Task statement 6.3 · Governance, risk and responsible use

6.3 · Organizational AI Policy

In Wrenfield Housing, the course example, the AI policy has two rules: no tenants' personal data in prompts unless anonymised first, and a person checks anything that goes to tenants, the board or the regulator. Those rules protect nobody until people follow them in daily work, know whom to ask, and speak up when something slips.

Why it's tested

The tempting habit is to treat the AI policy as the IT team's job, so anything the settings allow must be fine. Settings cannot see what someone pastes into a chat, and they do nothing at all in a personal account. Hiding a slip fails too, because it leaves the data where it landed and the gap open for the next person.

KnowledgeWhat an AI policy holds, and who owns it

An organisational AI policy is the written set of rules for how staff use AI tools at work. It usually says which tool and plan to use, what data may go into a prompt, who checks the output, how a new kind of use gets approved, and what to do when something goes wrong. Governance, the processes around those rules, covers who owns the policy, who approves requests, how slips are recorded and when the policy is reviewed. The policy needs a named owner, because a rule nobody owns is never updated and nobody knows whom to ask. At Wrenfield the policy owner is the governance manager. The IT team hold the Owner role on the Team plan and turn the policy's decisions into settings. This rules out any answer where each team sets its own rules.

Following the policy in daily work comes down to five habits, in the order a task meets them:

  1. Use the approved tool and plan: the organisation's account, not a personal one.
  2. Check the use is covered: the policy names this kind of task, or the owner has approved it.
  3. Ask before a new kind of use: a short request to the policy owner, before you start.
  4. Work within the rules: what data may go in, and who checks the output before it leaves.
  5. Report a slip the same day: to your team leader and the policy owner, saying what went where.

Interactive · the path the policy gives each situation

Three paths through Wrenfield's AI policy, each read from top to bottom. A task for Claude starts with one question: is this kind of use already approved? Yes leads to the Team plan account, where rules 1 and 2 apply and a person checks output before it leaves. No or not sure leads to a short request to the governance manager, who approves or refuses it. A slip, such as tenant data in a personal chat, leads the same day to telling the team leader and the governance manager, then the chat is deleted, the team gets a short reminder, and nobody is blamed. Paths not chosen are dashed and labelled not this path. A TASK FOR CLAUDE SOMETHING WENT WRONG A task you want Claude's help with Is this kind of use already approved? YES Team plan account rules 1 and 2 apply BEFORE IT LEAVES A person checks it tenants, board, regulator NOT THIS PATH NO, OR NOT SURE Short request to the governance manager DECISION Approved: added Refused: it does not go ahead NOT THIS PATH A slip tenant data, personal chat SAME DAY Report it team leader + governance mgr THEN Chat deleted, reminder no blame NOT THIS PATH Each path runs top to bottom. None of them is to carry on and say nothing.

Choose a situation and follow the path the policy gives it, from top to bottom.

Approved use: work in the Team plan account under rules 1 and 2, so a person checks anything going to tenants, the board or the regulator.

New use, or not sure: send a short request to the governance manager before you start. An approved use is added to the policy; a refused one does not go ahead.

A slip: tell your team leader and the governance manager the same day. The chat is deleted, the team gets a short reminder, and nobody is blamed.

Solid: the path you are on. Dashed and labelled NOT THIS PATH: the paths for other situations.

Asking first matters because a new kind of use can need data the policy keeps out, and the policy owner can spot that before anything happens. At Wrenfield a request is a short form to the governance manager: the task, the data that would go in, and who will check the output. When you are unsure whether a use is covered, treat it as new and ask, because a quick answer costs far less than a slip.

KnowledgeWhat Team and Enterprise admins can control

On the Team and Enterprise plans, an organisation's admins can turn parts of the policy into settings. Admins and above add members, while Owners enable each connector, the link that lets Claude reach a service such as Google Drive. Settings narrow what is possible but cannot judge what a person types, so the habits above still carry the policy. No setting reaches a personal account, which is why the organisation's account is the first habit (6.2 covers how each plan handles data). The grid shows each control and its plans, checked 2026-10-15:

Members and rolesTeam, Enterprise

Admins and above add members; Owners change roles. Custom roles are Enterprise only.

Policy useOnly approved staff get an organisation account.

ConnectorsTeam, Enterprise

An Owner or Primary Owner enables each connector, and each person still signs in. Owners can set a connector's actions to Always allow, Needs approval or Blocked.

Policy useOnly approved services reach Claude.

Project sharingTeam, Enterprise

Owners can turn project sharing off for the organisation, and on Enterprise for specific roles.

Policy useWork stays with the people meant to see it.

Organisation instructionsTeam, Enterprise

The standing instructions an Owner sets for every chat in the organisation (5.3).

