How to Roll Out Claude Enterprise: The 5 Key Decisions
Claude
Inteligencia Artificial
Antropico

Rolling out Claude Enterprise well comes down to five decisions, made in order: Structure & Identity, Access, Governance, Spend and Visibility. Each has a named owner, a small set of choices and a cost if you change it later. Four settings are hard to undo: domain claiming, how many organizations you run, how groups map to roles, and data retention. Get those right in the first 30 days, before most people get access, and the rest of the rollout is about adoption. This guide explains each decision from Generation Digital's experience of rolling out AI with over 200 clients in the US, UK and Canada, using the rollout approach Anthropic recommends.
Free download: the Claude Enterprise Rollout Kit. A decision workbook for all five decisions, the 90-day plan, and an editable decision record and Excel tracker. Get the kit →
Key takeaways
Five decisions, in order. Structure & Identity → Access → Governance → Spend → Visibility. Each decision scopes the ones after it.
Name an owner for every decision first. Spend usually belongs to whoever commits the budget, and Visibility to whoever answers for data risk.
Protect yourself from the four hard-to-undo settings. Domain claiming can't be reversed. Topology, group mapping and retention are expensive to change once people are working.
Groups do almost all the work. Nearly every control attaches to a group, and the union rule means a narrower group can never take access away.
Turn visibility on before access. The Compliance API can't record the weeks before you switched it on.
Treat every cap increase as a question. Model defaults, effort levels and instructions usually move spend more than raising caps.
Claude Enterprise rollout at a glance
Decision | What you decide | Usually owned by | Hard to undo? |
|---|---|---|---|
1. Structure & Identity | How many Claude organizations you run, and which groups live inside them | Claude Owner with your identity team | Yes: topology, domain claiming and group mapping |
2. Access | Which surfaces (Claude chat, Cowork, Claude Code, Claude for Microsoft 365) and connectors each group gets, and how deep | Claude Owner with team leads and the data-risk owner | No: granting later is cheap, removing breaks workflows |
3. Governance | Who can build and share skills, plugins and projects, and how freely | Claude Owner with your AI centre of excellence | No: loosening is free, tightening needs care |
4. Spend | Where caps sit and who approves more | Whoever commits the budget | No: caps change any time |
5. Visibility | What you record and measure, and how long content is kept | Whoever answers for data risk | Yes: deleted content can't come back |
Why Claude Enterprise rollouts succeed or stall
Rolling out Claude Enterprise is much more than putting a tool behind single sign-on. Done well, it changes how teams work. Done badly, it creates a short spike of curiosity, a long tail of unused seats and a security team that's nervous about what it can't see.
The gap between buying AI and scaling it is well documented:
Nearly two-thirds of organisations have not yet begun scaling AI across the enterprise (McKinsey, The state of AI, November 2025).
Gartner predicts over 40% of agentic AI projects will be cancelled by the end of 2027, citing costs, unclear value or weak risk controls (Gartner, June 2025).
Only 36% of employees feel adequately trained on AI, and 54% would use AI tools even if they weren't authorised (BCG, AI at Work, June 2025).
78% of AI users bring their own AI tools to work (Microsoft and LinkedIn, Work Trend Index, May 2024). That's one reason domain claiming and a sanctioned rollout matter.
In our client work, rollouts rarely stall because of the technology. They stall because a decision had no owner, a setting was made in a hurry and was hard to reverse, or nobody could show the value at the end. The five-decision method fixes all three: every decision has an owner, the hard-to-undo settings are flagged up front, and visibility is designed in from day one.
Start with your rollout objective
Before you touch a setting, agree what your organisation is trying to achieve. For some it's broad transformation: a new way of working everywhere. For others it's narrower: one function, one workflow or a commitment leadership has already made. The same five decisions apply either way. The objective shapes how you make them.
A useful objective has two halves:
What success looks like, stated so you can measure it. For example: every business unit using Claude weekly by the end of the quarter.
The constraint that must hold. For example: no security escalation from the regulated finance team, even if that slows the rollout there.
Keep it to one organisation-wide ambition. Differences between units come later, through phasing and the groups you measure, not through separate goals. When a choice would speed up adoption but put the constraint at risk, the objective tells you which way to go.
Who owns each decision?
A decision with no owner is one of the quickest ways a rollout stalls. One person may own several decisions, and titles vary too much to be a guide. Look for what the owner is accountable for and can commit.
