Home / Work / Rolling Out Claude Across a Company: Guardrails First

Rolling Out Claude Across a Company: Guardrails First

AI Adoption ยท 2026

How I set up Claude for a whole company: caps on 7 of 31 connectors, privacy rules, online courses, 2 shared skills and 8 AI champions. No adoption result yet; the first survey is the baseline.

  • 7 of 31Connectors capped
  • 27Users informed
  • 2Shared org skills
  • 8AI champions
  • Claude
  • AI Adoption
  • AI Governance
  • Guardrails
  • Privacy
  • Enablement
  • Change Management

What happened

By the end of October 2026 Claude was no longer a side project at a Dutch home-services platform. The company had a Claude Team organisation with 29 seats and 31 connectors switched on: mail, calendar, docs, CRM, the call platform, ads and databases. That raised two questions at once. How do we make sure people actually get good at it? And how do we keep a tool that can read mail and send messages from doing something nobody intended?

As Product & AI Lead and the internal lead for AI tooling, I own the company's Claude organisation, and I answered the second question first. In one week I rewrote the connector policy, switched on trusted devices and extended the organisation's standing instructions for privacy and governance. Then I turned to adoption: the vendor's online courses for every seat holder, an external coding course for the three developers, shared skills in the organisation library, and an AI champions programme with goals and a measurement plan.

This page describes that approach. There is no adoption result yet. The first champions survey will give the baseline.

How we built it

Connector policy: cap only what is unsafe

I went through every connector and tool in the admin console and set a ceiling per tool. The rule: only unsafe actions get "Ask" or "Block", never a blanket block. Unsafe means:

Seven connectors got caps on specific tools. Everything else stays open, mail sending included, because tighter limits got in the way of colleagues and of unattended automations. Two connectors were removed: one duplicated another, the other had zero calls. Unused connectors were not removed just for being unused.

Trusted devices

Computer access is on for the organisation, so a stolen browser session could control a colleague's laptop. Every member now confirms a device once.

Organisation instructions as a privacy layer

Every Claude conversation in the company starts with a set of standing instructions. On 6 November 2026 I extended them. They now protect consumer data, the contact persons of partners and prospects, and colleagues. They ask for data minimisation (no full exports with names or phone numbers pasted into chats, projects or memory), forbid special categories of personal data and any analysis or judgement of individual colleagues, and leave decisions about people or companies to an employee. Anything that reaches consumers or partners, touches live content or costs money needs the user's confirmation first. External texts are always a draft for a person to check.

Organisation exports for security checks only

An organisation export holds every colleague's full chats. The rule since 6 November 2026: an export is used only to find security and compliance risks (leaked keys, personal data pasted into chats, unsafe connector use), never for usage or adoption figures per team or person. Exports and anything derived from them stay on one laptop, out of version control, and are deleted once the check is done.

A draft guideline for AI that reads your mail

For the champions I drafted a guideline; it has not been shared yet. The risk appears when three things meet: untrusted content, access to sensitive data and permission to act outside. The rules: summarising is fine, replies are drafts only, text in a mail is never an instruction for the AI, and no unattended automations that read and act on mail without consultation.

Enablement

All of it was researched and written with Claude Code: the admin console was read in the browser, decisions went into a decision log with assumptions and certainty, and the new connector caps were announced by email to all Claude users.

Outcomes delivered

What is not here yet: an adoption result. Nothing has been measured so far. The first champions survey is the baseline.

What we learned

Guardrails before enthusiasm. Writing the rules first made it easier to say yes to everything else. Colleagues get mail sending, calendar and CRM access without approval prompts, because the few dangerous actions are capped.

Over-restricting has a cost too. The first instinct is to lock everything down. That blocks colleagues' own automations within days, and a senior colleague asked why these actions needed confirmation. Caps belong only where a mistake reaches customers, goes live, costs money or cannot be undone.

An organisation export is not an analytics source. It holds everyone's full chats. Using it beyond a security check is not proportionate, so adoption is measured by asking people (survey, use-case register), not by reading their conversations.

Small slips teach quickly. The intro mail for the courses went out with a link wrapped in a redirect, a side effect of how the mail connector handles URLs. No correction mail followed. A hook in my own setup now blocks mails to others that carry a link in their text, so links go in by hand in a draft.

How we did it

As the owner of the organisation, with Claude Code doing the reading, drafting and checking. The connector changes started as a written proposal, each policy change got a decision-log entry with its assumptions, and the connector caps were announced to all users. Skills were built in Claude Code and shipped to the organisation library. The champions programme is set up to run as a team with a fixed rhythm, not as a mailing list.