Using AI assistants like ChatGPT and Claude with ISO 27001: what's safe and what isn't

Yes, you can use an AI assistant such as ChatGPT or Claude to help run an ISO 27001 information security management system (ISMS), and it can save a small team a lot of time. It's good at drafting, summarising and pulling answers out of records you already have. It isn't a substitute for the people the standard holds accountable: ISO 27001 needs management decisions, named owners and records that someone has actually reviewed.

The risk isn't using AI. It's using it carelessly: pasting confidential ISMS content into a tool nobody has assessed, trusting clause numbers it made up, or letting "the AI said so" stand in for a decision. This guide covers where it helps, where it bites, and how to connect an assistant to your ISMS in a way you could explain to an auditor.

Where AI genuinely helps

The work in an ISMS that eats time is mostly reading and writing. That's where assistants earn their keep.

• Drafting. First drafts of policies, procedures, risk descriptions, treatment plans and meeting minutes. A draft that describes your business is much faster to fix than a blank page.
• Summarising. Turning a long incident email thread, a supplier's security pack or an audit report into a short summary with the actions pulled out.
• Answering customer security questionnaires. Most questions have already been answered somewhere in your policies and records. An assistant can find the relevant text and draft an answer, as long as it's working from your approved records rather than guessing. (See our guide to customer security questionnaires.)
• Chasing. Listing what's overdue, who owns it and what's coming up, so the routine follow-up doesn't depend on one person remembering.
• Explaining. Translating a control into plain English for the person who has to do it.

In each case, the output is a draft or a suggestion. A person still checks it and decides.

Where it's risky

Pasting confidential ISMS content into an unassessed tool

Your risk register, incident records and network details are some of the most sensitive documents you hold. They list your weaknesses. Pasting them into a personal AI account is sending information to a cloud service, and ISO 27001 expects you to have thought about that:

• Annex A 5.10 (acceptable use of information and other associated assets) expects rules for how information may be used. An AI tool is an "associated asset" like any other.
• Annex A 5.23 (information security for use of cloud services) expects you to decide how you acquire, use, manage and exit cloud services. An AI assistant is a cloud service.
• Annex A 5.12 (classification of information) gives you the labels. If something is "confidential", your rules should say whether it can go into an AI tool, and which one.

Check your AI provider's terms before anything sensitive goes in: whether your inputs can be used for training, how long they're kept, where they're processed, and what a business or team plan changes. If personal data is involved, UK GDPR applies too; the ICO's AI guidance is a good starting point.

Copy-paste goes stale and leaves no trail

Even with an approved tool, copy-paste has two quiet problems. The assistant only sees the snapshot you pasted, so it can answer confidently from last quarter's risk register. And nothing records what was shared, by whom, or what came back, which is awkward when clause 7.5 (documented information) asks you to control your ISMS records.

Made-up clauses and controls

Assistants are fluent, and fluency isn't accuracy. Common mistakes include:

• citing the 2013 Annex A numbering (such as "A.9.2.3") instead of the 2022 structure of 93 controls in four themes (organisational, people, physical, technological);
• inventing a clause or a requirement that sounds plausible but doesn't exist;
• describing a control you don't have as though you do, which is worse in a questionnaire answer than leaving a gap.

Check every clause and control reference against your own copy of the standard before it goes into a policy, your Statement of Applicability or anything a customer will read.

Letting the AI "approve" things

This is the one auditors will notice. ISO 27001 is built around accountable people:

• Clause 5.1 requires top management to demonstrate leadership and commitment to the ISMS.
• Clause 5.3 requires top management to assign responsibilities and authorities for roles relevant to information security.
• Clause 6.1.3 requires risk owners to approve the risk treatment plan and accept the residual risks.
• Clause 9.3 requires top management to review the ISMS at planned intervals.
• Annex A 5.2 and 5.4 cover information security roles and management responsibilities.

An assistant can prepare all of that work. It can't own a risk, accept one, approve a policy or sit in a management review. If your records show an AI tool making those decisions, you don't have the accountability the standard asks for.

What MCP is, in plain English

The Model Context Protocol (MCP) is an open standard that lets an AI assistant connect to another system and use it through a defined set of tools, under that system's own permissions. Instead of you copying text out of your ISMS and pasting it into a chat, the assistant asks the ISMS directly: "list the overdue tasks", "show me the open incidents".

For an ISMS, that's a better model than copy-paste for four reasons:

1. Scoped. The system decides what the assistant can see and do, not the person pasting.
2. Current. Answers come from today's records, not a stale snapshot.
3. Revocable. Access runs on a token you can switch off, rather than text that's already left the building.
4. Audited. Each call can be logged by the system it touches, so there's a record of what was accessed and changed.

MCP is only a protocol, though. How safe it is depends on how the system on the other end implements it, which is what the checklist below is for.

A checklist for connecting an AI assistant to your ISMS safely

Use this whether you're connecting to ClearIMS or anything else.

Two practical tips. Start with read-only access for a few weeks and see how people actually use it before granting anything more. And add AI assistants to your supplier and cloud service records like any other service, so they come up in your reviews.

How ClearIMS does it

ClearIMS has an MCP server, so you can connect your own AI assistant, such as Claude or ChatGPT, to your ISMS on any paid plan. An organisation admin has to turn it on, and it's off by default. Each person creates their own named access token, which expires (90 days by default) and can be revoked at any time. Your assistant can see and do only what you can in ClearIMS, under the same role checks as the app.

Tokens are read-only by default, covering things like an overview of your ISMS, risks, policies, tasks, incidents and system health. A read-and-write token can do everyday actions: log an incident, add an incident timeline entry, create a task, add evidence notes to a task, record a system health check, comment on a policy that's in review, and propose a risk reassessment. Approvals, sign-offs, releases, deletions and settings are never available to an assistant: they stay with your people. Every call is recorded in the audit trail, and changes made this way are marked "via AI assistant (MCP)".

You don't need your own assistant to get the benefit, either. ClearIMS works like a virtual ISMS manager: it emails people about their outstanding and overdue actions, and its AI drafts customer security questionnaire answers from your approved records, with sources. People still approve; the AI never approves anything. See how connecting your own AI assistant works.

We make the software. Your certification is yours, awarded by your certification body.

Take the 2-minute readiness check, or book a 20-minute call.