Decision | Look for someone who… | What they need from you |
|---|---|---|
Structure & Identity | Holds the Claude Owner role, runs your identity provider (Okta, Microsoft Entra ID or Google Workspace), and knows what your directory can express | Your proposed group structure and the directory groups it relies on |
Access | Leads the teams using Claude and can speak to their daily work, plus whoever signs off data risk for connectors | Which groups get which surfaces and connectors, in what order, and the workflow behind each write grant |
Governance | Owns change management and enablement, such as an AI centre of excellence, with security if needed | The governance posture options and what each asks of people |
Spend | Can commit budget: a finance partner or the executive whose budget line it is | Proposed caps and the usage story behind them |
Visibility | Answers for data risk: security, legal or compliance | The options, your recommendation and what each costs to change |
Two of the five usually land with someone other than the Claude Owner: Spend goes to whoever commits the budget, and Visibility goes to whoever answers for data risk. Involve both early. In our experience, the rollouts that slip are the ones where finance or security first hear about Claude when the bill or the audit question arrives.
How we handle it: in our first-week workshop we map an owner to every decision and run a short intake. Which functions are in wave one? Is there a named data-risk owner? Who runs the identity provider? Does any function handle regulated data? The answers tell us which decisions need extra care.
Prerequisites: six checks before you configure anything
Structure & Identity is the decision that's hardest to change later, so prepare for it. Six prerequisites matter. Four are must-dos: three block group setup outright, and one is insurance against locking yourself out.
Prerequisite | Why it matters | Must-do before setup? |
|---|---|---|
At least two people hold the Owner role directly, not only through a group | Stops a misconfigured group sync locking your admins out of the settings that would fix it | Yes: lockout insurance, before you configure SSO |
Domain verified | Single sign-on can't be configured until your domain is verified | Yes: blocks group setup |
Identity provider connected and SSO enforced | Claude reads groups from your identity provider. No connection, no sync | Yes: blocks group setup |
Provisioning app set up in your identity provider | The directory side of the sync. Without it, groups don't reach Claude | Yes: blocks group setup |
Naming convention for Claude groups | Easy to agree before groups exist. Renaming later needs a resync | No: run in parallel |
Billing owner identified | Needed for the contract and the Spend decision | No: needed before Spend |
Two roles deserve special care. The Primary Owner can do things no other role can, such as exporting organisation data or handing over the Primary Owner role, and is never removed automatically by a sync. Keep that person out of day-to-day admin. Give every other admin the least powerful role that covers their job.
Domain claiming: the one setting you can't reverse
People may already be using Claude on personal or team accounts created with their work email. Until you claim the domain, those accounts and everything in them sit outside your organisation. Verifying proves you own the domain. Claiming is a separate step that brings existing accounts on the domain into your organisation and routes future sign-ups through SSO.
It can't be reversed. Plan it as a one-way door.
Prerequisites: domain verified, organisation creation restricted, SSO enforced (not just configured) and provisioning live.
Add everyone to your identity provider first. Anyone on the domain who isn't there is locked out once the claim starts.
People get a 30-day window to bring their conversations and projects across or start fresh. Custom skills never transfer, and connected apps must be reconnected, so warn people to save anything they rely on.
Turn on "Restrict organization creation" anyway. Once your domain is verified, it stops new personal accounts or organisations being created on it, even if you never claim.
How we handle it: we draft the member notice, agree the date with HR and internal comms, and send it at least two weeks before the claim starts. In engineering-heavy organisations we usually find dozens of existing accounts, so the notice matters more than people expect.
Free download: the Claude Enterprise Rollout Kit. A decision workbook for all five decisions, the 90-day plan, and an editable decision record and Excel tracker. Get the kit →
The four settings that are hard to undo
Most Claude Enterprise settings are cheap to change. Four are not. Decide them deliberately, write them down, and get the right owner to sign them off.
Setting | Why it's hard to undo | Decided in |
|---|---|---|
Domain claiming | Brings accounts on your domain into your organisation. Once on, it can't be reversed | Prerequisites |
Organization topology | Merging or splitting organisations later means re-provisioning every affected member | Decision 1 |
Group structure and role mapping | Changing a live mapping shifts access for everyone it covers. Reversing it is a second mass change, not an undo | Decision 1 |
Data retention | You can change the period, but content deleted under a shorter window is gone for good | Decision 5 |
Decision 1: Structure & Identity
Your decision: how many Claude organisations you run, and which groups live inside them. Every other control in the rollout attaches to the groups you define here, so this is the root of everything else.
One organization or many?
