Skip to content

Security & trust

Patient data, and the boundary nobody crosses.

Accrual only works if a site's governance review says yes. This page is written for that review — including the parts we have not finished, which are stated rather than omitted.

The boundary

Sponsors never receive patient data. Not anonymised records — none.

This is the question every site asks first, so it is answered first.

The site controls

The site — an NHS trust, health board or independent provider — is the controller. Accrual is a processor acting on the site's documented instructions, under an executed Article 28 agreement, for one named purpose.

Sponsors get aggregates

Counts, distributions and forecasts. No names, no identifiers, no individual-level data of any kind. Counts small enough to identify someone are suppressed, not reported.

Patients hear from their own team

Accrual does not contact patients and provides no mechanism for a sponsor to. The care team at the site is the only party that reaches out.

Controls in place

What is implemented today.

Access

Role-based access control, least privilege by default
Unique user identification, no shared production credentials
Multi-factor authentication required for all personnel
Access review on a fixed cycle and on every role change
Background checks for anyone with production access

Data

Encryption in transit (TLS 1.3) and at rest (AES-256)
UK-only processing and storage by default, agreed per site
Data minimisation enforced in processing, not by policy
Site-level category restrictions honoured technically, not by policy
Documented retention and deletion, per the data agreement

Operations

Audit logging across every read of patient data
DPIA support pack provided for the site's own assessment
Documented incident response with defined notification paths
Change management with peer review on production changes
Separate environments; production data never used in development
Subprocessors bound by downstream agreements no less protective

AI governance

What the models are allowed to do.

Accrual is an AI product operating on patient records in a regulated setting. These are the constraints your review will want in writing, and they are all in the site agreement.

Is patient data used to train models?

No. Records are processed to answer your site's questions and nothing else. They are not used to train or fine-tune foundation models, not pooled across sites, and not shared with a model provider for improvement. Zero-retention terms are required of any model provider in the path.

Does the AI decide who is eligible?

No, and it is architected so it cannot. Accrual assembles evidence and produces a ranked, cited shortlist. Eligibility is determined by an investigator against the protocol. There is no autonomous decision and no automated action taken on a patient.

How do you handle hallucination?

Every criterion determination must cite the span of the record that produced it, and a determination without a citation is suppressed rather than shown. Where the record cannot settle a criterion the output is 'needs review', not an inferred answer.

Can we audit what the model did?

Yes. Every read of a patient record is logged, and every determination retains the model version, the criteria set, the cited evidence and the coordinator's disposition. That trail is available to the site, and is what makes the system reviewable rather than merely explainable.

How is the model evaluated?

Against coordinator dispositions — confirmations and dismissals on real shortlists — per protocol and per site rather than on a general benchmark. We publish no accuracy figure today because we have not earned one, and an unqualified number would be worth nothing to your review anyway.

What about bias across populations?

A matching model that reads some populations' records less well would quietly worsen the exact disparity diversity mandates exist to correct. Performance is monitored by demographic stratum, and divergence is treated as a defect rather than a limitation.

Not yet in hand

What we have not certified.

Security pages usually imply more than they hold. Here is the gap, stated plainly, because a governance reviewer will find it anyway and would rather we said it.

SOC 2 Type II

Not yet attained. Controls are being built to the Trust Services Criteria with an audit planned, but we will not claim a report we do not hold.

NHS DSPT

Not yet submitted. Most trusts require Data Security and Protection Toolkit “Standards Met” before a data-sharing agreement is signed, so this is the gate we are working to first.

ISO 27001

Not yet certified. Controls are being built to the standard alongside Cyber Essentials Plus, but neither certificate is in hand and we will not imply otherwise.

Penetration testing

Independent testing is planned before general availability. Results will be made available to customers under NDA once they exist.

Reporting a vulnerability

Tell us, and we will not come after you.

If you have found a security issue, write to security@getaccrual.co.uk with enough detail to reproduce it. We acknowledge within two working days and will keep you updated until it is resolved.

We will not pursue legal action against anyone who reports in good faith, avoids privacy violations and data destruction, and gives us reasonable time to fix the issue before disclosing it.

Governance review

Send us your security questionnaire.

We would rather answer it up front than discover a blocker three months in. Tell us what your review requires.