Responsible Handover Factsheet
Responsible Handover Factsheet
1. Background
CattleGrid is a UK-registered AI API governance gateway. It was founded by three people who have spent careers inside enterprise technology, healthcare data, and product design, and who observed — from inside large organisations — the gap between AI adoption and the governance infrastructure needed to support it responsibly.
CattleGrid is not a conventional AI tool. It is a governance layer that organisations deploy between their people and AI providers. Its purpose is to make AI adoption auditable, policy-compliant, and sovereign without preventing it.
2. Purpose
Organisations are adopting AI at a rate their governance functions cannot match. In regulated sectors this creates live exposure, with client data and sensitive information passing through AI API calls without inspection, logging, or policy control.
CattleGrid addresses the gap between acceptable use policy (governance intent) and technical enforcement (governance reality). It intercepts outbound AI API calls, applies customer-configured data governance rules in real time, and produces an immutable audit trail of every interaction.
The intended outcome: organisations can adopt AI tools with confidence that sensitive data is governed at the point of use, and that they have the evidence to demonstrate it to regulators, clients, and auditors.
3. What CattleGrid Does — and Does Not Do
Naming scope boundaries clearly is part of how CattleGrid works. The following reflects the current product at September 2026 launch.
Within scope
- Real-time inspection of outbound AI API calls before data reaches the AI provider
- Block or redact configured data categories (PII, financial identifiers, privileged content, custom rules)
- Reversible tokenisation (A/D): PII held in memory during AI processing, never transmitted — restored on return
- Document inspection: PDF, Word, Excel, ODT, RTF files inspected before upload to AI tools
- Immutable, tamper-evident audit trail exportable for regulatory inspection
- Vetted provider registry — only approved AI providers accessible through the gateway
- Named-person alert routing for SMF, COLP, and DPO oversight
- Usage analytics dashboard for governance visibility
- Zero-retention architecture — prompt content inspected in memory, never written to disc
- Governance layer on UK and EU infrastructure — audit records outside US jurisdiction
Outside scope
- Does not inspect AI responses or govern the output layer
- Does not conduct Data Protection Impact Assessments (DPIAs)
- Does not maintain Records of Processing Activities (RoPA)
- Does not draft or negotiate Data Processing Agreements
- Does not govern AI accessed via browser (consumer web interfaces, e.g. ChatGPT.com)
- Does not govern Microsoft Copilot / M365 Copilot (proprietary infrastructure, no standard API interception point)
- Does not cover fully on-premise or self-hosted LLM deployments
- Does not produce ISO 27001 ISMS documentation or security audit reports
- Does not constitute legal, regulatory, or compliance advice
4. Data Handling
CattleGrid processes outbound prompt content and, where document inspection is enabled, file content. The following statements govern how that data is handled.
Prompt content | Inspected in memory against configured rules. Never written to disc. Not retained after inspection. Not used for product development or model training. |
|---|---|
Document content | Extracted in memory, inspected, discarded. Same zero-retention posture as prompt inspection. No document store. |
Tokenisation (A/D) | Where enabled: detected identifiers replaced with tokens before the AI call. Re-identification map held in process memory for that single round trip only. Destroyed on every exit path. No persistent key. Classified as pseudonymisation under UK GDPR Art 4(5) — the customer remains data controller throughout. |
Audit metadata | Volumetric and pattern data only — request frequency, rule trigger rates, block/pass ratios. No content-level data. Used in aggregate and anonymised form to improve detection performance. |
Audit logs | Owned by the customer. Stored on CattleGrid infrastructure under the customer's data retention configuration. Exportable at any time. On contract termination, deleted within 30 days with a deletion certificate issued. |
Data processor status | CattleGrid operates as a data processor under UK GDPR Article 28. Processing is governed by the CattleGrid Data Processing Agreement (DPA v1.2), available on request. |
Infrastructure | Civo UK (London) — production. Civo Frankfurt — backup and mirroring. Both non-US-domiciled. No CLOUD Act exposure on CattleGrid's governance layer. AI inference continues to reach AI providers (OpenAI, Anthropic, etc.), which are US-domiciled; CattleGrid does not change that. |
5. Technical Approach and Performance
CattleGrid is a rules-based gateway. It performs no ML inference in the blocking or redaction logic. Every detection decision is the output of an explicit, human-readable policy rule configured by the deploying organisation. This is a deliberate design choice: a compliance officer must be able to explain to a regulator exactly why a specific prompt was blocked, and with CattleGrid that explanation is always available.