In Claude Enterprise, an organization is the boundary around your members, groups, settings and data. Organisations don't share content: projects, conversations, artifacts and custom skills in one aren't reachable from another. They can share sign-in through a parent organization.
One organisation is the default. Only three situations justify a split:
Split trigger | What it means |
|---|---|
Contract: part of the company must sign its own contract with Anthropic | A separate legal agreement, not just a separate cost centre. Each organisation is its own rollout |
Identity: its identity provider can't join your main sign-in | For example, a separate tenant that won't federate. One organisation per identity boundary |
Data isolation: it must be walled off from the rest of the company's data | By regulation or contract, confirmed by your compliance owner |
Different caps, connectors or a stricter posture for one unit are not reasons to split. They're group-and-role work. If you change this later, you rebuild: nothing crosses the boundary, so splitting or merging means re-provisioning everyone who moves. Adding a separate organisation later is cheaper than pulling apart one you over-consolidated.
How people arrive: JIT or SCIM
Both run through your identity provider, and SSO comes first. Just-in-time (JIT) provisioning creates a member the first time they sign in, but groups don't sync, so you maintain them by hand. Directory sync (SCIM), available on Enterprise plans, pushes members and their direct group memberships ahead of time, so when people change teams, Claude follows. For any organisation above a few dozen people, we recommend SCIM: a control is only as accurate as the membership list underneath it.
Groups, roles and the union rule
A group is a named set of members, usually synced from your directory. A role is a set of permissions attached to a group. Set access by role, not person by person: change a role and everyone in that group changes at once. One group structure spans every surface your contract covers.
A custom role can set four kinds of access: capabilities (for example Claude Code or Cowork), model access, connector permissions and, for admin roles, admin permissions. These only apply to members on a custom role. Members on built-in roles keep everything that's switched on for the organisation.
Then there's the rule that catches most teams out. The union rule: when a member is in more than one group, they get the union of every group's permissions. A narrower group can't remove what a broader group grants. If an engineer in a regulated team also sits in the company-wide Engineering group, they inherit Engineering's broader connector access. To hold a tighter boundary, give those people their own group and keep them out of the broad one.
Group pattern | Choose it when | What it means |
|---|---|---|
Mirror the org chart | Controls only need to vary by department, not by risk | One group per department. Quick to set up, but a risk boundary that cuts across departments has nowhere to attach |
Tier by risk | A regulated function needs tighter controls than its parent department | A few groups by sensitivity. Everyday controls that vary by department have no group shaped for them |
Hybrid | Everyday controls fit departments, but one boundary doesn't | Department groups plus a dedicated carve-out group. Where most organisations with a regulated function land |
If you change this later: adding or renaming groups is low-stakes. Changing the live mapping between directory groups and Claude roles is not, because it changes access for everyone it covers at once.
How we handle it: we design the group structure with your identity team before any role mapping goes live, and test it against the union rule with real people: "Who sits in two groups, and what do they end up with?" That one exercise prevents most access surprises in regulated teams.
Decision 2: Access to surfaces and connectors
Your decision: which surfaces each group gets and when, and which connectors each group gets, read-only or read-write.
Surfaces: where people use Claude
A surface is a place people use Claude: Claude chat, Cowork, Claude Code or Claude for Microsoft 365, depending on your contract. Access works in two layers, and both must allow it. The organisation-wide toggle is the ceiling: off means nobody has it. The role grant decides who has it under that ceiling.
Claude Code needs one more step. Because it runs in your engineering environment, it also has a boundary configuration in Claude Code managed settings, usually written by your platform or engineering lead, who knows which repositories and network domains it should reach.
Approach | Choose it when | What it means |
|---|---|---|
Broad first, specialised where needed | Your objective doesn't call for holding back | General-purpose surfaces go to most groups first. Specialised ones go only where the work needs them |
Pilot, then widen | Your objective calls for testing with one team first | One pilot group gets access. What you learn shapes the broad rollout |
Phased in behind a review | A constraint means a group must wait | The grant arrives after the review clears, and the constraint holds in the meantime |
Models and effort. Every new conversation starts on a default model, and you control it, organisation-wide or per role. You can also limit which models a group can use and cap the effort level. Most people use whatever model they start on, so the default steers most of your spend. Coding teams may need the most capable model. Routine drafting usually doesn't.
If you change this later: granting a surface later is cheap. Taking one away is expensive, because the group loses a tool mid-work.
