How Sensitive Business Data Is Transformed for AI Training: What Changes and What Stays the Same
What happens to sensitive data when a company licenses it for AI?
Your business records may contain both useful operational knowledge and details that identify people. A privacy review helps determine what can be retained, what must change, and what should be excluded before a dataset is used.
This guide uses a simple shape example, micro1’s published approach, and a healthcare example from Tempus AI to explain the process.
Watch the whiteboard explanation
Changing a name is one part of the work
De-identification describes efforts to remove or reduce links to real people. Replacing a name with a fictional identity can help, but other details may still identify someone. A unique job, location, event, or combination of facts can matter too.
“Anonymized” should not be treated as a promise that changing names makes any dataset safe to share. The UK Information Commissioner distinguishes anonymization from pseudonymization: coded or replaced identities can still be personal data where re-identification remains possible. The applicable legal standard depends on the setting.[1]
The visual: what changes and what stays
Imagine the same person appears in an email, a support ticket, and a report. In this simplified example, every orange circle A becomes the same purple diamond X. The squares and triangles represent approved information that can remain useful.
Consistency helps preserve the connection between records. A request, the follow-up work, and the result can still form a coherent story. This is one of the goals described in micro1’s research on transforming identities across an entire collection of files.[2]
Illustrative original
Jane Smith reported a duplicate charge. The team checked the invoice, approved a refund, and confirmed the resolution.
Illustrative transformed version
Alex Taylor reported a duplicate charge. The team checked the invoice, approved a refund, and confirmed the resolution.
This short example changes only the name to make the idea visible. A real review would also consider account identifiers, attachments, contact details, and revealing context.
How micro1 describes handling company data
micro1 says companies can transfer original data through approved methods without first anonymizing it themselves. That means micro1 may receive sensitive originals. Its published partnership materials describe these controls:[3]
- 01 · CONTROL ACCESS
Keep processing restricted
Access is limited to authorized personnel, with company datasets kept separate during processing.
- 02 · TRANSFORM & REVIEW
Prepare data for use
Identifying information is removed or transformed before third-party sharing. Companies can review representative samples before use.
- 03 · FOLLOW THE TERMS
Apply retention rules
Raw-data deletion follows the agreed schedule. Deletion requests are handled under the engagement’s terms.
Its research describes automated detection, consistent identity replacement, and validation, supported by human review of samples. It reports imperfect benchmark results; those results should not be presented as a guarantee that every sensitive detail will be caught.[2]
Put the specific safeguards in the agreement. Confirm who receives originals, which content is excluded, and what review is available for your dataset. A provider’s general description is the starting point for that discussion.
Tempus AI licenses de-identified data
Tempus shows how data licensing also operates in healthcare. On April 23, 2025, it announced multi-year collaborations with AstraZeneca and Pathos using de-identified oncology data to develop a multimodal foundation model. The agreements included $200 million in data licensing and model-development fees—a combined amount covering both activities.[4]
Tempus’s technical description says data goes through de-identification before research use or third-party research licensing. It specifically identifies HIPAA Expert Determination in that pipeline and explains the importance of preserving connections across a patient’s history.[5]
What does that process involve?
Tempus’s Policy Center describes two methods: Safe Harbor and Expert Determination. These are alternative approaches; this does not mean every dataset goes through both.[6]
- Safe Harbor: remove the specified categories of identifiers, subject to the rule’s exceptions, and have no actual knowledge that remaining information could identify someone. HHS specifies 18 categories.[7]
- Expert Determination: a qualified expert assesses and documents whether identification risk is very small, considering the likely recipient and other reasonably available information.[7]
Tempus says its techniques include removing or masking identifiers, suppressing rare values, grouping details such as ages into ranges, and adding statistical noise where appropriate. It also describes ongoing reassessment, encryption, restricted access, and terms prohibiting re-identification.[6]
Relate this back to the shapes: some squares and triangles may need to change too. Useful medical relationships can remain, while specific details are generalized or removed. HHS notes that de-identification still carries a small, nonzero identification risk.[7]
Tempus is a separate healthcare example, not a description of micro1’s methods or accepted datasets. HIPAA’s framework applies to covered health-information settings; it is not a universal standard for ordinary business records. This example also is not a valuation benchmark for your company’s data.
Bring privacy into the agreement discussion
Use these questions to make the proposed arrangement concrete. Ask for answers that distinguish original files, transformed datasets, and any models or other outputs created from them.
- Scope: Which records are permitted, and which client materials, trade secrets, credentials, or other content must be excluded?
- Original access: Who can receive or inspect sensitive originals, including subcontractors, and where will processing occur?
- Transformation and review: What standard is used, who checks the output, and can your company review samples or request corrections?
- Permitted uses: Who may receive the prepared data, for which purposes, and with what limits on sublicensing or re-identification?
- Retention and responsibility: What gets deleted, when, and how is deletion confirmed? What happens if sensitive information is missed or an incident occurs?
Before you sign: rights, rewards, and risks
Privacy belongs alongside payment, permitted uses, exclusivity, and responsibility. Our negotiation guide organizes the conversation around what you give, what you get, and what obligations remain.
Read the agreement negotiation guideExplore your company’s data licensing opportunity
Start by describing the types of records you have, the work they document, and the sensitivity issues that need review. A data manifest can help organize that first conversation without including the underlying records.
Sources and further reading
- UK ICO: Introduction to anonymisation — identification risk and the distinction from pseudonymization.
- micro1: PII Transformation for Enterprise Datasets — published transformation goals, evaluation, and review methods.
- micro1: Company Data Partnerships, privacy and handling FAQ, reviewed November 4, 2026. Source: www.micro1.ai/data-partnerships. For the opportunity and referral route, use the DataHasValue Data Licensing page.
- Tempus: AstraZeneca and Pathos agreements — April 23, 2025 announcement.
- Tempus: The Tempus data pipeline — May 27, 2026 technical overview.
- Tempus Policy Center — de-identification methods and stated privacy controls.
- HHS: Guidance on de-identification under HIPAA — Safe Harbor, Expert Determination, and residual risk.
Provider descriptions reflect their published materials, not an independent audit. Educational information, not a privacy certification or legal advice. DataHasValue may earn referral compensation from qualifying partnerships reached through its referral links.
