NetDocuments Microsoft 365 Add-In Replaces ndOffice: Rollout and AI Governance Risks

The details below come from Legal IT Insider's Neil Cameron, who covered the event. NetDocuments paid his travel and accommodation and invited him to speak on the closing media panel. Product dates, benchmark figures and availability claims are therefore conference-reported or vendor-supplied unless stated otherwise.

The headline for Windows admins: the add-in replaces ndOffice​

Chief product officer Dan Hauck said the reimagined experience will be available to all customers on 12 November. It covers the new web experience, the Microsoft 365 add-in and co-authoring. Cameron reported that the new Microsoft 365 add-in has no desktop component, replaces ndOffice and ndClick on both Windows and Mac, and keeps co-authored documents inside NetDocuments throughout.

That matters because of how the current client works. NetDocuments' support documentation says ndOffice requires the VSTO runtime and .NET 4.8 or later. It lists Windows 11 with Office 2021 or Microsoft 365 as supported. It also says ndOffice is not supported on ARM-based Windows devices, such as certain Surface Pro models. For ARM environments, NetDocuments recommends the M365 Outlook add-in instead. A web-delivered add-in with no local component could remove a lot of that packaging and compatibility work.

The existing Outlook add-in shows what that model looks like in practice. It is a web-based add-in deployed through the Microsoft 365 Admin Center under Integrated Apps. It works across Outlook desktop, Outlook on the web and mobile, and needs no local installation. An MSP's write-up, which I'd treat as practitioner experience rather than vendor documentation, says the Outlook add-in has had feature gaps against classic tools. One example was saving attachments from incoming emails directly into NetDocuments. I have no evidence about how close the new add-in comes to full parity, so pilot it before retiring ndOffice.

Cameron also reported these points about the new experience:

• It already covers more than 90% of everyday user activity, up from about 70% when private preview opened in May. That is a vendor-side measure, and "everyday" is not defined.
• It runs alongside the classic interface on the same data and governance, so firms can migrate users in phases.
• It includes a redesigned home page, an omni-search that sends natural-language questions to Smart Answers, a clients view and a matters tree.
• It includes a matter overview with an AI-generated summary, deduplicated parties, key dates and an activity feed.
• It adds a new mobile app and a dark mode.

A sensible rollout checklist​

This is general guidance from me, not NetDocuments instructions:

1. Inventory where ndOffice and ndClick are installed, including any ARM-based devices.
2. Pilot the add-in with a small group before the 12 November general availability date, and test the workflows your fee-earners use most.
3. Check how co-authoring behaves for documents shared with clients or outside counsel.
4. Decide who owns Integrated Apps deployment in your Microsoft 365 tenant.
5. Keep ndOffice in place until the pilot confirms the add-in does what your users need.

The strategy: the DMS as the center of the wheel​

The conference's single theme was "context." Cameron reported that CEO Josh Baxter's architecture slide put the Legal Context Graph at the center. Harvey, Legora, OpenAI, Claude, Microsoft Copilot and others sat around it, all connected through the Model Context Protocol (MCP).

NetDocuments announced the graph in May. Its press release describes it as a system that maps how matters, documents and communications connect across hundreds of millions of records while respecting each firm's permissions and ethical walls. That is the vendor's own description, not independent validation.

Cameron cites the ILTA 2026 Technology Survey as finding that firms of over 700 lawyers run an average of nine sanctioned AI tools, and firms under 100 lawyers about four. I haven't independently verified those figures. Cameron also reports that NetDocuments' MCP server has been in public preview since 30 September and offers five tools, including save document. Harvey and a ChatGPT Enterprise plugin are reportedly live. Legora said its early-access beta connection was coming, with save-back to follow by the end of October.

Cameron adds a caveat: whether firms will accept the DMS as their point of control remains open, because Harvey and Legora are competing for the same ground. Many firms also run more than one repository after mergers, which no single DMS can fix by declaring itself the center.

