Foundry Agent Production Readiness
Decide whether a Microsoft Foundry workload is ready for production, with evidence per gate.
See the engagementView the Microsoft Marketplace listingPRIVATE AI
Choose the AI boundary before choosing hardware.
You do not need to understand how the model works. You need to know what it can see, where it runs and who decided that. Managed Private AI chooses the boundary on the data rules before anyone chooses hardware, then runs a governed pilot inside it.
01 Source02 Boundary03 Human decision04 Evidence
HOW IT WORKS
The control path in practice.
- Classify the task and the data it needs.
- Compare tenant-private and local processing boundaries.
- Run a sample evaluation on the selected architecture.
- Document the operating owner, cost model, and recovery path.
WAYS TO START
The fixed engagements behind this page.
Governed AI Briefing
Security, legal and the business leave with one question, the pilot that answers it and the scoping call booked.
See the engagementTHE DECISION
What is private AI, in plain words?
AI that works on your documents inside a boundary you chose: your Microsoft tenant, a private deployment in Azure, or a device in your office, with the choice made on the data rules, not on the hardware.
You do not need to understand how a model works. You need to know what it can see, where it runs, who decided that, and how you would prove it. Managed Private AI is the method that answers those four questions before anything is bought.
THE MICROSOFT INSIGHT
Does every Microsoft AI feature keep data inside the same boundary?
No, and the differences are documented, so the boundary decision has to be made feature by feature.
Models provided by Anthropic as a Microsoft subprocessor, used in Copilot experiences including Cowork, are currently excluded from the EU Data Boundary, and tenants in the EU and the UK have them disabled by default. (Microsoft Learn, checked Oct 31, 2026)
Read the subprocessor and data boundary facts before the pilot and the client confidentiality committee gets a one-page answer. Read them after and the committee gets a surprise.
THE METHOD
How do you choose between tenant-private and on-premises AI?
By testing the workload, its data rules, the operating model and the Microsoft integration against each option before any purchase.
- Name the workload and the people who use it, and classify the data it touches.
- Test the tenant-private path first: Copilot, Cowork and Microsoft Foundry under your existing permissions and labels.
- Test a local or dedicated option only where the data rules require it, with the operating cost written down.
- Your written decision on the boundary, then a scoped pilot inside it.
THE GATE
Who sees the AI output before a client does?
A named professional, every time, because the permission row stops at may draft for anything that leaves the workspace.
Retrieval is limited to what the signed-in person may already open. Drafts carry their sources. A lawyer, a partner or a department head approves distribution and the approval is recorded with the document.
With the gate, privilege review is a step you can show. Without it, the first client-facing error is the one that defines the program.
THE FIRST RUNG
Where does a firm start if a ban on consumer AI tools is not working?
With a one-hour private AI briefing, then the free readiness check, then a one-day workshop and a four-week pilot inside the boundary you chose.
A ban moves the problem to personal phones and removes your view of it. Microsoft Defender for Cloud Apps can discover and block consumer AI apps on managed devices while a governed alternative is put in front of your people, which is the sequence the Hartwell Trust Bank episodes show on fictional data.
AFTER TWO WEEKS
What will be working after two weeks?
- The workload, its users and its data classes are written down and agreed.
- The tenant-private path has been tested on sample documents under your labels and permissions.
- Each boundary option has its operating model and its cost of ownership described, with no hardware ordered.
- You hold the boundary decision as a written recommendation with the pilot scope and its quote.
The work is delivered remotely inside your own Microsoft tenant, wherever your business is, and on site only when the work needs it.
HONEST ABOUT LIMITS
What this does not do yet
It does not stop a person pasting client text into a public tool on an unmanaged device. Microsoft Defender for Cloud Apps lists generative AI apps as a category, and an app tagged Unsanctioned is blocked on devices onboarded to Microsoft Defender for Endpoint, with up to three hours of propagation. (Microsoft Learn, checked Oct 31, 2026) Blocking works on devices onboarded to Microsoft Defender for Endpoint, so the boundary includes a device decision, not only a model decision.
It does not buy hardware or commit you to a model vendor: the method tests the boundary first and the design does not depend on one vendor's SDK.
It does not make a confidentiality claim for you: the pilot produces the evidence and your general counsel decides what the firm asserts.
THE ONE THING TO REMEMBER
Choose the AI boundary before choosing hardware, and make the choice on the data rules.
Said another way: the question is where the data may go, not which model is cleverest.
Said a third way: a boundary you can explain to a client is worth more than a server you can point at.
SHOW YOUR WORK
See the control path on a fictional bank.
Shadow AI at the first gate
Hartwell Trust Bank checks an outside AI use case with fictional data.
Coming
See every demonstrationDIRECT ANSWERS
Questions before you decide.
Does private AI mean on-premises hardware?
Not by default. A tenant-private path inside Microsoft 365 and Azure is tested first. A local or dedicated option is chosen only where the data rules require it and the operating cost is accepted in writing.
Can our data be used to train a model?
The pilot documents, for each feature, what the Microsoft terms say about training and retention, with the Learn source and the date, so the firm decides from the record rather than from a sales claim.
What happens to a banned consumer AI tool?
On managed devices it can be discovered and blocked with Microsoft Defender for Cloud Apps while a governed alternative is introduced. The ban alone only moves the use to personal devices.
How long until something is in use?
A one-hour briefing and the free check take a day. A one-day workshop sets the boundary. A four-week pilot puts a governed workflow in front of a small group, with the operating record kept.
How is it priced?
Quoted after a scoping call. The briefing, the workshop and the pilot are separate scopes, and nothing is widened until you decide in writing.
NEXT STEP
Start with a boundary you can explain.
Choose a focused check, or speak with the team about the workflow you need to improve.