Connectors: the value and the risk
Connectors give Claude access to other applications, such as your drive, wiki or ticket tracker, so people bring live context into a conversation instead of pasting it in. That reach is the value and the risk at once, so connector access is controlled at three gates:
Gate | What it controls | Set by |
|---|---|---|
Organisation | Whether the connector is available in your organisation at all | Owner or Primary Owner |
Role | Whether a group's role includes the connector, and at what depth | Owner or Primary Owner |
Member | Whether each person has connected their own account | The member, or centrally with enterprise-managed authorization |
For each connector, or each tool within it, a role sets Always allow, Needs approval or Blocked. Blocking the write and delete tools keeps the connector read-only. New custom roles start at Needs approval, so check what existing roles allow before you narrow them. Claude also inherits each person's own permissions in the connected system: if they can't see a file today, Claude can't see it for them.
Enterprise-managed authorization (Anthropic now calls it auto-connect) closes a common gap. Without it, you can scope a connector to a group, but it's only live once each person connects it. With it, access is provisioned through your identity provider, so everyone in scope gets the connector on first login. It's available only for connectors whose provider supports it, so check your list.
Choice | Choose it when | What it means |
|---|---|---|
Read-only, scoped to the groups that work in that system | Most organisations, most connectors | Immediately useful, lower risk, and no question about what Claude might change while you learn |
Write, phased in with sign-off | A group's workflow clearly needs Claude to write, and the risk owner agrees | Starts read-only, adds write once the need is proven. The risk sits with the owner who accepted it |
How we handle it: we start almost every client read-only, then add write access workflow by workflow, each one documented and signed off by the risk owner. Connectors are the most visible change in a rollout, so we announce each one and explain what changes.
Planning a Claude rollout? Book a free 30-minute consultation. We'll review your five decisions, flag the settings that are hard to undo, and outline your first 30 days. Book your consultation →
Decision 3: Governance of skills, plugins and projects
Your decision: how freely people create and share customisations, and whether one posture covers the whole organisation.
Three kinds of customisation answer the same question at different layers: how much of what one person sets up can reach everyone else?
Skills and plugins. A skill packages a workflow so Claude runs it the same way every time: a first-pass contract review against your legal playbook, a report in your house format, an incident review in your template. A plugin bundles skills with other components, such as connectors, so a toolkit installs in one go. A plugin can't give anyone a connector or capability their groups don't already include.
Projects. A workspace with its own context and files. Project sharing settings decide how widely a project can be shared.
Organization instructions. Up to 3,000 characters of always-on guidance included in conversations: who you are, how you're structured, tone for customer-facing writing, or rules such as citing the source document when quoting internal material.
Sharing creates real value: a skill that codifies how your best team works spreads that practice to everyone. The risk is sprawl: dozens of near-duplicate skills, untested workflows that others assume were reviewed, and sensitive data exposed by an over-broad share.
Posture | Choose it when | Settings in practice |
|---|---|---|
Open build, reviewed spread | You want free experimentation and have a centre of excellence or champions to vet what spreads | Anyone builds for themselves. Promotion to the whole organisation goes through review (skill sharing on, publishing "Requires review") |
Centralized | A staffed central team already owns how internal tools are built | The central team builds and vets everything. People install what's published (sharing off, publishing off, Owner provisioning only) |
Fully open | A small or high-trust organisation where review would only slow people down | Anyone builds, shares and installs. Vetting is social (all sharing on, publishing "Open") |
Approve-first (for a regulated group) | One group needs tighter control than the rest | Nothing built in that group is shared until a named reviewer approves it |
Customisations move through four distribution controls: Owner provisioning (an Owner adds a skill for everyone), skill sharing with named colleagues, sharing with groups, and publishing to the organisation library (Off, Requires review or Open). Skills also need code execution and file creation switched on, so check both settings together.
If you change this later: loosening is free. Tightening needs care, and two settings behave differently: turning off public projects makes existing public projects private, while a stricter publishing setting doesn't remove what's already published.
How we handle it: we match the posture to how you already govern tools your people build, and we set up the review gate with named reviewers, a simple quality checklist and a skills library that's curated, not crowded. We also draft your first organization instructions, usually three to five lines.
Decision 4: Spend caps and managing spend
Your decision: where caps sit, who approves more, and what happens before a cap moves.
