How we use AI.

1. Scope and effective date

This policy describes how Bambu Capital uses artificial intelligence across the systems we operate: our investor data room, our internal investment platform, and this website. It covers what our systems send to AI providers, what those providers are permitted to do with it, how long it is kept, and where a person is required.

It does not cover software our counterparties choose to run, and it does not cover AI features inside third-party tools we use for ordinary business purposes except where we name them here.

Effective date: 27 August 2026. Version 1.0.

We wrote this because we use AI heavily and we would rather say exactly how than let an investor discover it. If you are reading this alongside our founder's essay on the subject, this document is the precise version and it governs.

2. The decision rule

People decide. AI informs.

That sentence is common enough in our industry to be worthless on its own, so here is the mechanism that makes it enforceable rather than aspirational.

• An investment decision is recorded only as a ballot cast by a named partner, and every ballot is stamped with the kind of voter who cast it. Our systems record no non-human ballot.
• If a ballot on a decision gate were ever recorded as anything other than a person, the system refuses to settle that gate and nothing is written. The refusal is code, not procedure.
• A gate cannot be settled without the firm's own reasoning, in the firm's words, on the record. A tally alone will not settle it.
• Where a partner's reasoning was drafted from a meeting recording rather than typed, the record says so. A drafted reason and a typed reason are not stored as the same kind of thing.
• Approving drafted reasoning does not cast a vote. If a partner approves the words without choosing a vote, the system refuses.
• Every settled decision freezes the tally it was decided on, the committee that took it, and the state of the record at that moment. A decision that is later reopened is superseded and marked as such. It is not deleted.

3. What our AI is not permitted to do

• It does not vote, and it holds no decision authority of any kind.
• It does not write a change to any record on its own. It proposes a change, shows the exact before and after, and the change is written only after a person confirms it, through the same save path that person would have used by hand. If the underlying record moves between the proposal and the confirmation, the change is refused rather than applied to a stale picture.
• Every change it makes is stored with the value that was there before, so it can be put back.
• It grants no authority. Our assistant can do conversationally only what the same person could already do by clicking. It is gated by the same permission check as the screen itself.
• Our IC memo drafting assembles a draft from what is already on the record, watermarks it as a draft, files nothing, and writes no data. A partner edits it and attaches the final memo.
• Where a document's stored summary no longer matches the file we hold, it is withheld and the withholding is named, rather than presented as current.

4. What our AI does automatically

This section exists because section 3 would be misleading without it.

Some of our AI processing runs without a person triggering it:

• Documents are read when they arrive. Every document attached to a deal in our platform is sent to our AI provider in full and summarised automatically at upload. For PDFs this is the file itself; for other formats it is the extracted text.
• The data room search index rebuilds on a schedule, which sends portions of every indexed document to our infrastructure provider's AI service every few minutes.
• Room summaries refresh on a schedule, which includes core fund documents.
• Our data room assistant answers questions directly. When you ask it something, it answers you in the moment. A partner does not review that answer before you see it.

We record every one of those assistant exchanges and review them. That is a control after the fact, not before it, and we would rather say so than claim an approval step we do not have.

5. Who receives your information

We use two AI providers directly, and no others:

Anthropic (Claude), through Anthropic's commercial API. Receives, depending on the feature:

• Full document content: documents attached to a deal, uploaded meeting summaries, our fund raise deck, verbatim excerpts from data room documents when answering questions about them, and full meeting transcripts when drafting reasoning for a decision gate.
• Record data and cached summaries, without document text: deal records, sourced facts, screen verdicts, and the cached summaries described above. Our pre-mortem runs entirely in this mode and performs no document text extraction.
• Narrow public metadata only: for open-web research on a target company, its name and public domain, subject to the withheld list in section 8.

Cloudflare, through Workers AI on the platform our systems run on. Receives:

• Chunks of every indexed data room document, for the search index.
• Investor letter and portfolio-company update extraction.
• Certain assistant responses, as the primary path in some features and the fallback in others.

Both providers therefore receive data room content, for different purposes: Anthropic receives verbatim excerpts when the assistant answers a question about a document, and Cloudflare receives chunks of every indexed document to build the search behind it.

