Enterprise AI rollout checklist — SOC 2, ISO 27001, DPAs

Before anything else

What specific task is this replacing, and how will you measure it?
Who owns the rollout, and who reports the numbers?
What data will people realistically put into it — not what the policy says

The third question is the one that determines everything downstream. If the honest answer includes customer data or anything confidential, the contract requirements change completely, and you should establish that before you shortlist.

Vendor certification

Item	What to ask for	Why it matters
SOC 2 Type II	The current report, under NDA	The baseline for third-party risk management
ISO 27001	Certificate and scope	Often required where SOC 2 is less familiar
ISO 27017 / 27701	Where cloud or privacy scope matters	Relevant if you hold your own certification
Penetration testing	Summary of the most recent test	Shows whether security is operational or paperwork
Subprocessor list	Current list, with notification terms	You inherit their subprocessors

The data processing agreement

This is where deployments are actually won or lost. A marketing page saying your data is not used for training is not the same as a contractual commitment, and the difference becomes obvious the first time legal reads both.

Exclusion from model training, stated in the agreement rather than the FAQ
Retention period, and what happens to data at termination
Processing location, and whether you can choose a region
Incident notification timelines
Subprocessor change notification and your right to object

Identity and audit

SSO via SAML or OIDC against your existing identity provider
SCIM provisioning, so leavers actually lose access
Audit logs your SIEM can ingest, not just a dashboard
Admin roles that match how your organisation delegates

SCIM is the one most often skipped at pilot stage and most often regretted. Without it, account lifecycle becomes manual and your access review will find accounts belonging to people who left.

Personal data

If personal data will touch the tool, you need a lawful basis, a record of processing, and an answer for data subject requests that includes whatever the AI vendor holds. Under Japan's Act on the Protection of Personal Information, sensitive categories — medical history among them — require explicit consent before they go anywhere near a third party, and providing data to a vendor outside Japan carries its own disclosure and consent obligations.

What changes if the entity is in Japan

Whether there is a Japanese contracting entity, or you buy through a reseller
Whether invoice payment in yen is available — card-only is a blocker at many companies
Whether a Tokyo region exists, if residency is required
Whether support operates in Japanese, and in Japanese business hours

None of these are capability questions, and all four can stop a deployment that passed evaluation. Establish them early; they are cheap to ask and expensive to discover late.

Pilot structure

Twenty to thirty people, one month, one measurable task
A written policy on what may be entered, published before day one
A weekly channel for what worked and what did not
A decision date, agreed in advance, with the criteria written down

Can we run a pilot before legal review?

Only with non-confidential data and an explicit written scope. The risk is that a successful pilot creates pressure to scale before the contract questions are answered, and that is how shadow usage starts.

Business plan or enterprise plan?

Business plans usually cover training exclusion and basic SSO. Enterprise adds SCIM, audit logs, data residency choices and a negotiable DPA. If you need any of those four, the business plan will not survive review.

What about building on Azure OpenAI or Bedrock instead?

That is the right path when you are building an application or when residency in a specific region is mandatory. You take on the interface, the support burden and the model lifecycle in exchange.

Enterprise AI rollout checklist — SOC 2, ISO 27001, DPAs and personal data