How Claude Enterprise is billed
Claude Enterprise is billed as a per-seat fee, which covers access, plus usage measured in tokens at standard API rates. Check your own contract, and read Anthropic's Claude Enterprise consumption guide, which recommends rolling out agentic surfaces in waves with conservative caps. Agentic surfaces such as Cowork and Claude Code do more work behind the scenes, so they can use more per session. The key point for finance: usage, not just headcount, drives the bill, and the pool is shared, so one group's heavy month draws from everyone's.
Three levels of cap
Level | What it does |
|---|---|
Organisation ceiling | The most your organisation can use in a month. No group cap or override goes above it |
Group caps | A per-member limit every member of a group inherits, sized to that group's typical usage. You can also add a pooled monthly budget for the group |
Per-member overrides | One person's own limit, above or below their group's |
Members with no limit anywhere are uncapped. At the cap, a person's usage pauses until the month resets or a request is approved. When someone is in two capped groups, one organisation setting decides whether the higher or lower cap wins. Requests land with Owners and admins. Decide the escalation split when you set caps, not after the first request: who approves routine increases the same day, and what goes to the budget owner.
Size caps with consumption tiers. Define light, standard and power tiers by role type before rollout, so caps are assigned and adjusted consistently. A group cap answers "how much does one member of this group typically need?". It isn't a pooled budget. If usage is unknown, start tight and raise caps on evidence. If you see a growing list of overrides, those people need their own group.
Move spend before you raise a cap
Default model and effort level. If the default is the most capable model at high effort, that's your most expensive setup, and it's where most usage runs. A mid-range default, moderate effort and effort caps for routine work move more spend than most caps.
Organization instructions. Guidance at the point of use, such as "keep answers short unless asked" or "work from the files the task needs".
Reusable skills. A skill loads only when it's needed, which is cheaper and more consistent than re-sharing a long prompt.
Agentic habits. Tasks left running and scheduled Cowork tasks keep drawing from the pool. Make sure people know.
Review spend monthly, weekly in the first months or when a new surface launches. Look both ways: who's hitting caps, and who isn't using their seat. If you change this later: caps and defaults change any time, from that point on.
How we handle it: we build the cap table and escalation split with your finance partner before go-live, then run the first two monthly spend reviews with you. Our most common finding is the most capable model running routine work. Changing a role's default fixes it without touching the cap. We cover cost ownership in who owns the Claude budget.
Decision 5: Visibility, retention and adoption
Your decision: what retention window you set, whether the Compliance API is on and who reads it, and how you'll know adoption is on track.
The first four decisions configure what Claude does. Visibility tells you whether they're working, and lets you answer questions after the fact with evidence. Set it up before people get access. The gap between first access and switching recording on can't be filled later.
Record | What it answers | Use it for |
|---|---|---|
Compliance API | What was said: conversation content, files and activity events, including Cowork and Claude Code sessions. Only the Primary Owner can turn it on | Investigations, legal and regulatory requests, security review |
Audit logs | Who did what, when. Metadata only, no prompts or responses. Exportable for the last 180 days | Admin and access changes |
OpenTelemetry | How surfaces are running: usage patterns and performance | Your existing observability stack (optional) |
Analytics dashboard and API | Adoption and spend by group, model and person | Monthly reviews, chargeback, adoption tracking |
Retention is organisation-wide. By default, content is kept until people delete it. A custom window deletes content automatically, and the shortest is 30 days: there's no zero-retention option. Check which products your window covers: some newer surfaces, such as Claude Tag, are handled separately. If a regulated function needs a shorter window, everyone gets it. Set it against your existing data policy, with the policy owner's approval. If you change this later: the setting moves freely, but deleted content doesn't come back.
For content that should never reach Claude at all, propose inference hooks to your security team: every prompt goes to a server they control for an allow-or-deny decision before Claude answers.
Adoption signals: is the rollout working?
Signal | What it tells you | If it's low, look at |
|---|---|---|
Active members by group (breadth) | Whether adoption is spread or concentrated | Access and awareness: does the group have Claude, and was it properly enabled? |
Members returning week over week (breadth) | Whether people come back | Onboarding: what did the first session promise? |
Conversations per active member (depth) | How much people rely on Claude | Fit: do surfaces and connectors match their work? |
Skills and projects in use (depth) | Whether work compounds into reusable workflows | Governance: is the posture letting good work spread? |
Connector use (depth) | Which tools matter most | Connector scope: does it let connectors create value? |
Set a pace goal you can read on the dashboard. "Every granted group returning weekly within eight weeks" is a pace. "Good adoption soon" is a wish. Pair these leading indicators with real work outcomes, as we describe in our 90-day rollout plan.
