Privacy & Data Security

Our approach to information security

Lexical Labs operates a certified Information Security Management System (ISMS) and is accredited to ISO/IEC 27001. Our certification was most recently renewed following an independent surveillance audit in March 2026.

ISO 27001 certification requires us to establish, implement and continually improve documented policies, procedures and technical controls governing information security across the business — including risk management, access control, asset management, supplier management, secure software development, vulnerability management, business continuity, incident response, staff training and ongoing management review. Many of the practices described below form part of this ISO 27001-certified ISMS and are subject to regular internal review and independent external audit.

AI models and data usage

Underlying models and providers

The platform is built on frontier large language models, principally Anthropic’s Claude family, accessed both directly through Anthropic’s API and through Amazon Web Services’ Bedrock platform. These provide the AI inference that drives the workflows. We integrate the models under the providers’ enterprise and API terms rather than any consumer chatbot product — and it is precisely that arrangement that lets us make the data-handling commitments set out below. Where we work with a different inference provider, we ensure comparable protections are in place.

We do not use your data to train models

No client data is used to train or improve AI models. Under the terms governing our inference providers — AWS Bedrock and Anthropic — data passed through their APIs is explicitly not used to train or improve any model, and is not retained by them for that purpose (see AWS Bedrock model data use and Anthropic’s commercial terms, under which “Anthropic may not train models on Customer Content from Services”). Separately, as a business we do not use our clients’ data to train or improve models either; we hold it only to the extent necessary to deliver the agreed services.

Retention and deletion of inputs and outputs

Inputs and outputs are retained only for as long as necessary to provide the agreed services. In practical terms, a document you upload stays on the system until you choose to delete it; once deleted, it enters a deletion queue and is permanently (hard-) deleted within one month.

Our model providers apply comparable hard-deletion policies: Anthropic operates zero data retention, under which customer data is not stored at rest once the API response is returned (except where needed to comply with law or combat misuse), and AWS Bedrock uses a zero-data-retention model by default. Other third-party services integrated with the platform adhere to strict data policies that let customers control and delete their data:

• Supabase hosts the primary database for client data; it is SOC 2 Type 2, HIPAA and ISO 27001 compliant.
• Dagster stores execution metadata and logs used to operate and monitor the platform.
• E2B, which provides secure compute, is SOC 2 Type 2 and HIPAA compliant.

All data can be deleted at customer request; this is included in our standard terms.

Data security and hosting

Where your data is hosted

Client data is hosted on Supabase’s servers in the UK — this is the primary store of client and personal data. AI inference is provided separately from the UK via AWS Bedrock, with supporting infrastructure in the UK. Orchestration and secure-compute services (Dagster and E2B) are provided from servers in the EU.

Encryption

Data is encrypted at rest — a standard, confirmed feature of our Supabase hosting — and in transit, including the links between the database and the web application, between compute containers, and between our system and third-party inference endpoints.

UK GDPR and legal-sector compliance

Tiro Pro is designed to support compliance with the UK General Data Protection Regulation (UK GDPR) and other applicable UK data protection legislation. Many of our clients are law firms and legal teams handling highly confidential, legally privileged and commercially sensitive information, and the platform has been developed with those requirements in mind. Our approach includes:

• Data processor role — in most engagements we act as a data processor, processing client data only on the client’s documented instructions and solely to provide the agreed services.
• Data minimisation — we process only the information necessary to provide the service, and never for a secondary purpose such as model training.
• Lawful international processing — client data is hosted in the UK and inference is provided through AWS Bedrock under enterprise terms; where processing involves international transfers, these use appropriate safeguards under UK GDPR.
• Technical and organisational measures — encryption in transit and at rest, logical tenant separation, role-based access controls, audit logging, secure development practices and documented incident response.
• Data subject rights — we assist customers, where reasonably required, with requests relating to data subject rights under UK GDPR, including access, rectification, erasure and restriction of processing.
• Retention and deletion — client data is retained only as long as required to deliver the service or as instructed by the customer, and deleted documents are permanently removed under our documented deletion process.
• Legal confidentiality — documents processed by Tiro Pro frequently contain privileged and confidential material; the platform maintains strict confidentiality through customer-specific access controls, logical segregation of customer data and contractual confidentiality obligations on Lexical Labs and our key service providers.
• Contractual commitments — we incorporate standard data-processing and confidentiality terms into our customer contracts and provide reasonable assistance to customers in meeting their own UK GDPR obligations.

Access and permissions

Role-based access

Access is enforced at the data layer, not merely in the application. Authentication is handled by Supabase Auth; once a user logs in they receive a signed token carrying their identity (such as a user or team id), which accompanies every request to the backend. That token is validated at the database, where PostgreSQL Row-Level Security (RLS) evaluates policies on every read and write and returns only the rows the user is permitted to access. Because these policies are declarative and centralised in the database, the risk of accidental data leakage from application bugs is dramatically reduced — the database itself refuses to return unauthorised data. Within each client, access is divided into administrator and standard-user roles; administrators can invite new users, and we agree administrator assignments with each customer.

