Entrazure
Three constraints that make this sector different
Confidentiality is per matter, not per firm. A partner may be barred from seeing a file held two floors away, and the wall is a professional obligation rather than an internal preference. Almost every Microsoft 365 tenant is built on the opposite assumption: broad SharePoint sites, generous default sharing, and a decade of links sent to whoever asked. That gap is tolerable while search is mediocre. It stops being tolerable the moment an assistant can retrieve across the tenant on demand.
Billable time is the product, so friction is fatal. Every hour spent inside internal systems is an hour not billed, and adoption is decided in the first week by people with no appetite for a change program. The measure is not user satisfaction, it is minutes per matter against the way the work was done before. A tool that needs a training session before it saves anyone anything has already lost that comparison.
Your clients audit your AI before you finish evaluating it. Outside counsel guidelines and engagement terms increasingly specify where data is processed, whether content is used for training, how long prompts are retained and who at the firm may see the file. The firm’s risk committee needs a written answer, and that answer has to match the actual tenant configuration, because the client’s security team may ask for evidence rather than assurance.
What we build here
- Oversharing remediation before anything is enabled. A real audit of SharePoint and OneDrive sharing, broken inheritance and anonymous links, then sensitivity labels and DLP in Microsoft Purview, then Restricted SharePoint Search during pilot. Turning on Microsoft 365 Copilot in an unaudited tenant is the fastest way to create an internal incident, and in a firm with conflict walls it is the fastest way to create a reportable one.
- Identity and device posture that fits a firm. Microsoft Entra ID with conditional access, Privileged Identity Management, access reviews, and governed guest access for the client and co-counsel collaboration that happens whether or not IT sanctions it. Intune for devices that leave the building daily. This is cybersecurity and Entra ID applied to a partnership rather than an enterprise.
- An assistant over your own precedent and knowledge. Azure AI Search with security trimming so retrieval respects the permissions on the source document, inference inside your tenant, full prompt logging, citations back to the file. We choose between Copilot Studio and Azure AI Foundry on the merits of the case, and our AI engineering work always begins with what the system is forbidden to do.
- Document work that removes keystrokes. Document Intelligence over engagement letters, invoices, expense documentation and inbound client forms, feeding the practice management or time system through queued and retried integration rather than a nightly export nobody monitors.
- Governed reporting on realisation and utilisation. One certified semantic model in Microsoft Fabric instead of the four spreadsheets the management committee currently reconciles by hand before every meeting.
Where a firm needs its own Azure footprint for data outside Microsoft 365, that starts with an Azure landing zone sized to the firm rather than to an enterprise template.
An anonymized scenario
Anonymized scenario, not a named client.
A professional firm of about 180 people, four offices, wants a knowledge assistant over fifteen years of accumulated precedent and engagement files. The partners are enthusiastic and the risk committee is not, for the correct reason: nobody can currently state who has access to what.
The first six weeks produce no assistant. They produce a permissions picture: how many sites have broken inheritance, how many documents are shared with anonymous links, which walled matters are reachable by search today. Remediation follows, with sensitivity labels applied to the categories that matter and an owner named for each site. Only then does retrieval get built, security-trimmed against those permissions, with citations shown and every prompt logged. The pilot is one practice group, measured against a baseline of minutes spent locating precedent. The uncomfortable finding is the valuable one: the access problem existed before the AI project and would have surfaced eventually through a client audit instead.
Regulatory context we design against
Professional conduct obligations come first: law society rules on confidentiality and conflicts in each jurisdiction you practice in, CPA confidentiality requirements, and professional order obligations for engineering practice in Quebec, including record retention and sealing. Quebec Law 25 applies to client and employee personal information, with the privacy impact assessment expected before information moves outside the province, covered on our Law 25 compliance page. Clients in regulated industries will pass their own obligations to you through engagement terms, which is why SOC 2 Type II increasingly appears in requests for proposal aimed at firms that have never thought of themselves as technology vendors. We build the estate so each obligation lands as an enforced control with an exportable record, which is the subject of compliance architecture.
Questions we get
We want Microsoft 365 Copilot. What has to happen first?
A permission audit. Copilot surfaces everything a user already had access to, which in most tenants is far more than anyone believed, and in a firm with conflict walls that is an ethical problem rather than an inconvenience. We audit SharePoint and OneDrive sharing, remediate the worst of it, apply sensitivity labels, then pilot with Restricted SharePoint Search before opening it up.
Can an AI assistant respect our conflict walls?
Only if retrieval is security-trimmed against the source permissions, which most retrieval implementations skip. We build so the index enforces the same access the document library enforces, and so a walled matter is invisible rather than merely unlinked. If the wall depends on the model choosing not to answer, it is not a wall.
Our clients now ask where their data is processed. What do we tell them?
Something you can evidence. Inference runs inside your tenant in a region you chose, no client content trains a shared model, prompts and responses are logged and retained on a stated schedule, and access is limited to the engagement team. We write it as a one-page description your risk committee can send without editing.
How do we know whether an AI tool is worth keeping?
Measure against a baseline recorded before deployment: minutes per document reviewed, time to first draft, hours written off per engagement. In a firm where billable time is the product, a tool that saves six minutes per matter is a real return and a tool that adds two is uninstalled by Friday whatever the demo showed.
We are twenty people. Do we need Azure at all, or just Microsoft 365?
Often just Microsoft 365, configured properly, plus identity and device compliance. We will say so. Azure enters when you have data outside Microsoft 365 to govern, an integration between practice systems to build, or an AI workload that needs its own boundary. Selling you a landing zone you do not need is how a relationship ends at renewal.