How we protect a company's history when training AI

When a company licenses its history to train AI, the first question is about trust, not price. Here's exactly what happens to the data, and why we treat anonymization as a matter of detail.

TL;DR

Anonymization goes far beyond names, and every dataset is checked by people.
The raw copy is deleted after quality checks, and nothing is released until the company has approved the result.

The question every founder asks

Who will see years of Slack threads, emails and tickets, and what happens to them afterwards? We've worked with more than 30 companies, with teams of 50 to 500 people, and we hear that question in every first call.

Our answer: the raw data never leaves Ooak. Three values guide how we make that true:

Confidentiality

A license, not a sale. Nothing is shared that wasn't agreed.

Quality

Every dataset is reviewed by people, down to the details.

Customer care

One dedicated point of contact, from first call to delivery.

Custody: the raw copy stays in one place

Most of the risk sits in the raw copy, so we keep it in one place.

The company

SlackEmailDriveJiraGitHub

Ooak infrastructure · raw data never leaves this box

Encrypted storagescoped extract Anonymizationin-house pipeline Human QAreview + fixes

✕ Raw copy deleted after QA

Release

Company review ✓Anonymized dataset

Figure 1. Chain of custody. Only the anonymized dataset leaves Ooak, and only after the company has approved it.

We agree on the scope together, extract only that scope with read-only access, anonymize it on our own infrastructure, and have our team review the output. The raw copy is then deleted, and the company audits the anonymized dataset before anything is released. The whole process takes about 3 weeks.

Anonymization is in the details

We don't black things out: a redacted company is useless for training. We replace identities with consistent stand-ins and keep everything else, so a decision can still be followed from one tool to the next.

One decision, three tools

Slack · #ops

@julie.martin: the Acme renewal slips to the 14th, Mark is ok with it

From: julie.martin@ourco.com
Hi Tom, can you resend the latest invoice to Acme?

Jira · OPS-212

Renewal: Acme · due 14 · assignee J. Martin · reporter Mark Lee

Figure 2. Toggle the view. Identities change; the decision, the dates and the links between tools stay intact.

Replacing names is the easy part. The same person can appear ten different ways, and people can be recognized without their name at all. That's where most of our effort goes.

Where it hides	Original	Replaced with
Chat message	Jules	Clo
Email signature	Julie Martin · Head of Ops	Clara Moreau · Head of Ops
File name	Julie_notes_v3.docx	Clara_notes_v3.docx
Document author (metadata)	jmartin	cmoreau
Git commit author	julie.m@ourco.com	clara.m@company-a.example
Indirect identifier	"our only CFO in Chicago"	"our only CFO in Denver"

Figure 3. Identifiers hide well beyond the message body. Each one is found and replaced with the same stand-in, in every tool.

Under the hood

Parse every source while keeping its structure, including file metadata and commit history.
Detect identifiers with patterns, entity models and context-aware models for indirect identifiers.
Resolve every mention of the same person, client or project into one entity, across all tools.
Replace each entity with one stand-in, keeping formats intact.
Verify with automated leak scans, then human review.

Controls and the company's rights

Read-only access, isolated extraction jobs, encryption in transit and at rest, and time-stamped audit logs.
EU option: data processed and stored entirely in the EU, in line with GDPR.
Scope, duration and revocation are set in the license, and the company keeps ownership.
Sources can be excluded up front, such as private messages, customer databases or documents under NDA.

Hard questions to ask anyone who wants a company's history

Whoever the partner is, these are the questions worth asking. Here are our answers.

Who processes the raw data, and is any of it subcontracted?

Our own team, on our own infrastructure. No subcontractor ever touches raw content.

When is the raw copy deleted, and what happens to the mapping table?

Both are deleted once quality checks are done.

Does anonymization cover more than names?

Yes: nicknames and variants, signatures, file metadata, URLs, code and commit authors, attachments and indirect identifiers.

Is the output checked by people, or only by automated tools?

Both. Automated leak scans first, then human review of every dataset.

Does the company review and approve the output before it's used?

Always. Nothing is released without the company's approval.

Who is the contact if something comes up?

One dedicated point of contact, from the first call to delivery.