What Rolling Out Claude Across a Company Actually Involves

Turning Claude Enterprise on takes a purchase order and an afternoon. Then security asks what is retained, finance asks what "plus usage" will come to, and you find that people are still using personal accounts because nobody told them not to.

This article sets out the work that follows the signature, in roughly the order you meet it. It is written for whoever has to make the thing work for everyone else: IT admins, security and privacy leads, finance partners and department heads. Product details are as they stood in late October 2026, and settings change, so check Anthropic's release notes before you act. The quotations come from my book Claude Enterprise: The Field Guide, an independent guide that is not affiliated with Anthropic.

Decide what you want before you switch it on

The most expensive sentence in enterprise software is "let's switch it on and see what people do with it".

It sounds open-minded. It is a decision not to decide, and six months later nobody can say whether the rollout worked.

The alternative fits on one page. Call it an intent statement: three to five measurable goals, the non-goals, a line on risk appetite, the baseline measures, and a signature from an executive sponsor who will still be in post at the review.

Then fill in names. A deployment is governed by people before it is governed by policy, and the first job is a roster: the executive sponsor at the top, and four pillars of named roles beneath. Any line left blank is a risk you have already accepted.

Avoid giving all of it to a small AI team. They become the queue for every connector request, budget question and nervous email from legal, and within a quarter the organisation decides the tool is slow when the queue is. An ordinary chart of who is responsible, accountable, consulted and informed for each area is the remedy.

The first month should be boring. Anthropic's own rollout guidance is a reasonable skeleton: involve IT, security and procurement, do a security review, pilot with one role, review the settings, monitor for two to four weeks, then expand in waves.

Identity, data and connectors

What you are buying is a tenancy: one administered space in which every member signs in with the company identity, meets the same Claude in every tool, and leaves a record you can audit. A seat is one identity, not one application.

So single sign-on is the foundation. Three jobs follow from it.

Domain capture. Most organisations find, on the day they verify their domain, that Claude already has dozens or hundreds of accounts on it: free trials, subscriptions on personal cards, plans expensed quietly. Domain capture is how an Enterprise admin finds and claims them. Review the list first. The people active this week are your first conversation.

Provisioning. This decides how joiners get in, how movers change role and how leavers go out. The default is invite only. The alternatives are just-in-time provisioning or SCIM directory sync, each with or without group mappings from your identity provider.

The network edge. Personal accounts on company devices are a network matter more than a behavioural one. Tenant restrictions let your web proxy add a header that admits your organisation and refuses the rest.

With identity settled, security will open with training data. On Claude Enterprise, inputs and outputs are not used to train models by default. The commitment is in the commercial terms and documented on Anthropic's Trust Center.

The better follow-up is retention. By default an organisation keeps conversations until someone deletes them. An Owner can set a custom period, with a minimum of 30 days. Before you shorten it, list the projects whose last edit is older than the new period, because those are about to vanish and someone should sign that list.

Then come connectors, which are how Claude reaches systems you already run: mail, files, the CRM, the ticket queue. Without one, Claude knows only what is typed or uploaded. With one it can search live data and, if you allow it, act. A connector policy is a stack of layers, not a single switch. An Owner has to add the connector to the tenancy, a role has to permit it, and the member has to connect their own account before any data moves.

The moment that needs most care is when an agent goes to use a tool.

An agent is only as safe as the moment it reaches for a tool.

At that moment a session either asks the member or acts under a standing grant. Organisation-level switches decide whether members may grant "Always allow" for connector tools, and whether they may let Claude act without approval at all. Pick one connector that is already enabled and list its write tools. If any is on Always allow and nobody can say who decided that, set it back to needing approval.

The money is per seat, plus usage

Claude Enterprise is sold per seat plus usage at API rates. A seat therefore has a floor but no ceiling unless you give it one.

Spend limits are the ceiling. Since July 2026 you can set them at three levels: the organisation, a group drawn from your identity provider, and the individual member. They nest, so a member cannot exceed their group and a group cannot exceed the organisation. Pooled group budgets, added in October 2026, let a department hold one shared budget.

Treat the forecast as a range. Sketch a few user profiles and a model mix, set alerts before anyone hits the wall, and match model and effort to the task instead of defaulting everyone to the most capable option.

People are most of the work

A pilot exists to teach the organisation something it does not yet know. That is a different purpose from proving the sponsor right. Give it a baseline, enough weeks for the novelty to wear off, and success thresholds written down before the first login.

Then build a champion network: one person in every team of 25 to 50 who already likes the tool, with a little time, a shared channel and a direct line to the people who run the platform. Nobody learns this from a slide deck. They learn from the colleague two desks away.

Training is also a legal matter if you have staff in the EU. Since 2 February 2025, Article 4 of the EU AI Act has required deployers to ensure a sufficient level of AI literacy among their staff. There is no prescribed course or certificate. What counts is that the training fits the role and the risk.

Expect pushback around week six. People resist for four reasons that are worth separating: fear for their job, doubt that it works, the extra workload of a new tool, and worry about losing the craft. All four are reasonable, and only some are yours to fix.

Finally, the policy. Most organisations have an AI acceptable-use policy and most staff have never read it, because it runs to many pages.

A policy that nobody reads is not a control; it is a liability waiting for a hearing.

Two pages answering four questions will do: what data may go in, what needs human review, what must be disclosed, and what is forbidden. Read it aloud. If it takes more than six minutes, cut it.

Measure honestly, and plan for the bad day

Most business cases carry the same sum: hours saved per person per week, times seats, times a salary. It is nearly always wrong, and not from dishonesty. People are poor judges of their own speed, and a saved hour is rarely a recovered one. Count tasks and quality against a baseline, and keep one wave back for a few weeks as a comparison group.

Something will go wrong. A confidential document lands in a shared Project that the wrong group can see. A number in a board paper turns out to have been invented by the model. All are survivable if you decided in advance who looks in the log, who blocks the connector and who tells legal. Write that three-line runbook now.

Then review the settings every quarter: capabilities, connectors, network egress, retention and roles. Defaults change. The whole programme reduces to five verbs, repeated: decide, enable, teach, measure, review.

What to take from this

• Write one page of goals, non-goals and risk appetite, and get it signed, before a seat is provisioned.
• Put names against every area of ownership, and do not route everything through one team.
• Sort out single sign-on, existing personal accounts and provisioning first.
• Know your retention period, and check what a shorter one will delete.
• Treat connector write tools and standing approvals as the main risk surface.
• Seats have a floor and no ceiling. Set spend limits and alerts.
• Run a pilot that can fail, recruit champions, and keep the policy to two pages.
• Measure against a baseline, and review settings quarterly.

If you are doing this for real, Claude Enterprise: The Field Guide goes through each of these areas in detail, including contracts, audit logs and the Compliance API, skills and plugins, Cowork and its sandbox, Claude Code at company scale, and department-by-department rollout, with a "Try it" action in every chapter. It is 100 chapters and a 114-term glossary, as a 329-page PDF and an EPUB, and costs £9 on Gumroad: https://matsiems.com/books/claude-enterprise.