AI adoption strategy: the first 90 days with Claude

TL;DR

An AI adoption strategy for Claude is mostly identity work and three settings you cannot change later. Spend the first 30 days on provisioning, groups and spend limits, the next 30 measuring active use rather than licences, and the last 30 handing operations to a named owner. Skip the identity work and the rollout stalls at week six.

What a 90-day Claude rollout actually sequences

A Claude rollout is not a training problem. It is a provisioning problem with a training phase bolted on. The order that works puts the irreversible platform settings first, identity second, cohorts third and measurement throughout. The order that fails starts with a pilot group signed in through personal invites, because every later step then has to undo that.

The reason is mechanical. Claude decides how members join your organization through one setting, and that setting determines whether you can ever automate joiners and leavers. Anthropic’s provisioning documentation lists three modes:

Mode	Who gets added	Who gets removed	Plan availability
Invite only (default)	Admins add members by hand	Manually	All plans
Just-in-time	Users assigned to the Claude app in your IdP, on first login	Manually only. JIT never removes members automatically	Team, Enterprise, Console organizations
SCIM directory sync	Automatically from IdP assignment, without a login	Automatically when unassigned in the IdP	Enterprise plans and Console organizations with a parent or joined Enterprise organization

Invite only is the default, and it is the wrong default for anything past 20 people. SCIM is not available on Team plans, which makes the plan decision an identity decision rather than a budget one.

Which decisions are irreversible, and must come first

Four settings bind you for the life of the deployment. Decide them in week one, in writing, with the owner named.

1. The plan. Team tops out at 150 seats and has no SCIM, no custom roles, no Compliance API and no audit logs. Enterprise starts at 20 seats and adds all of them. Moving later means re-provisioning every member.
2. The provisioning mode. Switching from invite only to SCIM mid-rollout means reconciling two member lists by hand. Resetting a SCIM connection is destructive.
3. The email attribute. The field your IdP sends for SCIM and the field it sends in the SAML assertion must be the same, as our SSO and SCIM guide sets out. Anthropic’s Entra ID guide calls a mismatch here “the most common cause of login failures”, and the Okta guide names the usual culprit: SCIM reading user.login while SAML reads user.email.
4. The inference route. Whether prompts run on the first-party Claude API, on Bedrock, on Google Cloud or on Foundry decides residency, contracts and network design. Our data residency guide covers what each route actually commits you to.

Everything else, including which teams go first and what the training looks like, is reversible and should not slow week one down.

Days 1 to 30: identity, groups and the first cohort

The first month is plumbing. You need a verified domain in Claude’s identity settings, an Owner or Primary Owner to do the work, and the configuration values from the WorkOS setup flow that Claude uses for SAML and SCIM.

Order of operations that avoids rework:

1. Verify the domain and configure SAML SSO with the email attribute fixed.
2. Add the SCIM connection using the same email field, and assign only the pilot group in the IdP.
3. Create the groups you will govern by, because model defaults and spend limits can be set by group.
4. Turn on usage credits and set an organization default spend limit before the first member signs in.
5. Add the pilot cohort to the IdP group, and confirm they appear in Claude without a login.

One trap deserves its own line. If you run just-in-time provisioning with group mappings, a user whose sign-in carries no mapped group is removed from the organization at their next login. An incomplete group claim therefore looks like a random deprovisioning event. If you use group mappings on Okta, the SAML configuration needs a group attribute statement named groups.

The pilot cohort size that produces usable data is 15 to 30 people from two different functions. Smaller than that and the usage numbers are noise. Larger and you are rolling out, not piloting.

Days 31 to 60: measure adoption, not licences

Seats assigned is a procurement metric. It tells you nothing. The three numbers that predict whether a Claude programme survives its first budget review are weekly active members, tasks completed per active member and spend per active member.

Metric	Where it comes from	What a bad number looks like
Weekly active members	Organization analytics, filterable by SCIM group	Under 40 percent of assigned seats by week eight
Daily active developers and sessions	The Claude Code analytics dashboard	Developers who ran one session and never returned
Spend per member, per model	Spend report CSV, one row per person per model, refreshed daily with a one-day delay	Opus traffic on routine work, which means the model default was never set
Effective spend limit and period-to-date spend	GET /v1/organizations/spend_limits/effective on Enterprise	Everyone on null, which means unlimited

The spend report is the one that changes behaviour, because it is per person and per model. It appears once usage credits are turned on, and usage inside the seat allowance is not metered in dollars. On Team and Enterprise plans each member’s allowance resets on a rolling five-hour window and a weekly window, and it is shared across Claude chat, Cowork and Claude Code.