Authentication

User accounts are protected by robust authentication, with multi-factor authentication (MFA) being introduced as a further layer of account security.

Audit logging

The platform maintains extensive audit logging. Every operation on every document is recorded together with the identity of the user who made the change and the time it was made. Logs can be made available under the auditing provisions of our terms.

Confidentiality and data segregation

How your data is segregated from other clients

The platform uses a shared-schema, multi-tenant architecture with logical data isolation enforced by database-level Row-Level Security. All customer data is stored in shared tables, but RLS policies apply to every access operation, restricting visibility and modification to records within a user’s authorised scope. This provides strong tenant isolation while retaining the efficiency of shared infrastructure — each client and their users can only access data relevant to that client.

Cross-client exposure through the AI

By design, client data cannot be exposed to another client through the AI. Such exposure runs contrary to the fundamental principles of the models’ information- and data-security design, and we believe there is no material risk of it occurring (see AWS Bedrock data protection).

Risk and liability

Safeguards against inaccurate or misleading outputs

Our primary safeguard is an extensive evaluation suite — combining human- and AI-driven tests — that measures the quality of Tiro Pro’s analysis against gold-label datasets and other objective standards. These evaluations, or “evals”, are part of a process of continuous iteration towards higher quality; moving steadily in that direction is itself part of the service. Reliability evals focus on technical behaviour, such as whether output matches what we expect for a given input and how robustly the system handles errors; quality evals assess the substance of outputs against the task in hand. Above the tooling sits expert human judgement, both legal and technical, and a disciplined approach to choosing the right model for each task. These safeguards make outputs as reliable as the technology allows — not infallible.

Liability for AI-generated content

We do not accept liability for errors arising from AI-generated content. Large language models are tools; like any tool, they are only as good as the skill with which they are used, and they augment the work of lawyers rather than replacing professional judgement or the need for human checking. We take care, in the product’s design and our explanatory materials, to present them as tools and to be clear that responsibility for the final work product rests with the professional using them. Our terms and conditions include a skill-and-care warranty to back the services we provide.

Insurance

We carry appropriate cyber and professional indemnity (PII) insurance.

Testing and validation

Accuracy and reliability

Every AI-backed feature is covered by the evaluation suite described above. We are candid that LLM technology is non-deterministic, so we cannot represent that outputs will always be correct — no more than a human lawyer could — and we therefore work on a best-efforts basis, underpinned by our expertise in the technology and in the law. Some features carry particularly high test coverage: the templating feature used for drafting, for example, relies heavily on hard-coded logic with a low degree of non-determinism and is tested frequently, by both machine and human, for correctness.

Benchmarking and references

We have an established base of clients using our legacy contract-review software, which provides contract review for repetitive contracts. Tiro Pro is newer and still in a phase of initial testing; feedback to date has been very positive, and we are happy to share benchmarking, validation studies and client references in due course.

Incident management

Detecting and responding to incidents

As part of our ISO 27001 accreditation, Lexical Labs maintains a documented information security incident management process designed to ensure the rapid detection, assessment, containment, investigation and remediation of security incidents:

• Detection and monitoring — continuous monitoring of cloud infrastructure, applications and security events; centralised collection and review of system logs, authentication events and application activity; and automated alerts for suspicious or anomalous behaviour.
• Assessment — reported or detected events are triaged by authorised personnel and classified by severity and by impact on confidentiality, integrity and availability; where personal data may be involved, an impact assessment is undertaken immediately.
• Containment — affected systems are isolated to prevent further compromise and preserve evidence, and credentials are revoked or rotated where appropriate.
• Investigation — root-cause analysis determines the source, scope and impact; relevant logs and evidence are retained in line with our obligations, and third-party providers are engaged where necessary.
• Recovery — services are restored in a controlled manner once the threat has been removed and controls verified.
• Post-incident review — a formal lessons-learned review tracks corrective and preventive actions to completion and updates documentation, procedures and training.

Breach notification

We notify customers without undue delay after becoming aware of a confirmed security incident that has affected, or is reasonably likely to affect, customer data. For incidents involving personal data, notification is typically made within 24 hours of confirmation, unless legal or law-enforcement considerations require otherwise. Notifications include, where available, the nature of the incident, the categories of data affected, the likely impact, the containment measures taken and any recommended customer actions, with ongoing updates as more information becomes available. Where required by law, we support customers in meeting their own regulatory notification obligations, and a summary report can be provided on request once an incident is closed.

Future development

New features will not materially change our data-handling and security arrangements, which are already sufficient for the use cases and roadmap features we anticipate. Nothing on the roadmap involves standing up new infrastructure or sending data to a new destination — it all operates within the architecture and data flows described here — so while the product keeps developing, the security and data-handling posture set out on this page remains stable.

Contact us

If you have any questions about this policy or how we handle your data, please contact us at privacy@lexicallabs.com.