Before any AI automation is built, the first thing a real estate firm wants to know is where its data goes. Rent rolls, tenant records, IC papers and draft valuations are commercially sensitive, and some of that information is personal data under UK GDPR. It is a fair question, and one that can be structurally built into automations to complement the existing security that Claude Code already provides under the terms of your Team or Enterprise contract.
I build on Claude Code, Anthropic's tool for running AI on your own computer. The protection comes in two layers. The first is built into Claude and backed by your contract with Anthropic, which gives you certainty over how your data is handled. That does not remove the need for a second layer inside the automations themselves, so that a user cannot accidentally release sensitive information outside your organisation. Further controls can be built into the connectors Claude uses (its links to your email, SharePoint and other systems) to reduce the risk of data being overwritten, or of information being sent before it has had a full review.
What Anthropic provides
On business accounts (Team, Enterprise and API), Anthropic's Commercial Terms rule out using your prompts, files or results to train its models. Conversation data is deleted after 30 days, and Enterprise clients can choose zero data retention, where nothing is kept once the answer is returned.
Data is encrypted in transit and at rest, and Anthropic is audited against SOC 2, ISO 27001 and ISO 42001, the first international standard for managing AI. Claude Code runs on the computer it is installed on: your files stay there, and only the text needed for each request is sent to the model. By default, it asks before it changes a file, runs a command or fetches anything from the internet.
What I add to every build
Least access
Each automation is given one named folder or system and nothing more. Commands run in a sandbox, a walled-off workspace, with internet access limited to specific websites set within the plugin (the package that configures each automation), such as your SharePoint or HM Land Registry. Where passwords are needed, they sit in the computer's secure store, never in the automation's files, and access is read-only wherever possible.
Human approval
Sending an email, uploading a file or sharing data outside your business always needs a named person to approve it. Incoming documents, emails and web pages are treated as information, never as instructions, so text planted in a document (known as prompt injection) cannot redirect the automation. The automation drafts; a person decides.
Audit trail
Every action is logged with the time, what was touched and who approved it, and any change to the rules is logged too. Where an automation handles personal data (tenants, buyers, staff), a Data Protection Impact Assessment is completed before go-live, as the ICO expects. When the work ends, I sign out of every account, disconnect every link to your systems and delete the local working files.
A straight answer on risk
No one can honestly promise that any system, AI or otherwise, carries zero risk. That is why the protections are layered: if one fails, the next still stands between your data and the outside world.
The full briefing sets out Anthropic's eight protections and my twelve steps over five pages. If you would like a copy, or would like to go through how each step applies to your own processes, email adam@re-ai.co.uk. Follow RE·AI on LinkedIn for Issue 03.
Adam Moore MRICS · Founder, RE·AI
Sources: Anthropic, Claude Code security, data usage, sandboxing and settings documentation; Anthropic Commercial Terms of Service and Trust Center; Information Commissioner's Office, Guidance on AI and data protection. Checked October 2026.