CCAO-F 6.3 — Organizational AI policy

Which rules apply, and in what order

Narrowest wins

Law and regulationData protection, sector rules, contracts with clients
Anthropic Usage PolicyApplies to every user on every plan
Your employer’s AI policyApproved accounts, approved data, required review
Organization instructionsOwners set standing guidance for every chat
Project and personal setupInstructions and preferences for your own work

Read down: each layer can only narrow the one above it, never widen it. Most workplace questions are decided by the middle two layers.

Who can set what

On Team and Enterprise plans, governance is not an abstraction — it is a set of named roles with named controls. Owners and Primary Owners can set organization instructions: standing guidance, up to 3,000 characters, that Claude follows across every conversation in the organisation. They live in organisation settings, under organisation and access. Where they conflict with an individual user’s own instructions, Claude favours the organisation’s; where the organisation is silent, the individual’s still apply.

On Enterprise, a Primary Owner or Owner can also configure custom data retention. By default data is retained indefinitely unless a period is set; the minimum is thirty days and longer periods are chosen in thirty-day increments. Shortening a period schedules everything outside it for permanent, irreversible deletion, carried out by a daily background process that runs at midnight UTC and can take several days to work through. Retention actions are recorded in audit logs. And across Team and Enterprise, the Primary Owner can request data exports that may include conversations, uploaded files and usage patterns, and can remove a user’s access.

Control	Who sets it	What it actually does
Organization instructions	Owner / Primary Owner, Team and Enterprise	Standing guidance in every chat, up to 3,000 characters
Custom data retention	Owner / Primary Owner, Enterprise	Deletes data past a set period; minimum 30 days
Data exports	Primary Owner	Produces conversations, uploaded files and usage patterns
Audit logs	Organisation administrators	Record retention changes and deletion events
Removing access	Primary Owner	Ends a user’s access to the work account
Incognito chat	Any user, any plan	Skips history, memory and training — not an exemption from policy

What a usable AI policy actually says

Most policies fail not because they are wrong but because they are unreadable. A policy people follow answers six questions in language they recognise. Which accounts may be used — almost always “the company Team or Enterprise account, never a personal one”. Which data may go in, by classification rather than by example. Which tasks need a named human reviewer before the output leaves the building. When AI involvement must be disclosed to a customer, candidate or patient. What has to be recorded, and where. And who to ask when the policy does not cover the case in front of you.

That last question is the one most policies omit and the one staff need most. A rule with no escalation route turns every unanticipated situation into a private judgement call, made at speed by whoever is under deadline pressure. Naming a person — a data owner, a line manager, a compliance contact — costs one line and prevents most of the quiet workarounds.

A policy page that gets followed

Unusabletext

AI Usage Policy

Employees must use AI
tools responsibly and
in line with company
values.

Confidential data must
be handled with care.

Outputs should be
checked for accuracy.

Usabletext

AI Usage Policy
1. Company Claude
account only. Never a
personal account.

2. No customer names,
account numbers or
health data in prompts.

3. Anything sent to a
customer: named
reviewer signs off.

4. Note in the file that
AI was used.

5. Not covered? Ask
Priya, the Data Owner.

The weak version is not false — it is simply unusable at the moment somebody has a document in one hand and a deadline in the other.

Worked example · Shadow AI at a professional-services firm

A 200-person consultancy has no AI policy. A partner discovers that eleven consultants have been using personal Claude subscriptions to summarise client documents, because it saves hours and nobody said not to.

Name the real problem. It is not that they used AI. It is that client material has been processed through accounts the firm has no agreement over, no visibility into and no ability to export or delete.
Fix the account, first. The firm opens a Team account so work happens under an organisational agreement, where Anthropic acts as a data processor for the firm rather than under consumer terms with eleven individuals.
Write the six answers. Approved account, permitted data classes, which deliverables need partner review, when the client is told, what goes in the engagement file, and who to ask.
Make one rule technical. An Owner sets organization instructions reminding every chat that client names are not to be used and that draft advice must be reviewed by the engagement partner.
Amnesty, then enforcement. Staff are asked to move existing work across without blame. Enforcement starts after the tool they are being asked to use actually exists.

