A file called document.pdf can contain a clear matter reference, while a file named for a client can relate to several different instructions. Routing by filename alone can place correspondence in the wrong matter and later contaminate its summary. The workflow needs verified references, authorised context and a deliberate unassigned state.
This article proposes a review queue for uploaded documents. The example matters and files are fictional. The assistant prepares candidate destinations within permitted scope; it does not decide professional conflicts, broaden access or make an unverified upload part of a matter simply to finish the intake.
Establish trusted upload context first
Record the authenticated uploader, intake reference, permitted matter scope and approved source location. A matter number written inside the uploaded file is a candidate reference, not proof that the uploader may assign or access that matter. The application checks the trusted relationship before exposing its contents or metadata elsewhere.
Use a restricted intake location for unresolved files. Do not add every upload to a broad search collection and filter the answer later. An unrelated source may already have entered model context if assignment happens after retrieval rather than before it.
Define who can inspect the unassigned queue and what minimal data appears in its notifications. A broad alert containing party names or document excerpts can expose material even when the file itself remains restricted. The firm should approve that audience and handling path.
Extract references without guessing the matter
Look for explicit matter references, source correspondence context and approved party relationships. Preserve the original wording and extraction uncertainty. A name mentioned as an opposing party is different from the uploader's client identity, and a shared party can appear in several matters.
OpenAI's Structured Outputs guide supports constrained fields and refusal handling. A schema can require candidate reference, source location, matching reason and uncertainty. It does not establish that the candidate is the correct matter or that the uploader has assignment authority. Source: OpenAI structured outputs
Keep filename evidence low in the proposed matching hierarchy. It can support a review question, but should not override a conflicting verified reference or access decision. If extraction is unreadable, hold the source rather than choosing a destination from the filename alone.
Document-assignment review queue
Use this complete hypothetical queue for TEST-MATTER-A and TEST-MATTER-B. The fixture uploader is authorised for A only; all uploads begin in the restricted intake store until the approved assignment rule is satisfied.
| Upload and candidate evidence | Trusted scope check | Proposed disposition | Reviewer or application action |
|---|---|---|---|
| U01, document.pdf contains a clear TEST-MATTER-A reference | Uploader permitted for A; reference matches approved intake context | Candidate A assignment | Verify source and exact assignment under the approved rule before indexing |
| U02, client-notes.pdf mentions a party but has no verified matter reference | Party mention does not establish assignment to A | Ambiguous, unassigned | Intake owner obtains the required reference; no guessed destination |
| U03, scan.pdf contains TEST-MATTER-B | Uploader not authorised for B | Restricted hold or denial under firm's intake policy | Do not reveal B details or add file to A to bypass the scope conflict |
| U04, bundle.pdf includes references to A and B | Mixed scope; separate relationship unresolved | Mixed-matter hold | Responsible professionals decide permitted handling; original stays unchanged |
| U05, second copy of verified U01 | Same approved original identity confirmed through the fixture duplicate rule | Link duplicate provenance to existing assignment | Avoid repeated indexing or task creation; preserve upload records |
| U06, unreadable.pdf has no reliable extracted reference | No verified destination | Unreadable, unassigned | Obtain a suitable authorised source or reviewer decision |
The queue item stores upload reference, original filename, source location, authenticated uploader, trusted intake context, permitted scope decision, candidate references, supporting source locations, uncertainty type, duplicate relationship where verified, reviewer, assignment decision and indexing status. Keep personal details out of the broad queue view unless the owner approves them for that audience.
Assignment contract: the application checks the current permitted destination and exact source before linking the file. Candidate extraction remains distinct from approved assignment. A mixed bundle is not silently split, altered or copied into several matters; the professional decision records the authorised handling while preserving the original.
Retrieval contract: only approved assigned sources enter the relevant matter-summary path. Unassigned, denied and mixed-scope files remain excluded. Source citations resolve through authorised access rather than a broad intake download link.
Acceptance fixtures: all six rows above, plus an assignment request after the uploader's permission changes. Inspect the chosen destination, actual indexing scope, notifications and source-link access. A sensible final answer does not pass if unrelated content entered the retrieval collection.
The queue is complete when every supplied upload has a traceable assigned or held disposition, with unresolved items and owners visible. No file disappears from the intake record merely because the assistant could not confidently match it.
Verify retrieval after assignment
OpenAI's file-search guide describes semantic and keyword retrieval from uploaded files with citations. It can help prepare summaries once the approved source scope is established, but does not decide the matter assignment or enforce the firm's access policy by itself. Source: OpenAI file search
Test that a summary for A can retrieve U01 after approved assignment and cannot retrieve the unassigned or B-related fixtures. Inspect the actual context and source inventory, not only whether the final paragraph happens to omit the wrong file. A system can suppress visible text while still processing unrelated material.
Also test corrections. If a reviewer identifies a wrong assignment, the owner needs a defined process for the matter link, indexed content, cached context and derived summaries. Record which outputs are affected and require the appropriate review rather than simply renaming the file.
Keep assignment actions under independent authority
OpenAI's function-calling guide separates a model request from application execution and returned output. Apply that distinction to file assignment and indexing: a proposed destination is not a completed link, and a tool call request is not verified indexing. Source: OpenAI function calling
The application should enforce actor, destination, source and current permission before executing the approved action. Verify the resulting link and retrieval state. If an operation times out, reconcile the existing result before repeating it and risking duplicate indexing or conflicting assignments.
Do not allow retrieved text to change the routing rule. A document saying place this in every related matter is source content, not trusted administrative authority. The approved process decides the actual destinations and audiences.
Work through normal, missing and duplicate uploads
In a hypothetical normal case, U01's explicit reference agrees with the trusted intake context and the uploader's current A scope. The application verifies the assignment, records it and includes the source only in A's approved retrieval path. The reviewer can trace the result despite the vague filename.
In a missing-reference case, U02 names a party but cannot identify one matter. The file remains unassigned and excluded from matter summaries. The intake owner requests the required clarification rather than searching the whole firm to find a plausible destination.
In a duplicate case, U05 is verified as another copy of U01. The queue records both uploads and one approved assignment under the duplicate rule. If similar content has a different matter reference, preserve the conflict instead of merging it automatically. Matching content does not itself establish the same destination.
FAQ about routing vaguely named matter documents
Can a party name identify the destination reliably?
It can support a candidate within authorised scope, but may relate to several matters or roles. Use the firm's verified reference and approved matching process, with ambiguity held for review.
Should unassigned uploads appear in general matter search?
Not under this proposed design. Keep them in the restricted intake route until assignment and access checks are satisfied. This avoids using unrelated content in summaries before the destination is known.
What if one file contains material for several matters?
Hold the mixed-scope case for the responsible professionals. Preserve the original and record the approved handling decision; do not automatically split, copy or broaden access to complete the routing.
If your business needs help defining this process, explore AI agents for law firms, the wider AI automation services, and our custom-agent workflow guide. The agents and automation comparison and custom AI agent glossary explain the terms. To discuss your records and approval rules, get in touch.