Security pack
Claude and your documents
IMS Creator writes the documents a management system needs (the context analyses, policies, registers, plans and procedures) with Claude, Anthropic's AI model, and reads drawings with it. This section says exactly what is sent to Claude, when, on whose account, and what happens to it there. It is the part security reviews ask about most.
On your own Claude account
Claude work runs on your company's own Claude account, never on ours, and you choose the provider:
| Provider | What you give IMS Creator | Where Claude runs | Your agreement |
|---|---|---|---|
| Anthropic API | An API key from your Claude Console | Anthropic's infrastructure, where Anthropic's terms and your account's settings say | Anthropic's Commercial Terms and Data Processing Addendum |
| AWS Bedrock | An AWS region and an access key | The AWS region you choose (an Australian region is available) | Your AWS agreement |
| Google Vertex AI | A Google Cloud project, a region and a service account key | The Google Cloud region you choose (an Australian region is available) | Your Google Cloud agreement |
So the provider that processes your content is your supplier, under your contract, with your data settings. You can see every call in your provider's own console, set spend limits there, and revoke the key there at any time. For processing in Australia, choose Bedrock or Vertex AI in an Australian region.
A site can use its own Claude connection instead of the company's, for example to keep one site's costs apart.
Your key
- Checked with one small call when it is entered, then never shown again: the app shows only the provider, the region, the key's last four characters, who set it and when.
- Encrypted with your company's (or site's) own key and decrypted only for the moment of a call made for your work.
- If the provider refuses it (revoked, out of credit, over its limit), Claude work stops at that site, the reason is shown, and your company admins are emailed once. Unusual use (a daily spike against the past fortnight, or use out of hours) is also emailed to them.
- Every call's token counts are recorded for your site, so your admins can see what Claude costs them.
What is sent to Claude, and when
Claude is called only when someone at your company asks for it (writing or reviewing a document, reading a batch of drawings, proposing asset register entries), or by a scheduled job your admins set up. What each call carries:
| Work | What is sent | What comes back |
|---|---|---|
| Writing a document (context, risks, objectives, plans, procedures) | Your site's or company's profile (what you entered about it); the documents it builds on; any documents you chose to give it in full (for example an annual report); the instructions for that document | A draft, for a person to review, edit and approve |
| Reviewing a document | The document, and what has changed since it was written | Proposed changes, for a person to accept or not |
| Reading drawings | Page images of the drawings in the batch (JPEG and PNG), and close-ups of title blocks and handwriting | Title block details, tags, a summary |
| Asset information (classes, failure modes, maintenance programs) | The asset classes, their attributes, and a summary of the work order history for those classes | Proposed entries, for a person to accept |
Not sent: passwords, authenticator secrets, other companies' information, your people's sign-in history, the audit log. Each call carries only what that piece of work needs.
Web research. When writing context documents, Claude can search the web and read public pages to research your site and its environment (with every fact cited). These searches run on your provider's side (Anthropic's web search and web fetch tools), and the search terms Claude writes can include your site's name and location. Web research can be turned off for a site in its settings; Claude then works from your profile alone and lists what should be checked.
What Claude cannot do. Claude is given no tools that act on IMS Creator or your systems: its answers are text, kept as proposals that a person reviews and approves. Nothing it writes is run as code. Uploaded documents and drawings are given to it as material to read, not as instructions, and its answers come back as a form with fixed fields, which is what is stored.
What your provider keeps
This is set by your provider's terms, not ours. As published at the time of writing (November 2026); check the current terms through the links below.
Anthropic API
- No training on your content. Anthropic's Commercial Terms say: "Anthropic may not train models on Customer Content from Services." Customer Content means your prompts (inputs) and Claude's answers (outputs).
- Retention. Anthropic's documentation on API data retention says that prompts and outputs are not retained by default, except for its covered models (its most capable, Mythos-class models, such as Claude Fable), which keep prompts and outputs for 30 days for safety work, deleted automatically after that, with no Anthropic personnel able to read them by default. IMS Creator uses Claude Opus, Sonnet and Haiku, which are not covered models, unless your admins choose a covered model for a site.
- Kept longer only when flagged or required by law. Content flagged by Anthropic's automated trust and safety systems may be kept for up to 2 years, and when the law requires it.
- Zero data retention. Anthropic offers zero data retention by arrangement with your organisation. IMS Creator works with it: it uses the Messages API, prompt caching and web search (all eligible) and does not use the Batch API or the Files API (which keep data for a time), on the hosted service. If you need web fetch inside a zero data retention arrangement, check its current eligibility with Anthropic.
- Prompt caching. When a long document (such as an annual report) is given to Claude for several steps of one piece of work, IMS Creator marks it for caching, so it is read once. Anthropic holds the cached form in memory only, for 5 minutes after last use, and caches are never shared between organisations.
AWS Bedrock: AWS states that Bedrock does not store or log your prompts and completions, does not use them to train models, and does not share them with model providers: Anthropic has no access to them.
Google Vertex AI: governed by your Google Cloud terms for generative AI (including Google's commitment not to train on your data without permission); see Google Cloud's documentation on Vertex AI data governance.
Sources: Anthropic Commercial Terms · Anthropic API and data retention · Data retention for covered models · Prompt caching · Web search tool · Amazon Bedrock data protection · Vertex AI data governance
On our side
- What Claude returns is stored in your company's own databases, like everything else you write in IMS Creator.
- While a batch of drawings is being read, Claude's answers are held on the server's encrypted disk until they are collected into the register, then deleted.
- We do not use your content, or Claude's answers about it, for anything but providing the service to you: not to train or tune any model, not to improve prompts with your material, not shared with anyone.
Choices you have
| You want | Do this |
|---|---|
| Claude processing in Australia | Connect AWS Bedrock or Google Vertex AI in an Australian region |
| No retention at the provider at all | Anthropic API with zero data retention arranged with Anthropic, or Bedrock |
| No web searches about your sites | Turn off web research in the site's settings |
| A limit on spend | Set it at your provider; IMS Creator also emails your admins about unusual use |
| To stop Claude at once | Remove the connection in IMS Creator, or revoke the key at your provider |