AI Policy
Which model providers we use, what reaches them, what never does, how long our inputs are kept, and where a person stands between a model and anything we publish.
1. Why This Document Exists
We publish readings of how other organisations describe their own use of artificial intelligence. A company that does that and stays quiet about its own use has not earned the reading. This document answers, about us, the questions our calculator asks of everyone else.
2. The Providers We Use
We use large language models from two providers, and from no others:
| Provider | What We Use It For | Country |
|---|---|---|
| Anthropic PBC | Research drafting, summarisation and analysis of published source documents | United States |
| OpenAI OpCo, LLC | Research drafting, summarisation and cross-checking of published source documents | United States |
Both are named on our Sub-Processors page. If we engage a third model provider we will name it here and there before it processes anything.
3. The BSI Calculator Uses No Model At All
This is the point most often assumed wrong, so it is stated first. The BSI Calculator is a rule-based reader, not an AI product. It fetches pages an organisation has published, strips the markup, and applies a fixed, versioned rule set to the text. No model is called, no prompt is constructed, and nothing you type into it is sent to Anthropic or to OpenAI.
That is why a calculator result is reproducible. The same pages read against the same rule set produce the same reading, and the rule set carries a version identifier so that two readings taken months apart can be told apart from each other.
4. What Is Sent To A Model, And What Is Not
Models are used in our research work: reading published law, published guidance and published corporate documents, and drafting analysis from them.
4.1 What We Send
Published source material, our own notes and drafts, and our own method documents.
4.2 What We Do Not Send
We do not send client confidential material to a model without the client's written instruction. We do not send personal data belonging to a data subject who is not a public officeholder acting in that capacity. We do not send calculator inputs or calculator results. We do not send credentials, keys or internal system data.
5. Training
We do not use customer data to train models. We do not train models of our own, and we do not fine-tune a provider's model on client material.
Our accounts with both providers are configured so that our inputs are excluded from model training. We do not permit any provider to train on what we send, and we treat a change to that position by a provider as a change of sub-processor, notified under section 6 of our Sub-Processors page.
6. Retention Of Prompts And Model Inputs
Prompts and model inputs are retained by each provider for a limited period for abuse monitoring, set by that provider under its own terms, and are then deleted. We keep no separate copy of a prompt, and no prompt is stored in our database.
Model outputs that become part of a research deliverable are retained as part of that deliverable, on the retention terms in section 11 of our Privacy Notice.
7. Automated Decision-Making
We carry out no automated decision-making that produces a legal effect concerning any individual, and we carry out no profiling. No model decides anything about a person here.
A calculator result is not a decision about the organisation either. It reports what could be established from published documents and what could not, and the second half is stated as plainly as the first.
8. A Person Stands Between The Model And The Page
No model output is published without a person checking it against the source it claims to rest on. Every factual claim in a research deliverable carries a quotation, a source address and the date it was retrieved, so that a reader can verify it without trusting us.
Where a claim cannot be evidenced that way, it is recorded as not measured rather than softened into something that reads as a finding. An unverified claim is the failure mode this business exists to expose, and it does not get a pass because a model produced it.
9. Accuracy, And What A Reading Is Not
A calculator reading is an observation of published documents at a moment in time. It is not legal advice, it is not an audit, it is not a certification, and it is not a statement about what an organisation does internally. An organisation may hold a control it has not published; the reading says the control was not found in what was published, which is a different sentence and is written as one.
10. Governance Of This Policy
This policy is owned by the founder, who is accountable for it. It is reviewed when a provider changes, when the method changes, and at least once a year.
Our security controls are described in our Security Overview. What we collect and why is in our Privacy Notice. Our obligations when we process data for a client are in our Data Processing Terms.
11. Changes To This Policy
This is version 1.0, published on 17 August 2026. When we change it we publish the new version at this address and move the version number and the date. A change of model provider is always a material change and is stated as one.
12. How To Reach Us
Write to ror@blacksnowintel.com. Our other published documents are indexed on the Legal page.