We also use Fellow, a third-party meeting notetaker, which holds the recording and transcript of meetings we record with it; transcript content from Fellow is sent to Anthropic when we draft reasoning for a decision gate. Fellow's published sub-processor list names Anthropic, OpenAI, Google, Groq and Together.ai for its own AI features, and AssemblyAI and Recall AI for recording and transcription, so those vendors can be indirect recipients of meeting content held by Fellow, on Fellow's systems rather than ours; Fellow does not publish which vendor handles which feature, so we cannot tell you which of them has actually processed our meetings. Fellow states that customer data is not used to train AI models, but that statement sits on its sub-processor and product pages rather than in its privacy policy or terms, and we do not hold a signed data processing agreement with Fellow that says so; Fellow's default is also to keep recordings and transcripts indefinitely until an administrator sets a deletion schedule.

We use no other AI provider directly. Specifically, our own platform sends nothing to OpenAI, Google, Amazon Bedrock, Cohere, Mistral, or any other model vendor; the only route by which any of them can reach our content is the Fellow route described immediately above.

6. Training

No AI provider we use is permitted to train models on your information.

• Anthropic's Commercial Terms of Service state that Anthropic may not train models on customer content submitted through its commercial services. (Commercial Terms of Service, effective 17 June 2025.)
• Cloudflare states that it does not use customer content submitted to Workers AI to train any model, or to improve Cloudflare or third-party services. (Workers AI data usage documentation, updated 21 April 2026; Service-Specific Terms, updated 2 June 2026.)

Two clarifications we think matter:

We use commercial API endpoints only. We do not use consumer AI products for firm data. Consumer plans at these vendors have different defaults; those defaults do not apply to us because we do not use those products.

Nothing of yours is training any model of ours either. We are accumulating a record of our own investment decisions and the reasoning behind them, and we expect one day to use it to help our own systems reason about our own judgment. That record is our reasoning, not your documents. No model, ours or anyone else's, is being trained on your information today.

These are the terms that govern commercial API use, and our systems call the commercial API exclusively. We hold the executed agreements and our account records on file.

7. Retention

At our providers. Anthropic's standard retention for commercial API traffic is 30 days. There are carve-outs an investor is entitled to see in the same sentence: content flagged by trust-and-safety processes may be retained for up to two years, and safety classification scores for up to seven years. We do not currently hold a zero-data-retention arrangement.

At Bambu. We keep what our systems generate: document summaries, extracted document text and its search index, assistant questions and answers, and decision records. Today we keep these indefinitely, until we delete them on request. We are building a retention schedule and this section will state it when we have one.

8. Open-web research

Five of our research features let the model compose and run its own web searches. We constrain what we put into the prompt; we do not pre-approve each query the model then writes.

When we research a target company, an enforced list withholds our economics and our criteria: our EBITDA floors, enterprise-value bands, tolerances, operating-margin targets, platform slots, criteria sets, team assessments, and the firm's own name. A scanner refuses the request if any of them appear. This applies to our sourcing, company-research and comparables features.

9. The data room, specifically

If you are an investor or a counterparty in our data room, this is the section that concerns you.

• Document content reaches an AI provider. When you ask our assistant about a document, verbatim excerpts of it are sent to a model. Separately, portions of every indexed document are sent to our infrastructure provider to build the search index.
• Your questions are recorded. Every question asked of the assistant, and every answer given, is stored with your email address, your organisation and your role, and is readable by the Bambu team. We did this deliberately: we are responsible for what our assistant tells you, and we cannot be responsible for something we cannot see.
• Answers are not reviewed before you receive them.
• Document opens and downloads are audited, as they are in any data room.

10. What we do not do

Each of these is verified in our own source code, not merely intended:

• We send no end-user identifier to any AI provider. No provider we use can attribute a request to a particular person at Bambu or at a portfolio company.
• We route no AI request through any third-party gateway or prompt-logging proxy.
• We use no third-party analytics, error-tracking or session-recording service anywhere in our platform.
• We write no prompt, document text or model output into our application logs.
• We use no consumer AI product for firm data.

11. Your rights, and how to reach us

You may ask what we hold about you, ask us to delete it, or ask how a particular output was produced. Requests are handled by a person, not a form. Write to [email protected].

12. How this document is kept honest

This policy was written against our source code, and every factual claim in it was verified against that code or against a provider's contract on the date above.

Code changes. This document is re-verified against the platform quarterly, and on any material change to how we use AI. When we find a divergence we correct the document and note what changed.

We hold ourselves to one rule in particular: if this document says a control exists, that control is enforced in software, not in a habit.