Infrastructure: a move to AWS​

CTO John Motz described a two-and-a-half-year re-platforming onto AWS. In the new design, a document arriving triggers extraction, profiling and a matter summary with no human interaction. Germany and Australia are live, with the UK due in November and the Americas in January. Cameron reports a design target of five-nines availability and capacity for ten times current US volume. Those are targets, not measured results.

The benchmark: a cost story more than a quality story​

NetDocuments' Legal Context Engineering Benchmark holds the model and agent fixed and varies only the context. It uses 300 questions across ten matters built from public filings and dockets. The company reports a 48% lower cost per correct answer with the graph enabled. Scott Kelly said that figure has since improved to 55%.

Cameron points out that quality moved far less on the frontier model, from 73.9 to 75.4. He also notes these limits:

• AI models wrote the questions and reference answers.
• An AI grader marked the answers, with human spot checks.
• The corpus is frozen.
• Each matter's score is still an average of its answers.

NetDocuments has published its method, which Cameron calls uncommon among vendors. He says it is working on making the benchmark testable inside individual firms. Treat the results as vendor evidence for this setup, not proof for your own workload.

The risk Cameron raises: "semantic drift"​

Cameron's main analytical point is that a DMS acting as an AI context layer can distort what AI tools see. He names two forms:

• Selection drift. Wrong or missing machine-inferred tags silently exclude documents from retrieval. An answer can be faithful to what was retrieved while important evidence was never considered.
• Interpretive drift. AI-generated summaries and conclusions get filed and then treated as source material, so early errors become harder to tell apart from evidence.

These are Cameron's analytical terms. He does not report that any customer has suffered such an incident.

The save-back feature makes the concern concrete. Documents saved by outside agents inherit the workspace profile and are processed like everything else. Kelly told Cameron that an external agent's authorship is recorded in document history and that partner tools cannot write into the graph itself. Cameron's objection is that a record is not a safeguard. The history can say a document was written by AI, but it can't make the next AI ask or care.

He also cites NetDocuments quoting the MCP specification at the conference: the protocol itself cannot enforce security principles at the protocol level. Ethical walls, DLP and audit trails work inside the DMS. Once content reaches another tool, the receiving tool governs it, if anything does.

Controls Cameron suggests​

These are his recommendations, not a list of shipped features:

• Review the activity log regularly for what AI has added to matters.
• Add a profile field that grades material as AI-generated, AI-assisted, human-verified or external.
• Require approval for every save-back and use AI-access DLP rules for sensitive matters.
• Make verification real. Record how a fact was checked, not just that someone clicked the green tick.
• Sample AI profiling for missing tags as well as wrong ones, and check summaries against source documents quarterly.
• Name one owner with authority to set the switches.

For admins, the useful principle is to treat every MCP connection as a data egress path. Each one needs the same scrutiny as any other integration touching client data.

The regulatory backdrop​

California's SB 574 was signed on 30 September 2026 and takes effect on 1 January 2027. Baker McKenzie's analysis says it requires attorneys to personally verify citations in court filings, regardless of AI use. It also requires reasonable steps to check and correct AI-generated work and disclosure of AI use in court submissions. It restricts entering confidential or personally identifying information into AI systems unless access is limited to the attorney and authorized persons bound by confidentiality. Baker McKenzie notes the law leaves open the form and scope of disclosure and what counts as restricted access. Knowing what AI has touched is becoming a compliance question as well as a housekeeping one.

The bottom line​

• The most concrete item for Windows and Microsoft 365 admins is the 12 November availability of the add-in. It is meant to replace ndOffice and ndClick on Windows and Mac.
• The benchmark's strongest result is on cost, not answer quality, and it comes from the vendor's own test setup.
• Provenance recording is useful, but it only governs downstream AI if receiving tools honor it. That is not yet assured.
• If you run NetDocuments, pilot the new client, audit MCP and save-back settings before turning them on, and assign someone to own AI governance.