Takeaway: Shadow AI is a supply problem before it is a discipline problem. People route around a ban that leaves the work undone; they rarely route around a sanctioned tool that is quicker than the workaround.

Worked example · Writing organization instructions that earn their 3,000 characters

The Owner of a Team account at a housing association wants the organisation’s rules to be present in every conversation, not just in a document nobody opens.

Start from what is universal. Anything that varies by team belongs in a project, not here — the instructions apply to every chat in the organisation.
Write behaviour, not aspiration. “Never include a tenant’s name or address in a draft; use ‘the tenant’ and leave placeholders” is a rule Claude can follow. “Respect tenant privacy” is not.
Add the house vocabulary. The organisation’s terms for its own processes, so drafts come back in the language colleagues use.
Leave the rest out. Safety rules and Anthropic’s Usage Policy are not overridable here, and length is capped at 3,000 characters — spend them on what is specific to this organisation.
Tell people it exists. Staff who do not know why Claude keeps leaving placeholders will assume it is broken and work around it.

Takeaway: Organization instructions are where a written policy becomes a default. They do not replace the policy, the training or the review — they just stop the most common slip happening by accident.

Configured guardrails and the gap they leave

There is a real difference between a rule that is written down, a rule that is configured, and a rule that is enforced. Organization instructions are guidance Claude follows; the documentation is explicit that they are not a way to override safety guidelines, and equally they are not a technical barrier that makes a forbidden action impossible. Retention settings genuinely delete data. Data exports and audit logs genuinely create visibility. Nothing in the product stops a determined person pasting something they should not.

That gap is why governance standards pair configuration with two human things: training, so people know the rule, and review, so a second pair of eyes sees output before it matters. Deciding which outputs need that second pair of eyes is covered in 2.4; the judgement about which tasks belong to Claude at all is 6.1.

The third human thing is the record. Governance standards, in every field, come down to being able to show afterwards what was done and why — the same accountability idea the data-protection principles state directly. In practice this is modest: a line in the case file, the engagement record or the ticket saying a draft was AI-assisted and who reviewed it. It costs seconds, it is the first thing an auditor or a complaining customer asks for, and it is the part teams skip first because nothing breaks when they do.

A governance review, item by item

Passes: Work happens on the organisation’s account, not personal onesTeam plan in place; personal use discouraged in writing
Passes: Organization instructions carry the two rules that matter mostNo tenant identifiers; review before sending
Check: A retention period is set deliberatelyEnterprise default is indefinite until someone chooses
Fails: Staff know who to ask when the policy is silentNo named escalation contact anywhere in the document
Missing: Use of AI is recorded where the work is recordedNothing in case files says a draft was AI-assisted
Fails: New joiners are told the rules in inductionPolicy circulated once by email, eight months ago

Notice the pattern: the configured items pass, the human ones lag. That is the usual shape of a first review.

Traps the wrong answers are built from

Tempting but wrong	Do this instead
Using a personal account for work because the company has not provided one	Raise the gap; work under the organisation’s agreement, where the employer is the controller.
Assuming an incognito chat puts work outside company policy	Policy applies to the task, not the chat mode, and work-plan incognito chats still appear in exports.
Writing a policy of principles with no named contact	Answer six concrete questions and name who to ask when it is silent.
Treating organization instructions as enforcement	Treat them as a default that prevents accidents, backed by training and review.
Leaving Enterprise retention on its default and calling it a decision	Choose a period deliberately; the default is indefinite retention until someone sets one.

You should now be able to

Order the rules that apply to a work task and identify which one governs.
Name the administrative controls available to Owners and Primary Owners, and what each does.
Judge whether an AI policy is usable by testing it against six concrete questions.
Recognise shadow AI as a supply problem and describe the sanctioned-tool response.
Distinguish a rule that is written, a rule that is configured, and a rule that is enforced.

Organizational AI policy