# How your company's data is de-identified before an AI company sees it Source: https://workflowdividend.com/de-identification.html Last reviewed: 2026-10-23 Author: Sean Key, Applied Clinical Judgement (refers companies to micro1 and may receive referral fees) You're probably wondering what happens to your customers' and employees' names if you license your company's data. This guide covers how they're removed, why the useful part survives, where the risks are, and what to ask before you sign. It's written for you and for the lawyer or privacy officer you'll pass it to. ## Key points - AI models are being trained to do real work, so they need to see how work happens across your systems. - Names and contact details are swapped for consistent stand-ins, such as made-up names. The links between records survive; the identities don't. - No method makes re-identification impossible. Good programs blur revealing details, scrub free text and ban re-identification by contract. - micro1 doesn't publish the full method behind its Company Data Partnerships program, but it has published test results for its tool. The nine questions at the end of this guide will get the rest in writing. ## What training an AI model involves Think about how a new hire learns a job. The manual helps on day one, but they really learn by working through real cases: what came in, what the team discussed, what was decided and how it turned out. AI models are now trained the same way. The first generation learned language from public websites. The models being built now learn how skilled people handle tasks like a claim or a contract revision, and that takes worked examples of real business tasks. Those examples sit inside your company. That's why AI companies will pay for controlled access to them. ## Why a linked case is worth more than one document A policy document gives a model the rule. It can't show judgment. What's valuable is the chain of events: the request, the discussion, the decision, the record and the result, spread across email, chat, tickets and your CRM. *Text version of this diagram:* **On its own:** a refund policy, saved as a PDF in a shared drive. It shows the rule, but nothing about how anyone applied it. **Linked across your systems:** 1. Email: a customer asks for a refund outside policy. 2. Slack: the team debates whether it's a fair exception. 3. Ticket: a manager approves it and notes why. 4. CRM: the refund is logged and the account flagged. 5. Outcome: the customer renews next quarter. The linked version shows the whole job: the request, the judgment call, the action and the result. That's what AI models learning to do real work are missing. Diagram 1. The same refund as a single document and as a linked case. ## How names are replaced while records stay linked Deleting every name would break the chain. Nobody could tell that the customer in the email is the one discussed in Slack and logged in the CRM. So the standard technique is consistent replacement. Each real person or company gets a stand-in, and the same person gets the same stand-in in every system. The stand-in can be a label like [Customer_A]. micro1 goes further and gives each person a realistic made-up identity: a name, email address and username. That has a side benefit. If the software misses a real name, it sits among thousands of invented ones, so nobody can tell it's real. *Text version of this diagram:* An illustrative example; every name here is invented. Each real person gets one made-up identity, used in every system. | System | Before: in your systems | After: what the AI lab receives | | --- | --- | --- | | Email | From: Jane Ruiz, Ruiz Family Dental. "Can I get a refund on my March invoice?" | From: Dana Whitfield, Cedar Lane Dental. "Can I get a refund on my March invoice?" | | Slack | "@mike can we refund Jane? She's been with us six years." | "@tom can we refund Dana? She's been with us six years." | | CRM | Jane Ruiz · refund $240 · approved by Mike Chen | Dana Whitfield · refund $240 · approved by Tom Becker | Because Dana Whitfield always stands for Jane Ruiz, the chain from request to decision survives. That chain is what makes the data valuable. And if a real name slips through, it's hidden among thousands of invented ones. Diagram 2. Consistent stand-ins keep records linked while removing who they're about. ### Anonymized, de-identified and pseudonymized: what each term means - **Pseudonymized:** Names are swapped for tags, and a separate, protected key could link them back. California's privacy law defines it in those terms: data that can't be tied to a person without extra information kept separately. - **De-identified:** A legal standard as well as a technique. Under California law the data mustn't reasonably link to a person, and whoever holds it must also commit publicly and by contract not to re-identify it. - **Anonymized:** Everyday shorthand for "can't be traced back". US law doesn't define it, so ask any company that uses the word what it means by it. ### What travels, what stays behind and what the model keeps Here is the whole process in one picture. The swap key is what makes your data pseudonymized rather than anonymized, so who holds it matters. The AI lab gets only the stand-in version. Training then keeps the working patterns, not the people. *Text version of this diagram:* 1. **Your documents**: email, Slack and CRM records mention real people, for example Jane Ruiz asking for a refund and Mike Chen approving it. 2. **The swap key**: each real person gets one stand-in, used everywhere. Jane Ruiz becomes Dana Whitfield and Mike Chen becomes Tom Becker. The key is the only link back to real people. It stays behind and is never sent to the AI lab. 3. **What the AI lab receives**: the same records with only the stand-in names, so the chain from request to decision still makes sense. 