Claude Code in Regulated Work

In a regulated organisation, what matters most is whether an AI agent's work can be controlled and demonstrated afterwards. This article describes working with Claude Code in five stages, from task to sign-off. For each stage, the matching control and the evidence that goes with it are set out, and a clear line is drawn between what Anthropic provides and what remains your responsibility.

Controls first, then capability: what has to be in place before an AI agent touches regulated work.

In a regulated organisation, two things matter most about an AI agent: that you can control what it does, and that you can show afterwards what it did.

Claude Code is an agentic framework that works inside a folder on your computer. The AI agent reads files, proposes changes, edits them and runs commands [1][2]. That is exactly what makes it useful, and exactly why it needs clear rules and careful controls in a regulated setting.

Five stages, each with its own control

The animation shows the path a task takes through these stages. For each stage you set up a control in advance. It is built into the system, so it holds every time.

  1. 1. TaskA person describes a small, well-defined job in plain language.
  2. 2. ScopeThe AI agent works in a defined folder, on a working copy of the data, with no personal data.
  3. 3. PermissionEvery edit and every command needs your approval, or falls under a rule you have set in advance.
  4. 4. EvidenceCommands, logs and test output show what actually ran.
  5. 5. Sign-offA named reviewer signs off the result. Only then is the change carried over to the original and recorded in version control.

The AI agent drafts. The evidence proves. A person decides.

How Claude Code can fit into regulated work. Each stage is a control you set up yourself, built into the system. Illustration with invented example data.

What each stage means in practice

StageControl you set upEvidence you keep
1. TaskA written, narrowly defined task. The task contains only data without any personal reference.The task description
2. ScopeA dedicated working folder, a working copy of the data and, where available, a sandbox for commands.The folder and the data version used
3. PermissionApproval prompts before each action, or allow and deny rules set centrally.The approvals given
4. EvidenceCommands and tests that produce checkable output, and monitored usage.Logs, test output, usage metrics
5. Sign-offA named reviewer and version control. Only after sign-off is the change carried over to the original.The reviewed change and its history

The permission stage is the one you notice day to day, because the AI agent asks. The evidence stage is the one that protects you: it replaces the AI agent’s own account with records a reviewer can check.

What Anthropic provides, and what stays with you

The split matters because regulated teams sometimes assume the product answers questions that are actually theirs to answer.

Anthropic, as the vendor, provides a documented security programme, with its certifications listed in its public Trust Center. It also provides product features that support your controls: permission prompts, a sandbox for commands, settings that can be enforced centrally, and usage metrics for monitoring [1][3].

You provide the decisions around its use: which data may be used, which folder the AI agent may work in, which rules are enforced and who reviews. Whether the tool needs formal validation for its intended use depends on your quality management system; that is a question for your quality assurance. And you remain responsible for reviewing proposed changes before you approve them [1].

There is also a people side. Article 4 of the EU AI Act requires organisations to ensure that staff who work with AI have a sufficient level of AI literacy [4]. A short, documented introduction and written guidelines are a sensible way to meet that obligation.

Before a pilot

Settle these points in writing before the first run:

  • The intended use, clearly described
  • Which data is allowed and which is excluded
  • Who owns the pilot and who reviews the results
  • Which permission rules apply and who may change them
  • Which records are kept and where

Then start small, with synthetic or public data in a copy of the working folder. If the five stages hold on a low-risk task, you have a pattern you can extend step by step. If gaps appear, you have found them early and at little cost.

From pilot to production

The pilot deliberately works on a copy and with non-critical data. You move into production once these conditions are met and documented:

  • Your data protection officer has decided which real data the AI agent may process.
  • Your quality assurance has decided whether and how the tool is validated for its intended use, and any required validation is complete.
  • The permission rules are enforced centrally.
  • The five stages have proven themselves in the pilot, as shown by the records kept.
  • The responsible person has approved the move and documented that approval.

The working-copy principle stays in place in production. The AI agent continues to work in its own workspace, such as a separate copy or its own branch in version control. A change reaches the original only once a named reviewer has signed it off at stage 5. That keeps your actual work protected at all times and makes every change traceable.

The sandbox is a safeguard for the commands the agent runs: it limits which files and network destinations a command can reach [1]. That is why it stays switched on in production as well.

Sources

  1. Anthropic. Security. Claude Code Documentation. code.claude.com
  2. Anthropic. Claude Code overview. Claude Code Documentation. code.claude.com
  3. Anthropic. Anthropic Trust Center. trust.anthropic.com
  4. European Parliament and Council. Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). Official Journal of the European Union. 2024 Jul 12. Article 4. eur-lex.europa.eu

AI Enablement

We enable you and your people to use AI properly, tailored to your skills, your processes and the tools of your working day. Your team grows confident in putting AI to effective use in daily work.

Book an introductory call or send us a message

Frequently asked questions

Does Claude Code need to be validated before we use it?

That depends on the intended use and on your quality management system. A tool that drafts internal notes is assessed differently from one that works with data feeding into a regulated decision. Involve quality assurance or compliance early and describe the intended use precisely. That description is the basis for the decision.

Which controls come with Claude Code, and which do we set ourselves?

The AI agent asks for permission before it edits files or runs commands, and it can run commands in an isolated environment (a sandbox). Organisations can also enforce settings centrally. Which rules apply, which data the AI agent may see and who reviews the result are decisions you make.

What does the EU AI Act have to do with this?

Article 4 obliges organisations that use AI to ensure that the people working with it have a sufficient level of AI literacy. Introducing an AI agent without training and clear rules misses the purpose of that provision. The law leaves the form of training open; a short, documented introduction and written guidelines are a sensible start.

Can we use Claude Code with patient or personal data?

That is a decision for your data protection officer. Until it has been made, work with synthetic or public data in a separate working folder. If the answer changes later, update the scope in writing and document the change.

How do we know what the AI agent did?

The reliable record is the command output, the test results, the approvals given and the change history in version control. The summary the AI agent writes itself describes its work. Only these records let a reviewer or an auditor check and trace what happened.

Read more