Every serious conversation about AI at a law firm arrives, usually within the first hour, at the document management system. That is where the documents are. iManage Work or NetDocuments holds the agreements, the briefs, the opinions, the closing sets, and the filed email going back decades. If AI is going to be useful to the firm, it has to read what the DMS holds.
But the DMS is not just a file store. It is where the firm's security model lives: matter-centric workspaces, need-to-know access on sensitive engagements, ethical walls raised under protective orders and lateral screening obligations. The filing structure and the permission structure are the accumulated output of decades of risk decisions. An AI layer that reads the documents while approximating the security model has not integrated with the DMS. It has bypassed it.
Treating the DMS as the system of record is therefore an architectural commitment, not a slogan. It means the AI layer holds no permissions of its own invention, publishes nothing the DMS would not permit, and defers to the DMS on every question of content, version, and access. Everything that follows falls out of that commitment.
Ingestion: take the metadata, not just the text
The mechanical half of integration is ingestion, and the standard is API-based, incremental, and metadata-preserving. API-based, because file-share exports and database snapshots sever the link to the source of truth. Incremental, because a firm's DMS changes by the minute and a nightly full crawl is neither fresh nor affordable at archive scale. Metadata-preserving, because the profile fields that attorneys and records teams have maintained for years (client, matter, document type, author, dates) are the highest-quality structured data the firm owns.
Those profile fields are gold for a knowledge layer. A model can guess that a document is a credit agreement; the DMS profile states it, along with the matter it belongs to and the client that matter serves. When ingestion carries the metadata into the graph, every downstream answer inherits it. Reframe builds its Context Graph this way: profile data forms the skeleton, extraction fills in the clause-level detail, and the result is described in the platform overview.
Scope selection matters too. Most firms do not ingest the entire archive on day one. They start with the workspaces of one or two practice groups, honor the same inclusion rules the records team already applies (client instructions, holds, destruction schedules), and widen as trust builds. Deletions and refilings must propagate as faithfully as additions: when a document moves matters, is reprofiled, or is destroyed under a records schedule, the knowledge layer has to follow.
Permissions: the wall that goes up at 3 p.m.
Here is the scenario that should anchor every integration review. A lateral arrives, a conflict is identified, and at 3 p.m. the general counsel's office raises an ethical wall around a live matter. From that moment, a screened attorney must not be able to reach the walled documents through any door: not the DMS client, not enterprise search, and not an AI assistant that ingested those documents last week.
Three mechanisms, layered, make that true. ACL mirroring keeps document and workspace permissions synchronized alongside the indexed content, so retrieval respects them wholesale. Event-driven updates close the timing gap: a batch sync that runs nightly leaves the wall porous until morning, while event-driven propagation closes it in minutes. And query-time enforcement is the backstop: at the moment of retrieval, the system checks current entitlements before a single passage reaches the model's context window or the user's screen.
The integration's own service account deserves the same scrutiny. A connector that crawls the DMS with broad administrative rights becomes the most privileged reader in the firm, so its credentials, scope, and audit logging belong in the security review alongside the user-facing controls. Least privilege applies to software exactly as it applies to staff.
AI adds a subtle failure mode of its own. A model that read a walled document yesterday can paraphrase it today unless enforcement gates what enters the prompt, not merely what appears in a results list. So ask two questions of any vendor: how long is the window between a permission change in the DMS and enforcement in the AI layer, and does enforcement apply to retrieval into the model rather than just to display? The enforcement model deserves a permanent place in your security review; ours is documented on the security page.
Versions, scans, and the email problem
Three unglamorous details decide whether attorneys trust what comes back.
- Versions. The executed agreement and draft seven live in the same workspace, and they disagree. An AI layer that is not version-aware will eventually quote a position the parties negotiated away, with a confident citation to the wrong paper. Version awareness (executed over draft, latest over superseded, lineage preserved) is a correctness requirement, not a refinement.
- Legacy scans. A meaningful share of any firm's archive is image-only PDF: signature pages, exhibits, closing sets from the years before born-digital practice. Without OCR in the ingestion path, the archive silently thins, and answers skew toward the recent without anyone noticing the skew.
- Filed email. A large fraction of the firm's actual reasoning lives in email threads filed to matters: why the clause moved, what the client instructed, what opposing counsel conceded at 11 p.m. Integration that skips filed email reads the conclusions and misses the reasoning, and on some matters email is the only record of the negotiation.
Writing back, carefully
Reading is half the integration. The DMS gets better when the AI layer writes back under controls. Profiling suggestions: the system proposes document type, matter association, and parties for unprofiled or misprofiled documents, and a records professional approves before anything changes. Auto-filing recommendations for email shrink the largest gap in most archives. Summaries and extracted term sheets can land as new documents in the workspace, or as metadata on the source, so the work product is findable by everyone with rights to the matter rather than trapped in one attorney's chat history.
The rule for write-back mirrors the rule for reading: the DMS remains authoritative. Suggestions flow in through the API, land as pending updates governed by DMS permissions, and take effect only when a person with authority accepts them.
Done well, write-back also pays the records team in a currency it values: higher email filing rates, cleaner profiles, fewer orphaned documents. The AI layer stops being one more consumer of the archive and becomes a contributor to its upkeep.
The anti-pattern: exporting the DMS
There is a shortcut that appears in vendor architectures and should end the evaluation when it does: exporting the document corpus to the vendor's own cloud store and running everything there. The copy starts stale and gets worse. Permissions fork from the source at the moment of export, so the 3 p.m. wall never reaches the copy. Versions drift as new drafts land in the DMS and not in the mirror. And the firm now operates a second, unmanaged archive of its most sensitive material, outside the systems its records policies and client commitments actually govern. The export also creates a copy that outlives the contract, which becomes a data destruction problem the firm inherits at renewal time.
The DMS is the system of record. An AI layer that cannot work subordinate to it will eventually work against it.
The right pattern keeps the DMS authoritative for content and permissions alike, keeps the AI layer subordinate to it, and keeps the whole arrangement inside the firm's boundary. That is how Reframe connects: to iManage or NetDocuments through their APIs, from inside the firm's own tenant, with the graph and the models governed by the firm's existing controls.
Done right, this integration becomes the foundation for everything else the firm wants from AI: knowledge management that runs on extraction rather than filing discipline, and the knowledge graph that turns a document archive into something the firm can actually query. Done wrong, no amount of model quality buys the trust back.