4. **What the model learns**: patterns, not people. For example, how a refund request is handled, who signs it off and how a long-standing customer is treated. Models can occasionally memorize rare text word for word, which is why names are replaced before anything reaches the AI lab. 5. **The license fee** comes back to your company, paid through micro1. Every name here is invented. Ask micro1 who holds the swap key, where it is stored and when it is destroyed. Get the answers in writing before you agree what is in scope. Diagram 3. A solid shape is a real person; an outlined shape is their stand-in. ## Where your data goes, step by step The order matters. You decide what's in scope. Software removes identifiers. You review what will be released. Only then does a de-identified dataset reach an AI lab. *Text version of this diagram:* 1. **Your systems** (you decide): Slack, email, CRM and documents. Nothing has moved yet. 2. **You choose the scope** (you decide): which tools, teams and date ranges are included. 3. **Automated extraction and de-identification** (platform step): identifiers are removed or replaced with tags. 4. **You review and approve** (you decide): you check samples before anything is released. 5. **De-identified dataset** (platform step): your workflows, without the identities. 6. **An AI lab trains its model** (platform step): it learns how the work is done, not who did it. micro1 says it agrees scope and security with you first, removes personal data where relevant, processes it in isolated pipelines and lets you review samples before use. Ask micro1 for its exact process in writing. Diagram 4. The steps outlined in rust are your decisions. ### micro1's description of its process Scope, security and anonymization are agreed before you start. Personal data is removed where relevant, some data is rewritten synthetically, processing runs in isolated pipelines, and you can review samples. Its September 2026 research paper describes realistic stand-in identities kept consistent across every file, with people reviewing samples of the output. Its June 2026 webinar says data is deleted 60 days after transfer by default. ## How de-identified data can be traced back to people The links that make your data valuable could also, in principle, help someone work out who a record is about. Removing a name doesn't help much if a job title, a town and a date still point to one person. *Text version of this diagram:* Each detail is harmless on its own: - A job title: "owns a dental practice" - A place: "Boise, Idaho" - A date and an event: "refund dispute, March 2025" Put together, they can point to one identifiable person, even with the name removed. Good de-identification blurs these details too, for example "a healthcare business in the western US" and "early 2025". Free text is scrubbed for names, signatures and phone numbers, and contracts bar recipients from trying to re-identify anyone. Ask each program which of these it does. Diagram 5. Small details combine. Good de-identification is designed around this risk. > "Just scrubbing is not enough." Grant Gooley, Founder & CEO of My Portal Marketing, which licensed its data to micro1, in an interview on micro1's channel ([watch it on my micro1 review](https://workflowdividend.com/micro1-data-partnerships-legit.md#video-gooley)). His team left client data and finances out of scope. That's why serious de-identification works in layers: 1. **Leave things out.** You set the scope. Client records, finances or a whole system can stay out of the license. 2. **Remove or tag direct identifiers:** names, emails, phone numbers, account numbers and addresses. 3. **Blur revealing details.** Exact dates become months, towns become regions, rare job titles become categories. 4. **Scrub free text,** such as email signatures, "say hi to Jane" and pasted phone numbers. It's the hardest layer, so press for detail here. 5. **Limit who sees what** through automated processing, isolated systems and no human review of raw data. 6. **Back it with contracts** that legally bar every recipient, the AI lab included, from trying to re-identify anyone. Even with all six layers the risk isn't zero, so be wary of anyone who claims it is. Together they make it very small, and "very small" is the standard US law sets. ## How well micro1's de-identification tool performs In September 2026, micro1's AI team published a research paper on its de-identification tool, flow-transform 1.0. No company in this market had published figures like these before, and they're worth reading closely. - **On a public test set,** the tool found about 95% of personal details such as names, emails and usernames. That's ahead of the rival tools micro1 tested it against. micro1 research paper, Sep 2026 (micro1.ai/research/pii-transformation-for-enterprise-datasets), checked Oct 23, 2026 - **On micro1's own harder test,** built from realistic but made-up business records, it caught about 89%, using an optional AI review step. Read the other way, roughly one detail in ten got through. - **People check samples of the output.** They also study failures and turn them into new test cases. - **The paper names what it can't yet do.