AI build policy

Your information is protected before AI ever sees it. This page explains how NowPatient builds and governs AI, in plain language first, with the detail for healthcare organisations and procurement teams further down.

• A governed layer, not shortcuts. AI in NowPatient runs through a purpose-built harness. We never paste patient data into public chat tools or wire model calls directly into application code.
• Privacy first. Identifiable details are de-identified or redacted before they reach an AI model.
• No training on your data without your permission. Our contracts prohibit third-party AI vendors from training their models on your data, and we only train or develop our own AI models with your health data if you have opted in, or if the data has been anonymised.
• Everything is logged and checked. Every AI request and response is recorded in an audit trail, and outputs are checked against clinical-safety and content guardrails.
• People stay in the loop. Human oversight is built in where it matters most.

The harness is a single governed layer that sits between NowPatient products and the underlying frontier AI models. Every AI feature we build uses it. That means access, data handling, safety checks and monitoring are managed in one place, rather than reinvented, and potentially weakened, in each product.

• De-identification before processing. Protected health information (PHI) and patient-identifiable data are de-identified or redacted before they reach a model.
• Minimum necessary. AI features receive only the information they need to do their job.
• Encryption. Data is encrypted in transit and at rest.
• Jurisdiction. Processing is held within agreed jurisdictions: primary UK hosting (AWS London / eu-west-2) for clinical and application workloads; some supporting infrastructure also in other AWS regions including the US; HIPAA-aligned controls and clinician BAAs where US clinicians onboard.
• No vendor training. Contractual terms prohibit third-party model providers from training on your data or keeping it after processing. Where PHI is processed by a provider, we maintain Business Associate Agreements / zero-retention arrangements.

Infohealth develops some AI models of its own. When we do this:

• Permission first. We use identifiable or coded health data to train or test our own models only with your explicit, opt-in consent, given separately from your care and from accepting our policies. Saying no never affects your care.
• Coded, not named. Training data is held under a code. Names, contact details, NHS numbers and free-text notes are removed, and the key linking codes to people is held separately by a small number of named staff.
• Secure environment. Researchers, including university partners, work inside our own secure environment, which lets analysis results and model files out but not row-level data.
• Release protocol. Before any coded data is shared outside Infohealth, it passes a release protocol (for example age bands, date shifting and suppression of rare values) and is shared only under a contract that bans re-identification and onward sharing.
• Fairness. Where you choose to tell us, we record characteristics such as skin tone so we can test and report whether a model works equally well for everyone.
• Assessment first. We complete a data protection impact assessment and, for research, obtain research ethics approval before any training or research use of personal data starts.
• Medical device rules. Any model intended for a medical purpose is developed under clinical risk management and is placed on the market only after it meets the UK medical device regulations and any other rules that apply where it is sold.
• Withdrawal. You can withdraw at any time. We remove your data from active datasets and future sharing, but cannot remove it from a model already trained or from results already published.

Access to systems, prompts and data is role-based and follows the minimum-necessary principle. Permissions are managed centrally and reviewed at least every six months.

Every AI request and response is recorded in a tamper-evident audit trail. This supports HIPAA audit-control expectations and NHS expectations for transparency and accountability, and allows us to investigate and respond if something goes wrong.

• Outputs are validated against clinical-safety and content guardrails before they reach you.
• Human oversight is built into workflows where AI output could affect care or decisions.
• AI features are tested before release and monitored after it. Models are evaluated before being adopted or changed, including on data that was not used to train them.

What our AI does, and does not do. NowPatient uses AI to support clinical care, not to replace it. AI does not diagnose, prescribe or make treatment decisions, and no AI output is acted on or shared with another healthcare professional until a registered clinician has reviewed it. Responsibility for every clinical decision remains with the clinician. We do not use AI to make automated decisions that have legal or similarly significant effects on patients. We tell patients when AI has been used in their care.

Because models, prompts and permissions are managed centrally, we can evaluate, update or swap a model, apply policy consistently across every product, and respond quickly to emerging risks without rebuilding each product.

NowPatient’s AI is designed around the following.

HIPAA does not offer a formal certification, and DTAC is an assessment rather than a badge. We describe our alignment accurately and are glad to share supporting documentation under NDA.

If you have questions about how AI is used in NowPatient, want to see supporting documentation, or wish to raise a concern, contact us:

• To withdraw consent to research or model training: your account settings, or email the Data Protection Officer.
• Email: dpo@nowpatient.com
• Data Protection Officer: Amish Patel BPharm (Data Protection Officer / Compliance)

We review this policy every 6 months and whenever our AI practices materially change.