Policy useA short policy reminder shapes every chat.

Audit logsEnterprise

Owners and Primary Owners export the last 180 days of organisation activity.

Policy useEvidence when the policy is reviewed.

KnowledgeYour policy sits on top of Anthropic's

Anthropic's Usage Policy applies to anyone who uses Claude, so an organisation's policy sits on top of it. The organisation can add rules, such as Wrenfield's rule 1, but it cannot allow what the Usage Policy forbids. As 2.4 showed, the Usage Policy requires a qualified person to review advice or decisions about an individual in high-risk areas such as housing. At Wrenfield those decisions stay with people (6.1). A new version takes effect on 12 November 2026. Each new version is a reason for the policy owner to reread the organisation's policy against it. Wrenfield's governance manager therefore reviews the policy once a year and whenever the Usage Policy changes.

SkillWorked example: the slip a new starter reported

In their second week, a new operations officer wanted help replying to a tenant's damp complaint. Their Team plan login was not set up yet, so they pasted the tenant's email into their personal Claude account, name and address still in it. That broke the approved-tool habit and rule 1 at once. The officer noticed within the hour. From there it could go two ways, and the wrong way is a quiet deletion:

# hidden
11:10  tenant's email pasted into a personal Claude account
11:50  chat deleted; nobody told
→ no record, no check on what else is needed, the next starter does the same
# reported
11:10  tenant's email pasted into a personal Claude account
11:55  officer tells their team leader and the governance manager
12:30  chat deleted; slip logged; data-protection lead asked if more is needed
→ a short reminder at the next team meeting; Team plan login on day one

The cause was a gap, not only a careless moment: the officer had a task and no approved account, so the personal account was the nearest tool. Reporting it turned one slip into a fix for everyone, because the governance manager could ask the data-protection lead what else was needed and close the induction gap. Nobody was blamed, since a policy that punishes reporting teaches people to hide the next slip. The officer then drafted the reply in the Team plan without the tenant's personal data in the prompt.

Ask one question: are this tool, this data and this kind of use already covered by the policy? Yes → go ahead within its rules. No or not sure → ask the policy owner before you start; if something has already slipped, report it the same day.
Takeaway
  • An AI policy says which tool to use, what data may go in, who checks the output and how new uses are approved, and it needs a named owner.
  • Team and Enterprise admin settings narrow what is possible, but only people's habits keep personal data out of prompts.
  • Report a slip early and without blame, and review the policy whenever Anthropic's Usage Policy changes.

check yourself

0 / 3 answered

A marketing coordinator wants Claude to turn customer reviews into social media posts. Their company's AI policy lists approved uses, this is not one of them, and the policy says a new use needs the policy owner's approval first. The company's Team plan is available, and the posts are due in two weeks.What is the most appropriate first step?

Select one.

In their first week, a finance assistant pastes a supplier's contract, bank details included, into a personal Claude account to summarise it. An hour later they realise that the company's AI policy allows only the company's Enterprise plan, and that it asks staff to report any slip promptly.What is the most proportionate response?

Select one.

A company on the Claude Team plan has a new AI policy. The policy owner asks an IT admin, who holds the Owner role, for two settings: staff may use only the approved Google Drive connector, and a short reminder of the policy's data rule should shape every chat in the organisation.Which TWO settings meet these requests on the Team plan?

Select two.

sources

  1. Claude Help Center, "Manage members on Team and Enterprise plans": Admins and above add members; roles Primary Owner, Owner, Admin and Member; custom roles on Enterprise only. Checked 2026-10-15.
  2. Claude Help Center, "Roles and permissions": Owners invite and remove Admins and Owners and modify roles; data retention controls on Enterprise only. Checked 2026-10-15.
  3. Claude Help Center, "Use connectors to extend Claude's capabilities": on Team and Enterprise an Owner or Primary Owner enables connectors; each person still authenticates; Owners can set actions to Always allow, Needs approval or Blocked, and users cannot override it. Checked 2026-10-15.
  4. Claude Help Center, "What are projects?": project sharing on Team and Enterprise; Owners can turn it off for the organisation, or on Enterprise for specific roles. Checked 2026-10-15.
  5. Claude Help Center, "Set organization instructions": Owners and Primary Owners on Team and Enterprise; Claude follows them in every conversation in the organisation. Checked 2026-10-15.
  6. Claude Help Center, "Access audit logs": Enterprise organisations only; Owners and Primary Owners export the past 180 days. Checked 2026-10-15.
  7. Anthropic, "Usage Policy": applies to anyone who submits inputs to Anthropic's products; effective 12 November 2026; Anthropic will update it as the technology and its risks evolve. Checked 2026-10-15.
  8. Wrenfield's governance manager, its request form and the new starter's slip are this course's example.