** Layered design files such as Photoshop and Figma, and faces and voices in recordings, aren't solved. These are micro1's own results, mostly on made-up data, because real company data can't be published for testing. The paper also doesn't say how it handles revealing details like dates, places and job titles. So the earlier advice stands: leave sensitive material out of scope, review the samples you're offered, and get the contract terms in writing. Source: Mahajan, Maas, Zhou and Joshi, "PII Transformation for Enterprise Datasets", micro1 research (micro1.ai/research/pii-transformation-for-enterprise-datasets), September 24, 2026. ## What US privacy law requires The US has no single federal privacy law for business data. Which rules apply depends on your sector and where your customers and staff live. Three matter most. ### California (CCPA, as amended by the CPRA) Information only counts as "deidentified" if it can't reasonably be linked to a person and the business holding it does three things. It takes reasonable measures to keep it that way. It publicly commits not to re-identify it. And it contractually requires every recipient to do the same. Virginia, Colorado, Utah and other states use a similar three-part test. ### Healthcare (HIPAA) Health information can be de-identified in one of two ways. Safe Harbor removes 18 listed types of identifier, including names, small geographic areas, most dates, contact details, and record and account numbers, provided there's no actual knowledge that what remains could identify someone. Expert Determination needs a qualified statistician to document that the risk of re-identification is very small. ### Finance and legal Financial firms have duties under the Gramm-Leach-Bliley Act, and law firms must protect client confidentiality and privilege. If that's you, bring your counsel in early. This is a general explanation, not legal advice. I'm not a lawyer. Laws change, so ask your counsel to confirm what applies to you. ## If your company is in the UK UK companies are covered by the UK GDPR and the Data Protection Act 2018, as amended by the Data (Use and Access) Act 2025. Licensing a pseudonymized copy of your data still counts as processing personal data on your side, so these rules apply even though the AI lab never sees a name. ### A data protection impact assessment (DPIA) Almost certainly required. AI training is a new technology, the volumes are large, and the people in your emails and chat wouldn't expect this use. Complete it before anything is exported, with your data protection officer if you have one. ### A lawful basis and a compatibility check Legitimate interests is the likely lawful basis, backed by a written assessment. Since February 2026, Article 8A of the UK GDPR sets out when reusing data for a new purpose is compatible with why you collected it. Data you originally collected with consent is much harder to reuse. ### Telling people first Update your privacy notices before any data is reused, brief your staff, and give people a chance to object before export. Since June 2026 you also need a way for people to complain to you directly. ### Transfers to the US If data leaves the UK, it needs a recognized safeguard: the UK Extension to the EU-US Data Privacy Framework if the recipient is certified under it, or the UK's international data transfer agreement. Ask micro1 which applies, and whether raw data leaves the UK before names are replaced. If your data includes people in the EU, the EU GDPR may apply too. This is a general explanation, not legal advice. Ask your adviser to confirm what applies to you. ## Nine questions to ask before you sign Forward these to whoever has to approve the deal. A well-run program will answer every one in writing. 1. Which identifiers do you remove or tag, and how do you handle free text in emails and chat? 2. Do you replace names consistently? Who holds the key linking tags to real names, and is it ever destroyed? 3. How do you generalize dates, locations and job titles? 4. Does any person see our raw data: your staff, your contractors or the AI lab? 5. Will the contract bar the AI lab and every other recipient from attempting re-identification? 6. Can we review samples and remove items before anything is released? 7. How long do you keep the data, and how do we request deletion? 8. What share of personal details does your tool catch, and was it tested on data like ours? 9. For health data: which HIPAA method do you use, and who carries it out? [Download the questions as a one-page checklist (PDF)](https://workflowdividend.com/assets/nine-questions-checklist.pdf) ## More guides - [Why AI companies pay for business data What they're buying, and why everyday work is valuable](https://workflowdividend.com/why-ai-companies-pay-for-business-data.md) - [Is it a scam? The evidence the program is real, and six warning signs](https://workflowdividend.com/is-it-a-scam.md) - [micro1 review What micro1 offers, and the concerns to weigh](https://workflowdividend.com/micro1-data-partnerships-legit.md) - [What is your data worth? Published figures and what sets the price](https://workflowdividend.com/what-is-my-company-data-worth.md) - [If your company has closed Who qualifies, and who can still sign](https://workflowdividend.com/closed-company-data.md) **About the author** Sean Key has been a freelance consultant for 22 years, mainly delivering data and digital projects into the NHS. He runs Applied Clinical Judgement, which connects professionals with AI training work, and helps established companies weigh up micro1's Company Data Partnerships program. He isn't a lawyer and isn't employed by micro1. If your company signs up through his introduction, micro1 may pay him a referral fee at no cost to you. [Sean Key on LinkedIn](https://www.linkedin.com/in/seandkey/) · [Book a call](https://calendly.com/seandavidkey/the-workflow-dividend) ## Sources micro1, "Company Data Partnerships" (micro1.ai/data-partnerships), accessed October 23, 2026. micro1, "How to earn up to $2M training AI models with anonymized company data", YouTube webinar, June 25, 2026. micro1, "The hidden asset every business already owns", YouTube interview with Grant Gooley. micro1, "What every business owner should know about AI data licensing", YouTube interview with a micro1 referral partner. Mahajan, Maas, Zhou and Joshi (micro1), "PII Transformation for Enterprise Datasets", micro1 research, September 24, 2026. California Civil Code § 1798.140(m) "Deidentified" and (aa) "Pseudonymize", 2025 code, accessed October 23, 2026. US Department of Health and Human Services, "Guidance Regarding Methods for De-identification of Protected Health Information", accessed October 23, 2026.