EHR-derived feed · Approved by Halvane’s privacy officer, Mon Oct 12, 2026
Store gate
Mask policy US-COM v14, store gate, one EHR-derived record
Field
What Ontelis holds
Store treatment
Class
Patient name
Not ingested
Do not ingest
Patient
Patient ID
P-7f3a
Pseudonymize
Patient
Birth date
1961
Generalize to year
Patient
ZIP
021
Generalize to three digits
Patient
Diagnosis
The ICD-10 code
Keep as code
Patient
Prescriber NPI
The NPI, refreshed by sync
Keep
Licensed third-party, HCP personal
Fill date
2026-W14
Week only, under Halvane’s Expert Determination
Patient
Safe Harbor alone would allow only the year for the fill date. Three-digit ZIP is kept only where that area has more than 20,000 people, otherwise 000.
One record from an EHR-derived feed, after the store gate. HCP is a healthcare professional; NPI is the National Provider Identifier.
Trusted by 2 out of the 5 largest life sciences companies
Every field is classed before it is stored.
Mask reads the labels you already keep in Purview, Collibra, BigID or Snowflake, and proposes a class for the rest. A data steward confirms each one. New fields stay out until classed.
Decide what Ontelis holds, per field.
Keep a value, point to it, swap it for a token, generalize it or drop it. Tokens are looked up in your own vault, never made by hashing a name.
Each reader sees only what their purpose allows.
Every request declares a purpose. The same record reaches an analyst with patient detail, a rep’s agent as counts only and Copilot not at all.
Masked data stays masked wherever it goes.
Treatments follow each value into context packages, decision records, the ledger and model prompts. Patient data goes only to endpoints inside your boundary, and no model is trained on it.
Named people change each rule.
Tightening takes one signature. Loosening patient data needs the privacy officer and a passing re-identification check. Ontelis staff can propose, never approve.
Prescribing data stays out of anything of value.
Speaker picks, consulting, advisory boards and grants cannot use prescribing-derived fields, and a lineage check proves it. Linkage and re-identification blocks hold even if you ask to lift them.
“What won over our compliance team wasn’t the speed. It was that the agent shows its work, stops at Draft, and tells you when it doesn’t know.”
Head of Marketing OperationsOne of the top five US biopharma companies
What Mask does not do
Certify HIPAA de-identificationYour privacy office or an outside expert decides; Ontelis stores and enforces the decision.
Capture consentThat stays in OneTrust or the CRM.
Act as company-wide data loss preventionMask controls what Ontelis stores, serves and sends.
Write your privacy policyYour privacy office sets the rules a Mask policy enforces.
Where Mask fits
Mask shapes each context package and sets which models an agent draft may use.
Ontelis Mask controls, field by field, what Ontelis stores, what each person or agent sees and what may be sent to an AI model or an outside agent. It's how data masking and pseudonymization work in Ontelis: a patient ID can be held as a random token, a birth date as a year and a patient name not at all. A data steward confirms each field's class, and named people at your company approve every rule change.
No, you keep them. Ontelis Mask starts from the labels you already hold in Microsoft Purview, Collibra, BigID or Snowflake tags, and proposes a class only for fields that have none, for a data steward to confirm. Mask's scope is Ontelis: what Ontelis stores, what it shows each person or agent, and what it sends to a model or an outside agent. It isn't company-wide data loss prevention.
No, masked isn't anonymized. In Ontelis Mask it's one way a field is shown to a reader, such as •••• in place of a prescriber's NPI. Most of what Mask does, like swapping a patient ID for a token, is pseudonymization, and pseudonymized data is still personal data under GDPR. Ontelis calls data anonymized in four cases only: it was dropped, it's held as metadata only, it's served as aggregates with small counts suppressed, or it's covered by a recorded Safe Harbor attestation or Expert Determination.
No. Whether a dataset counts as de-identified under HIPAA is your privacy office's call, or an outside expert's, by Safe Harbor or Expert Determination. Ontelis Mask records that decision and enforces it field by field. If your Expert Determination allows fill dates by week, for example, Mask holds the week and nothing finer. Mask also blocks any join between de-identified and identified data unless a recorded Expert Determination covers that exact join.
By default, patient data reaches only AI models inside your own boundary. Ontelis Mask sends patient-class values only to approved endpoints there, and the one exception is a zero-retention endpoint under a BAA that your privacy officer approves. Each request states its purpose first, so an agent gets only the fields that purpose allows. An outside agent such as Copilot, coming through the Agent Gateway, receives only the data classes your policy allows for that destination.
Yes, but only for three listed reasons: safety follow-up, the patient's own request or a legal hold. Reversing a Mask token takes two approvers, and the reason is recorded. The tokens are random, never a hash of a name and birth date that someone else could recompute, and the lookup table sits in your own vault under a separate key role. The people who operate Ontelis can't reverse a token.
A wrong class is caught by sampled audits of classified fields and corrected by your data steward. In Ontelis Mask that correction is a versioned change: you see who gains or loses access before it applies, and the stored data is reprocessed to match. The default limits the damage too. A column that's added or renamed after go-live is held as metadata only and stays out of what agents receive until a steward classes it.
The pseudonymization lookup table sits in the customer’s vault under a separate key role, so the people who operate Ontelis cannot reverse tokens.
Vendor documents
Ontelis provides: a SOC 2 Type II report, an ISO 27001 certificate, a HIPAA business associate agreement wherever Ontelis processes PHI, a data processing agreement with transfer clauses, and a sub-processor list.
Models
Ontelis builds no foundation model. No model is trained on customer data, except an opt-in adapter that never leaves the customer’s data plane.
Your controls
Ontelis’s posture as a vendor is kept separate from Govern, which holds your own controls.