Shadow AI at the first gate
Hartwell Trust Bank checks an outside AI use case with fictional data.
Coming
See every demonstrationLEGAL | SAMPLE WORKFLOW
Keep matter context within its permission boundary.
You do not need to understand how it works. You need to know what it can see, and who decided that. A fictional firm drafts a matter summary from approved documents: access checks run before retrieval, and a lawyer reviews the cited draft before it leaves the workspace.
01 Source02 Boundary03 Human decision04 Evidence
HOW IT WORKS
The control path in practice.
- Select the matter and confirm its document permissions.
- Retrieve only material available to the signed-in user.
- Create a cited draft with uncertain statements marked.
- Require lawyer review before distribution.
THE WOUND
What can the AI see, and who decided that?
Only what the signed-in lawyer may already open, and the firm decided it, in a permission row the client confidentiality committee can read.
You banned the consumer tools and the associates moved to their phones. A ban does not keep client data from AI; it removes your view of it. The alternative is private, governed and explained in plain English.
THE MICROSOFT INSIGHT
Does blocking a public AI tool on managed PCs stop data leaving?
On devices onboarded to Microsoft Defender for Endpoint, yes: an app tagged Unsanctioned in Microsoft Defender for Cloud Apps is blocked there, with up to three hours of propagation.
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)
Block and offer a governed alternative, and the associate keeps working. Block alone, and the associate keeps working on a personal phone you cannot see.
THE SAMPLE WORKFLOW
How does a cited matter draft stay inside its boundary?
Access checks run before retrieval, the draft cites its sources and marks uncertain statements, and a lawyer reviews it before it leaves the workspace.
- Select the matter and confirm its document permissions.
- Retrieve only material available to the signed-in user.
- Create a cited draft with uncertain statements marked.
- Require lawyer review before distribution, and record it.
THE SIGNAL
How do you know when AI use turns risky?
Microsoft Purview Insider Risk Management ships a Risky AI usage policy template that scores prompts and responses containing sensitive information and browsing to generative AI sites.
Microsoft Purview Insider Risk Management ships a Risky AI usage policy template that scores prompts and responses containing sensitive information in Microsoft 365 Copilot and agents, and browsing to generative AI sites. (Microsoft Learn, checked Oct 31, 2026)
With the policy, a risky prompt becomes an alert with an owner. Without it, it becomes a conversation with a client.
THE FIRST RUNG
Where does a COO start without a technical background?
With a one-hour private AI briefing in plain words, then the free readiness check, then a one-day workshop and a four-week pilot inside the boundary the firm chose.
For the general counsel, the briefing states the model subprocessor and data boundary facts plainly, with the Learn source and the date, so the committee decides from the record.
AFTER TWO WEEKS
What will be working after two weeks?
- The consumer AI tools in use are discovered on managed devices and the block is tested, not assumed.
- The matter workflow's permission row is written and the lawyer approver is named.
- The confidentiality committee has one page in plain English: what the AI can see, who decided, how it is proven.
- You hold the pilot scope and its quote as a written decision.
HONEST ABOUT LIMITS
What this does not do yet
It does not keep every Microsoft AI feature inside the same data boundary. 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) The briefing states which features apply to the firm and the boundary decision is made feature by feature.
It does not make a confidentiality or privilege claim for the firm: the pilot produces the evidence and the general counsel decides what is asserted.
It does not cover personal devices the firm does not manage, which is why the governed alternative matters more than the ban.
THE ONE THING TO REMEMBER
A ban removes your view of the problem; a governed alternative gives it back.
Said another way: private, governed, plain English, in that order.
Said a third way: the committee should be able to read what the AI can see without reading how it works.
SHOW YOUR WORK
See the control path on a fictional bank.
DIRECT ANSWERS
Questions before you decide.
Will associates have to learn a new tool?
No. The workflow runs inside Microsoft 365 where they already work, reads only what they may open, and puts a cited draft in front of the reviewing lawyer.
What happens to client data in the model?
The briefing documents, for each Microsoft feature, what the terms say about training, retention and the data boundary, with the Learn source and the date, so the firm decides from the record.
Can we see a working example?
Hartwell Episode 1 shows a fictional bank blocking a public AI tool on a managed PC, live, and the demos section lists the clips as they are published.
Who reviews before anything leaves the workspace?
A named lawyer, every time. The permission row stops at may draft for anything client-facing, and the review is recorded with the document.
How is it priced?
Quoted after a scoping call. The briefing, the workshop and the pilot are separate scopes and nothing is widened until the firm decides 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.