Legal services
AI inference you can use for confidential client data.
UK data residency. No US sub-processors. No training on client communications.
Legal teams using commercial AI APIs to process client correspondence, contracts, and advice face a question their providers cannot fully answer: what happens to the data, and who might see it. Marigold runs inference on private AWS in London. No data reaches a US-based model provider. Nothing is retained for training. For privileged communications and confidential client data, the architecture makes the answer straightforward.
Commercial AI APIs are third-party data processors. Under the SRA Code of Conduct, the data you send through them is data you are responsible for. Marigold gives your information governance team a clear, documentable answer to where the data goes, who can access it, and whether it is retained.
Data handling properties
01
UK data residency
Inference runs on private AWS in London. Data does not leave that region as part of inference, storage, or queuing. No cross-border transfer occurs.
02
No US sub-processors
No inference request reaches OpenAI, Anthropic, Google, or any other US-based model provider. The model runs from weights held on UK infrastructure. Your AI supply chain has one link.
03
No training on inputs
Client data passed to the inference API is not retained for training, fine-tuning, or any secondary purpose. Outputs are stored briefly for retrieval and then deleted.
04
Open-weight models
All models run from publicly documented open-weight checkpoints. Weights, architecture, and training provenance are inspectable. No proprietary black-box model receives your data.
05
Audit trail
Every inference job records status, timing, and identity within the UK region. Execution history is queryable. Access logs are retained in CloudWatch within London.
06
UK-registered company
Bay Information Systems is a UK-registered company. Any contractual relationship is with a UK entity subject to UK law.
SRA and privilege considerations
Know your data supply chain
The SRA Code of Conduct requires solicitors to keep client affairs confidential and to have effective systems for doing so. Where AI tooling processes client data, the tooling is part of that system. Sending client data to a US-based API constitutes a transfer to a third-party processor in a foreign jurisdiction. Under most standard commercial terms, that data may be subject to retention, review, or training unless an enterprise agreement explicitly prohibits it.
Whether any specific use of Marigold is appropriate for privileged communications is a matter for your firm's risk assessment and your IG team's judgement. What Marigold provides is a clear technical answer to the data handling question: UK-resident processing, no US sub-processors, no retention beyond the job window, and a DPA available on the Pro tier. We are happy to provide the technical documentation your IG team needs to complete that assessment.
| Property | Marigold | US-based AI API |
|---|---|---|
| Data processed in UK | Yes | No |
| UK-registered processor | Yes | No |
| No US sub-processor for inference | Yes | No |
| No training on client data | Yes | Varies / opt-out required |
| Audit log within UK region | Yes | No |
| Inspectable open-weight models | Yes | No |
Use cases
These are illustrative. Whether any specific use case is appropriate for your practice depends on the data involved, its classification, and your firm's risk framework.
Contract review and extraction
Extract key terms, obligations, defined terms, and critical dates from contracts at volume. Structured output for review, not replacement of review. No document content leaves UK jurisdiction.
Due diligence document review
Classify and summarise large document sets in M&A, litigation, or regulatory matters. Surface relevant documents against a query. Audit trail for every classification decision.
Client correspondence classification
Classify incoming correspondence by matter, urgency, and type. Route to the correct fee-earner without manual triage. All processing within UK jurisdiction.
Frequently asked
Does Marigold process data outside the UK?
No. Inference runs on private AWS in London by default. Data does not leave that region as part of inference. No US-based model provider receives your data.
Does Marigold train on client communications or inference inputs?
No. Inference inputs are not retained for training, fine-tuning, or any other purpose. Outputs are stored briefly for retrieval and then deleted on schedule.
Is Marigold appropriate for legally privileged communications?
Whether using Marigold for privileged communications is appropriate depends on your firm's risk assessment and SRA compliance framework. What Marigold provides: UK-resident processing, no US sub-processors, no retention beyond the job window, and a data processing agreement on the Pro tier. How those properties interact with privilege in your specific use case is a matter for your IG team to assess. We are happy to discuss the technical architecture in detail to support that process.
Is a data processing agreement available?
We are able to discuss data processing terms with practices that have specific requirements. We do not currently have a standard DPA document ready to issue. Contact us and we will work through your requirements directly.
What models are available?
The hosted registry includes instruct models for text generation and summarisation, embedding models for search and classification, and image-to-text models. All are open-weight with publicly documented provenance. See the full registry.
AI inference with a clear answer to every data handling question.
If you are evaluating AI tooling for a legal practice or in-house team, we would like to hear from you. Register your interest below or contact us directly.
Register interest
Tell us your practice type and use case and we will follow up directly.
Noted. We will be in touch.
Prefer to talk directly? ed@bayis.co.uk