How we handle it: we turn on the Compliance API and route it to a named reviewer before wave one, agree retention with your data-protection lead, and set the pace goal with business-unit leaders. When a group goes quiet, we diagnose the cause within days: usually access, awareness or a first session that didn't fit their work.
Free download: the Claude Enterprise Rollout Kit. A decision workbook for all five decisions, the 90-day plan, and an editable decision record and Excel tracker. Get the kit →
How the five decisions connect
Structure comes first, the next three attach to its groups, and Visibility reads across all of them. That's why a change ripples. Add a new group and it needs a surface grant, connector scope, a cap and a dashboard view before anyone in it goes live. Move one person into a regulated group, and their surfaces, connectors, governance posture and cap all change with them.
Regulated functions: one posture, not five exceptions
For a regulated team in finance, health or the public sector, check that its five answers form one consistent posture:
Structure: its own group, so the union rule gives it nothing broader than its own role.
Access: the same everyday surfaces as its neighbours, with specialised ones such as Claude Code held until visibility reporting is in place, and connectors read-only.
Governance: approve-first, while the rest of the organisation builds freely with a reviewer for anything that spreads.
Spend: a group cap sized like any other, with every increase reviewed.
Visibility: the Compliance API on with a named reader, and retention set to the regulated team's requirement for everyone.
One group handled through the ordinary system, with no special machinery. That's why it holds. If your whole organisation is regulated, that set is your baseline.
Turn the decisions into a kickoff plan
Review the hard-to-undo settings. Domain claiming, number of organisations, group mapping and retention: each decided, or on an owner's desk with a decide-by date.
Close open items. For each open decision, write what's blocking it, what unblocks it and a date. An owner's name alone is not a plan.
Sequence the plan. Structure → Access → Governance → Spend → Visibility on → go-live by group, each with an owner and a start date.
When a new Claude product arrives
Anthropic ships new surfaces often. Most of your configuration carries over, because it reflects your objectives, not the product: your organisation boundary, groups and roles, governance posture, caps and visibility. For anything new, ask three questions:
Who gets it? Which groups' daily work calls for it, and in what phase?
What does it carry? Does it bring its own settings or connectors that need fresh decisions?
Does it move risk? Does it give Claude new reach: a new class of data, more autonomy, or new people in front of it? If so, involve the risk owner.
Claude Tag, launched in beta in June 2026 so teams can work with Claude inside Slack, is a good example. It works under its own identity in a channel, with the channel's access, so everyone in that channel gets the same capability. Channel usage bills to the organisation with its own limits. That's new reach and a new spend line, so it goes to both the risk owner and the budget owner before you switch it on.
How the five decisions fit the 90-day rollout plan
The five decisions are the configuration half of a rollout. Our 90-day rollout plan is the people and value half. They fit together like this:
Phase | What happens | The five decisions |
|---|---|---|
Days 0–30: lay the foundations | Owners, prerequisites, security review, baseline | All five decided and configured, hard-to-undo settings signed off, Compliance API on before access |
Days 30–60: activate | Wave one trained, champions, weekly measurement | Surfaces and connectors phased in by group, first skills through the review gate, first spend review |
Days 60–90: prove and decide | Value against the baseline, readout, scale decision | Adoption signals read against the pace goal, caps re-based on real usage |
After day 90: scale | Roll out in waves | New groups go through the cascade, and new products go through the three questions |
For the leadership view, why Claude transformations stall and how to move from pilot to scale, see the Enterprise AI Transformation Guide.
Claude Enterprise in production: public examples from the US, UK and Canada
Claude is already running at scale in organisations of every size and in government. These are public deployments reported by Anthropic, its customers and government bodies. They are not Generation Digital clients, and the figures are as each organisation reported them.
