifiers are stripped, and the underlying transcript still has to be processed before that stripping happens. Fireflies' current no-training stance is a real, credible shift, but it depends on the customer trusting a policy that could change again, since it's a stated commitment rather than an architectural constraint. Granola and Otter's opt-out model means a firm's deal content is training-eligible by default, in de-identified form, until someone finds the setting and turns it off, and for Granola specifically, that default only flips automatically on Enterprise plans. Fellow's model removes the question entirely: there's no data eligible for training and no anonymization process to trust, because customer content was never in that pipeline to begin with.

Privacy policies change, so treat this table as a snapshot rather than a permanent reference, and verify current terms directly with any vendor before a deal call touches sensitive content.

Why deal notes carry more risk than typical meeting content

Not all meeting content is equal risk. A standup or a marketing sync carries little downside if a vendor's systems retain it longer than expected. Deal notes are different, and regulators are starting to say so directly.

In a July 2026 client alert, Skadden, Arps, Slate, Meagher & Flom addressed how AI tools that can access nonpublic information at scale intersect with existing insider trading and MNPI rules. The firm's core warning: broker-dealers and investment advisers must maintain policies and procedures reasonably designed to prevent the misuse of MNPI, even when no one ultimately trades on that information, and firms risk scrutiny if AI tools foreseeably could use restricted data improperly. Skadden recommends firms inventory their data restrictions, segregate and permission AI access to nonpublic information, and maintain audit trails before granting any AI tool access to sensitive conversations, a standard that applies directly to a notetaker sitting in on a diligence call, anonymization notwithstanding.

That regulatory framing lines up with what independent compliance advisers are seeing in practice. ACA Group, a compliance consultancy that works with regulated financial firms, has warned that AI notetaker tools can capture sensitive discussions and potentially MNPI before compliance has had a chance to review it, creating data leakage, client confidentiality, and inconsistent oversight risk across a firm. Once an AI notetaker produces a transcript or summary of a deal conversation, that record also becomes something examiners can request, which raises the stakes on how long a vendor retains it, and who else touches it, in the meantime.

How to evaluate a notetaker's training and retention policy

Applying Skadden's inventory-and-permission framework specifically to a meeting notetaker breaks into four concrete checks:

Training default, not just training option

Does the vendor train by default (even on anonymized data) with opt-out available, or does it never train at all, on any plan? As the comparison above shows, this varies significantly even among well-known tools, and "opt-out available" still means the data was eligible for training until someone flipped a switch. "Anonymized" also isn't the same guarantee as "never processed for training in the first place."

Retention controls, including zero-day options

Can the firm configure how long recordings, transcripts, and audio persist after a meeting, down to deleting raw content immediately after processing while keeping summaries and action items? A zero-day retention setting means raw content never has a window in which it could be trained on or breached.

Access permissioning and redaction

Can the firm restrict a deal call to only the deal team, redact specific segments after the fact, or pause and resume recording mid-call when the conversation turns to MNPI? This is the segregation and permissioning piece Skadden points to, applied specifically to meeting content.

Archiving and audit trail integrations

Does the notetaker integrate with the compliance archiving systems (Global Relay, Smarsh) the firm already uses for other communications, so deal notes fall under the same review and audit process as email and chat?

How financial firms typically deploy AI notetakers on deal calls

Deal teams that have worked through this evaluation tend to land on a consistent pattern:

• Botless or bot-based capture depending on the call. External diligence calls with a target's management often prefer no visible bot joining as a participant, while internal IC discussions may use either method depending on firm policy.

• Zero-day retention for the most sensitive calls. Raw audio and transcript deleted immediately after AI processing, with only the summary and action items retained, so there's no window where raw MNPI content sits on a vendor's servers.

• Compliance review built into the workflow, not bolted on after the fact. A workspace-wide way to search, flag, and audit meetings, with AI trackers surfacing potential MNPI or PII mentions for a compliance officer to review, creates a defensible trail if a regulator or LP ever asks how the firm handles sensitive conversations.

• Archiving integrations wired into existing compliance infrastructure, so meeting records land in the same review pipeline as email and messaging rather than becoming a system compliance has to check separately.

Where Fellow fits

Fellow is the AI meeting assistant and notetaker built with a strict zero-training policy: Fellow never trains AI models on customer notes, transcripts, audio, or summaries, on any plan, with nothing to opt out of because customer content is never eligible for training, anonymized or otherwise.

For deal teams specifically, that policy pairs with a set of controls built for regulated, MNPI-sensitive work: configurable data retention including a zero-day option, in-meeting pause and resume so a conversation can drop out of the transcript the moment it turns sensitive, post-meeting redaction, and a compliance portal that lets a compliance officer search, flag, and audit any recorded meeting across the workspace, with AI trackers surfacing potential MNPI concerns and a fully logged review trail.

Fellow also integrates with Global Relay archiving, so deal call records land in the same audit pipeline as the rest of a firm's regulated communications. Fellow is SOC 2 Type II, GDPR, and HIPAA compliant, and firms in private equity use it specifically because deal notes need this level of control by default, not as an add-on.

This table maps directly to the practical controls the July 2026 Skadden client memo recommends firms consider when giving AI tools access to nonpublic information:

Conclusion

Not every AI meeting notetaker handles deal notes the same way, and the gap between "trains on anonymized data by default with an opt-out" and "never trains, nothing to opt out of" is exactly the kind of distinction regulators are now asking firms to account for. A firm that checks a vendor's training default in writing, confirms zero-day retention and granular access controls, and verifies archiving integration with existing compliance infrastructure has done the diligence the tool itself deserves. Fellow was built around a zero-training default and compliance controls designed for exactly this kind of regulated deal work.