Detection method | Configurable rule-pack engine: regex patterns, keyword matching, and context-aware rules. Centrally managed and versioned. AI-model-assisted detection is in development. |
|---|---|
Standard rule packs | UK and EU PII categories (National Insurance numbers, passport numbers, credit card numbers, NHS numbers, financial account identifiers, email addresses, phone numbers). Legal rule pack (privilege markers, client identifiers). Financial services rule pack (FCA-relevant categories). Additional packs: in development. |
Custom rules | Customer-configurable via the rule builder. Organisations are expected to review and supplement standard rules for their specific context. CattleGrid provides rule-writing guidance and professional services for organisations requiring bespoke rule development. |
Performance monitoring | Detection rate, false positive rate, and false negative rate logged per deployment. Thresholds agreed with customers at onboarding. Monitored continuously and surfaced in the compliance dashboard. |
Explainability | Full. No opaque model parameters. Every blocking decision references the rule that triggered it, visible to any authorised user of the dashboard. |
Autonomy level | Low. CattleGrid enforces rules defined by the deploying organisation. It does not make independent decisions. All policy changes require authorised administrator action. Future versions may introduce ML-assisted detection as an opt-in module — any change to the autonomy profile will be documented in advance. |
Formal benchmarking | In development. Sector-specific test suites for legal, financial, and healthcare-adjacent contexts are being completed ahead of general availability. Current coverage uses standard UK and EU PII format sets. |
6. Legal and Regulatory Alignment
CattleGrid supports (and does not replace) compliance with the following frameworks. The role column indicates whether CattleGrid's contribution is Direct (primary technical control), Partial (one dimension of a broader obligation), or Supporting (provides evidence relevant to compliance).
Framework | Role | What CattleGrid provides |
|---|---|---|
UK GDPR Art 32 | Direct | Technical measures for security of AI data processing. Zero-retention architecture, in-transit inspection, audit trail. |
UK GDPR Art 26/28 | Direct | Operates as a compliant data processor. DPA v1.2 governs the relationship. Deletion certificates on termination. |
EU AI Act Arts 26 / 26(5) | Partial | Supports deployer obligations: oversight logs, input-layer governance. Does not address output suitability or conformity assessment. |
FCA AI Update / Multi-Firm Review 2024 | Direct | Shadow AI visibility, immutable audit trails, third-party AI oversight evidence — the three supervisory expectations documented in the FCA's 2024 review. |
FCA Consumer Duty | Partial | Client data controls supporting Consumer Duty outcome obligations. Data dimension only. |
SMCR / SMF accountability | Supporting | Named SMF alert routing. Exportable governance reports for Senior Manager oversight trail. |
SRA Code (COLP / Privilege) | Direct | Privilege boundary enforcement. Client identifier rules. COLP alert routing. |
ISO 27001:2022 Annex A | Direct | Technical control evidence for access control, encryption in transit, audit trail, incident alerting, and operations logging in AI API context. |
Cyber Essentials | Supporting | AI API access control layer as part of broader security posture. CattleGrid certification in progress. |
Responsible Handover Framework (SaS / Wellcome) | Direct | CattleGrid has completed a full tool proposer self-assessment against the Framework. That document is available on request. |
7. Certification and Documentation Status
The following reflects CattleGrid's status at September 2026 launch. This table will be updated as certifications are achieved.
Item | Status | Notes |
|---|---|---|
Cyber Essentials | In progress | Self-assessment complete. Submission pending. |
ISO 27001 | In progress | Annex A evidence register produced. Certification trajectory active. |
ISO 42001 (AI Management) | Planned | On roadmap following ISO 27001 completion. |
Algorithmic Impact Assessment | Planned | Committed before regulated sector deployment. |
Data Processing Agreement (DPA v1.2) | Available | Available to prospective customers on request. |
System Overview documentation | Available | Available to prospective customers under NDA. |
Responsible Handover self-assessment | Complete | Full tool proposer self-assessment. Available on request. |
8. How CattleGrid Uses the Responsible Handover Framework
CattleGrid occupies an unusual position in relation to the Framework. It is simultaneously a tool proposer (a commercial product completing this factsheet) and a governance intermediary — the technical enforcement layer that organisations deploy after completing a responsible handover of any AI tool.
In practice, this means the Framework is now part of how CattleGrid structures conversations with regulated sector clients approaching AI adoption. The Framework's structured question set provides a principled entry point for discussing what governance obligations an organisation faces, what AI tools they are considering, and where CattleGrid fits as a control layer.
CattleGrid views the direction of travel toward a mandated responsible handover standard — as discussed with the Institution of Engineering and Technology — as consistent with and complementary to its own governance positioning. We are an active supporter of this work and have committed to participating in the interactive software beta programme as an early tester.
9. Contact
Technical contact: rob@cattlegrid.uk
Data protection contact: brian@cattlegrid.uk
General enquiries: hello@cattlegrid.uk
Website: cattlegrid.uk
Legal notice
This factsheet is provided for information purposes and to support adoption due diligence. It does not constitute legal, regulatory, or compliance advice. Organisations should seek independent advice on their specific compliance position. CattleGrid is a data processor as defined under UK GDPR Article 28 and does not provide legal advice on data protection obligations.
CattleGrid Ltd | Company No. 17020793 | Cheltenham, UK