Organisation | Region and sector | What's public |
|---|---|---|
US and global · professional services | Claude available to more than 470,000 people, with a Claude Centre of Excellence (October 2025) | |
North America · insurance | Claude Enterprise for its 20,000+ employees, reporting an average of 2.5 hours saved per employee per week (February 2026) | |
North America · software, mid-market | 89% active usage among licensed employees within eight weeks, then licences doubled from 1,000 to 2,000 | |
North America · security software | 88% weekly active usage across about 1,500 employees, with a company-wide Cowork rollout in 17 days | |
US and global · consulting | Claude for all 19,000 employees, with more than 7,000 active within weeks of access (August 2026) | |
UK · banking | A knowledge assistant used by more than 16,000 colleagues, and Claude Code for half of developers by the end of 2026 (October 2026) | |
UK · financial services | Claude Enterprise company-wide, reporting 70 hours saved weekly for analysts and a payback period under three months | |
UK · professional services | Claude for all UK partners and staff, rolled out from June 2026 | |
Australia · data and AI, mid-market | Claude Enterprise for all 1,200 employees, with 91% weekly usage | |
Norway · marketing, mid-market (400+ staff) | Reports 30% less time on routine tasks and 40% faster proposal development | |
US · public sector | About 500 government organisations use Claude, and the GSA OneGov agreement extends discounted Claude Enterprise pricing to federal, state, local and tribal buyers (October 2026) | |
US · state government | A statewide agreement used by departments including the DMV, with the same pricing for cities and counties (July 2026) | |
UK · central government | A GOV.UK agentic assistant pilot with the Government Digital Service, starting with employment support (January 2026), following a 2025 memorandum of understanding | |
Canada · provincial government | Uses Claude to test provincial applications against security and privacy standards (July 2026) |
The pattern behind the fastest results is consistent: whole functions rolled out in waves, connectors to the tools people already use, and governance designed in from the start. That's the five-decision method in practice.
Common Claude Enterprise rollout mistakes
Claiming the domain before everyone is in the identity provider, and locking people out.
Only one Owner, assigned through a group that a sync error can remove.
Splitting into several organisations for reasons that groups and roles could handle.
Ignoring the union rule, so a regulated team inherits broad access through a company-wide group.
Granting write access to connectors on day one, before a workflow needs it.
No review gate for skills, so untested workflows spread and people assume they were checked.
The most capable model as everyone's default, then raising caps to pay for routine work.
Turning on the Compliance API after go-live, leaving a gap you can never fill.
Measuring logins instead of work, so the day-90 readout has nothing to prove.
Do you need a Claude partner?
You can make these decisions yourself, and this guide and our free kit are designed to help. A certified partner earns its place by reducing risk and getting you to value faster:
Going it alone | With Generation Digital | |
|---|---|---|
The five decisions | Learned as you go, often after a setting is hard to undo | Made in week one from a method we've run many times, with every hard-to-undo setting signed off |
Security and compliance | Security reviews stall on unknowns | Evidence ready for your security review: identity, connectors, retention, Compliance API |
Adoption | A spike, then a long tail of unused seats | Role-based training, champions and fast help for quiet teams |
Spend | Caps raised to fix symptoms | Consumption tiers, model defaults and a monthly review with finance |
Proving value | No baseline, so no clear answer at day 90 | A baseline at day 0 and a readout your sponsor can stand behind |
Why Generation Digital
A certified partner in Anthropic's Claude Partner Network, using the rollout approach Anthropic recommends.
Over 200 clients across mid-market, enterprise and public sector in the US, UK and Canada on AI transformation.
Human and business-led. We combine configuration with training, change management and adoption, because seats only create value when people use them well.
Every size of organisation. Enterprise and mid-market rollouts, plus right-sized support for smaller businesses through Claude for Small Business.
Public sector ready. UK public bodies can buy our services through G-Cloud 15.
Local teams in London, New York and Toronto.
Explore our Claude Enterprise rollout service, Claude training, Claude Code enablement and Claude governance work.
What happens in a 30-minute consultation
Your objective and constraints: what success looks like and what must hold.
Your five decisions: where you are on each, and which hard-to-undo settings need attention now.
Your first 30 days: the owners, prerequisites and sequence to get to wave one safely.
You leave with a clear next step, whether or not you work with us.
Planning a Claude rollout? Book a free 30-minute consultation. We'll review your five decisions, flag the settings that are hard to undo, and outline your first 30 days. Book your consultation →
Frequently asked questions
What are the five decisions in a Claude Enterprise rollout?
Structure & Identity (how many Claude organizations you run and which groups live inside them), Access (which surfaces and connectors each group gets), Governance (who can build and share skills, plugins and projects), Spend (where caps sit and who owns the escalation) and Visibility (what you can measure and how long data is kept). Make them in that order, because each decision scopes the ones after it.
Which Claude Enterprise settings are hard to undo?
Four: domain claiming, which can't be reversed; organization topology, because splitting or merging organizations means re-provisioning members; the group-to-role mapping, because changing it shifts access for everyone it covers; and data retention, because content deleted under a shorter window is gone for good.
How long does a Claude Enterprise rollout take?