Budget differently for a coding seat than a chat seat. Claude Code averages about 13 USD per developer per active day and 150 to 250 USD per developer per month across enterprise deployments, and stays under 30 USD per active day for 90 percent of users. Our Claude Code pricing guide for teams breaks that down per setup.

Days 61 to 90: controls, defaults and the handover

The last month converts a pilot into an operation. Three controls do most of the work.

Model defaults. Admins set the default model that new conversations start with in chat, Cowork and Claude Code, and can restrict which models specific roles may use. This is the single cheapest cost control, because it stops routine work defaulting to the most expensive model without anyone having to be told. Our AI FinOps guide covers the rest of the spend levers.

Spend limits by group. Limits resolve from a hierarchy: a per-user override beats a group limit, which beats a seat tier, which beats the organization default. The source field on each member’s effective row tells you which level applied. A group limit is a per-member default, not a pooled group budget, so 50 people in a group with a 200 USD limit can spend 10,000 USD between them. Organization-level alerts reach admins at 75 and 90 percent of the cap; members see in-app notifications at 75 and 95 percent and can request an increase.

A named owner. Not a committee. One person who owns the member list, the spend limits and the model defaults, with a deputy. Our centre of excellence guide describes the four functions that role has to cover before it becomes a department.

What the AI Act asks before you let staff in

The AI literacy obligation has applied since 2 February 2025, and the Act became broadly applicable on 2 August 2026. For a deployer rolling Claude out to staff, that means the training is not optional and it has to leave evidence: who was trained, on what, when, and what the material said about the limits of the system.

This is the part of a rollout that gets deferred to month four and then fails an audit. Fold it into days 31 to 60 while the cohort is small enough to train properly. Our Claude training guide sets out the role-based curriculum and what records to keep, and the AI Act deployer obligations guide covers which duties attach to you rather than to Anthropic.

Run it yourself or outsource?

Task	Keep in-house	Sensible to outsource
IdP configuration and SCIM	Yes. Your identity team owns the IdP and the attribute mapping	Only the first connection, as a paired session
Group and role design	Yes. It mirrors your org chart	No
Model defaults and spend limits	Yes, after the first month	Initial baseline, if nobody has done it before
Training delivery	Partly. Internal champions outperform external trainers after the first wave	First two waves and the material
Adoption reporting	Yes. It is a recurring internal report, not a project	Building the first report and its queries

The honest split: identity and governance stay in-house because they are permanent. Material, baselines and the first training waves are worth buying once.

FAQ

How much does a 90-day Claude rollout cost?

Licences dominate. Claude Enterprise is 20 USD per seat per month billed annually plus usage at API rates with a 20-seat minimum; Team Standard is 20 USD and Premium 100 USD per seat per month annually. For 50 seats on Enterprise that is 12,000 USD a year before usage. Add Claude Code usage at 150 to 250 USD per developer per month for the developers in scope. Internal effort for the first 90 days is typically 20 to 30 person-days across identity, security and the programme owner.

Should we start with Team or Enterprise?

If you need SCIM, audit logs, the Compliance API or custom roles, start with Enterprise, because none of those exist on Team and moving later means re-provisioning. Team is the right answer only below roughly 50 seats with no automated joiner and leaver requirement. See our Team plan guide for where the ceiling sits.

JIT or SCIM: which provisioning mode should we use?

SCIM if your plan allows it. It adds and removes members from IdP assignment without a login, which is the only mode that closes the leaver gap. JIT never removes members automatically, so every departure becomes a manual ticket. JIT is the fallback on Team plans, where SCIM is not available.

What is the most common reason a Claude rollout stalls at week six?

Login failures from a mismatched email attribute between SCIM and SAML, followed by group claims that are missing from the SAML assertion. Both look like random access problems to users and both are configuration errors in the IdP, not in Claude.

Do we need to train everyone before switching Claude on?

Train everyone who gets a seat, before they get it. The AI literacy duty has applied since February 2025, and retrofitting training records to people who have been using a system for months is the part auditors notice. Keep the cohort small enough that this is possible.

Can we roll out Claude Code and Claude chat at the same time?

You can, but budget and measure them separately. They draw on the same seat allowance but behave differently: one Claude Code debugging session can consume more than a day of chat, because each turn carries file contents, tool calls and multi-step reasoning.

Sources

1. Anthropic: Setting up JIT or SCIM provisioning (28 October 2026)
2. Anthropic: SSO setup with Microsoft Entra ID (28 October 2026)
3. Anthropic: Claude pricing plans (28 October 2026)
4. Claude Code: Manage costs effectively (28 October 2026)
5. Anthropic: Spend Limits API (28 October 2026)
6. European Commission: Regulatory framework for AI (28 October 2026)