Most mid-market and enterprise rollouts reach a scale decision in about 90 days. The five configuration decisions and the prerequisites sit in the first 30 days, before most people get access. Smaller organisations with a simple identity setup can move faster.
Who should own a Claude Enterprise rollout?
The Claude Owner, usually in IT, leads the rollout and holds most of the decisions. Two decisions usually escalate: Spend goes to whoever commits the budget, and Visibility goes to whoever answers for data risk, such as security, legal or compliance. Each decision needs a named owner before work starts.
Should we run one Claude organization or several?
One organization is the default. Only three situations justify a separate one: part of the company must sign its own contract with Anthropic, its identity provider can't join your main sign-in, or it must be walled off from the rest of the company's data by regulation or contract. Tighter controls for one unit are handled with groups and roles, not a second organization.
What is the union rule in Claude Enterprise?
When a member belongs to more than one group, their permissions are the union of every group's role. A narrower group can't take away what a broader group grants. To hold a tighter boundary, put those members in their own group and keep them out of the broad group.
Should Claude connectors be read-only or read-write?
Start read-only for most groups. Read access is immediately useful and lower risk. Add write access later, group by group, once a workflow clearly needs it and the risk owner signs off. Claude also inherits each member's own permissions in the connected system.
How do you control Claude Enterprise costs?
Set an organization ceiling, group caps sized to typical usage and a few per-member overrides, then use the levers that move spend before you raise a cap: a mid-range default model, sensible effort levels, organization instructions and reusable skills. Review spend monthly, and treat every cap-increase request as a prompt to look, not an automatic yes.
When should we turn on the Compliance API?
Before members get access. The gap between first access and turning recording on can't be filled later. Route it to a named reviewer in your existing security or compliance review, and set retention to match your existing data policy.
Do we need a Claude partner to roll out Claude Enterprise?
No, but a partner reduces risk and time to value. A certified partner has made the five decisions many times, knows which settings are hard to undo, and brings the training and change programme that drives adoption. Generation Digital has worked with over 200 clients on AI transformation across the US, UK and Canada.
Does this approach work for public sector and regulated organisations?
Yes. A regulated function is handled through the same system as everyone else: its own group, narrower access and connectors, an approve-first governance posture, reviewed spend, and the Compliance API on with retention set by the risk owner. UK public sector bodies can buy Generation Digital's Claude services through G-Cloud 15.
What's in the free Claude Enterprise Rollout Kit?
A decision workbook for the five decisions (owners, prerequisites, groups, surfaces, connectors, governance, caps and visibility), the 90-day rollout plan with gates, an editable decision record in Word, and an Excel tracker. It's free from Generation Digital.
Get your Claude Enterprise rollout right first time
Download the free Claude Enterprise Rollout Kit to work through the five decisions with your team, or book a 30-minute consultation with a Claude specialist at Generation Digital.
Sources
Anthropic, Claude help centre articles on domain capture, SSO and SCIM, role-based permissions, group spend limits, data retention, the Compliance API, audit logs, connectors, skills and organization instructions (accessed 22 Nov 2026): support.claude.com
Anthropic, Claude Enterprise consumption guide (accessed 22 Nov 2026)
Anthropic, How am I billed for my Enterprise plan? (accessed 22 Nov 2026)
Anthropic, Claude Code settings and monitoring (accessed 22 Nov 2026)
Anthropic, Introducing Claude Tag (June 2026)
Anthropic, Deploying Claude Enterprise with Confidence (Claude Partner Network training), adapted by Generation Digital
McKinsey, The state of AI in 2025 (November 2025); Gartner (June 2025); BCG, AI at Work 2025 (June 2025); Microsoft and LinkedIn, Work Trend Index (May 2024)
Customer and government references: as linked in the table above (accessed 22 Nov 2026)
About the author: Graham Mackay is Founder and CEO of Generation Digital, a certified partner in Anthropic's Claude Partner Network. He leads AI transformation programmes for mid-market, enterprise and public-sector organisations across the US, UK and Canada.
Aplique esto en su negocio
¿Está implementando Claude o inteligencia artificial en su empresa?
En Generation Digital somos socios certificados de la red de partners de Anthropic Claude. Hable con uno de nuestros especialistas para planificar su implementación, capacitar a su equipo y facilitar la adopción.
Recibe noticias y consejos sobre IA cada semana en tu bandeja de entrada
Al suscribirte, das tu consentimiento para que Generation Digital almacene y procese tus datos de acuerdo con nuestra política de privacidad. Puedes leer la política completa en gend.co/privacy.

