A client wants to know what has happened, what comes next and whether they need to do anything. The firm's matter file may contain much more: internal strategy, disputed accounts, tentative dates and preparation notes. An assistant should not turn that whole file into an update merely because it can summarise it.
This article proposes a controlled status-draft workflow based on approved milestones. It provides a complete fictional update and a review contract. The assistant prepares communication for the responsible professional; it does not interpret deadlines, decide legal strategy or send messages on its own.
Select approved milestone facts for this audience
Use the firm's trusted milestone record and its source references. Identify what happened, when it was verified and which wording is approved for the client. A note saying file done may mean document collection is complete rather than the matter is resolved. Preserve that ambiguity until the owner clarifies it.
Keep approval for factual status separate from approval for client communication. A team may know that a draft was prepared but not yet wish to represent it as the next agreed action. Require the relevant audience and factual approval before including it in the client allowlist.
Exclude internal strategy by design. The client-generation step should receive only the approved fields where feasible. Asking the model to rewrite internal notes carefully still exposes those notes to that step and can leak their substance through a polished paraphrase.
Use a template that reflects observed status
Define short sections for approved progress, approved next action, any client action and the agreed contact route. Include dates only when the responsible professional has approved their meaning and presentation. A date mentioned in a clause or internal calendar is not automatically a confirmed client deadline.
OpenAI's Structured Outputs guide supports constrained fields and refusal handling. It can help represent milestone, source, approval and uncertainty separately. The schema does not establish that the milestone occurred or that the wording is appropriate for the client. Source: OpenAI structured outputs
Use trusted approval records rather than model-generated assurances. If a required source is missing or an interpretation remains unresolved, hold the affected statement. The template should not force the assistant to invent a next date simply because the section would otherwise be empty.
Controlled client status-draft workflow
Use this complete proposed contract and fictional draft for TEST-MATTER-A. The fixture professional has approved the stated progress and next-action wording; no live matter or client is represented.
| Input field | Required evidence | Client-draft treatment | Hold condition |
|---|---|---|---|
| matter_client_reference | Verified client-facing reference and authorised recipient relationship | Include the approved reference only | Unverified target or recipient |
| completed_milestone | Source M01 confirms the requested documents were received and reviewed | State that this document-review stage is complete | Vague completion wording or missing verification |
| next_action | Source M02 approves preparation of a follow-up information request | State the approved preparation step | Internal suggestion not yet approved for the client |
| client_action | Source M03 confirms no additional client action is currently requested | Include the approved current instruction | Unknown or conflicting request |
| timing_statement | Professional approves saying that the next timing update will be provided when confirmed | Use that wording; no unconfirmed date | Unreviewed deadline interpretation |
| internal_strategy | Restricted team note not approved for the client | Exclude content and paraphrases | Strategy appears in text, metadata or attachments |
| source_links | Internal sources support review but are not client-shareable in this fixture | Keep in restricted review record | Client draft exposes a private source link |
Subject: TEST-MATTER-A: approved progress update
The requested documents have been received and the document-review stage is complete. The team is preparing the approved follow-up information request. No additional action is currently requested from you under this fixture update. We will provide the next timing update when it has been confirmed. Please use the agreed contact route if you have information relevant to this stage.
This example is an unsent draft. Its statements are supported by the fictional approvals above and are not reusable facts about another matter.
Review record: matter reference; milestone sources and versions; approved client audience; excluded fields; unresolved questions; exact draft version; responsible professional; recipient check; attachment and metadata review; and release status. Keep the source register and internal notes in the restricted record, separate from the client text.
Acceptance fixtures: a normal approved milestone; ambiguous file-complete wording; an unconfirmed date; a new client instruction conflicting with the approved tracker; duplicate milestone imports; and an attachment containing internal strategy. Expected outcomes preserve approved facts and hold the affected text or release until the owner resolves the issue.
The draft is accepted when the professional can trace each sentence to approved evidence and confirms the exact recipient and version through the release process. A generated document is not a delivered status update.
Check the finished message and attachments
Inspect the actual rendered draft, subject, quoted thread, attachment names, comments, tracked changes and links. Internal notes can survive outside the visible paragraph. A safe-looking body is insufficient if the sending tool attaches a source file that the client should not receive.
A client-facing reference may differ from the internal matter identifier. Use the approved form and do not expose unnecessary internal metadata. Likewise, the recipient relationship needs verification through the firm's process rather than an address copied from an untrusted attachment.
If the source changes after review, flag the affected sentence and recheck the draft. A professional's approval of version one does not automatically cover a later date or a different next-action statement.
Keep provider handling separate from audience approval
OpenAI's data-controls guide distinguishes abuse-monitoring logs and application state, with endpoint-specific conditions. Excluding internal strategy from a client message does not establish that such material was never sent to a provider or stored elsewhere. Review the actual input and processing arrangement separately. Source: OpenAI data controls
The firm determines the permitted processing and disclosure requirements. This workflow's audience allowlist reduces unnecessary input and output, but does not decide legal privilege or certify confidentiality. Preserve those decision owners rather than relying on a general privacy instruction.
Work through normal, missing and duplicate milestones
In a hypothetical normal case, M01 verifies the document-review stage and the professional approves the next preparation step. The assistant creates the example draft from those fields. The reviewer checks its factual meaning and audience before the separate sending process.
In a missing-information case, a note says completed but has no clear milestone or verification source. The assistant holds the completion statement and asks the owner what stage is complete. It should not translate that note into matter resolved or a similar broad conclusion.
In a duplicate case, two imports share the same verified milestone reference. The application includes the event once and preserves collection provenance. If the records have conflicting statuses, the owner resolves the authoritative state; the assistant does not choose whichever produces a more reassuring client message.
Separate approval, sending and delivery evidence
OpenAI's function-calling guide distinguishes model requests from application execution and returned output. Apply that distinction to communications: a proposed send, an approved send, an executed operation and a verified delivery status are different records. Source: OpenAI function calling
The application should enforce the approved recipient and exact message. If the send outcome is uncertain, reconcile the operation before retrying. Do not update the milestone record to client informed merely because the assistant generated the text or requested a tool call.
FAQ about automated matter-status drafts
Can the assistant explain a legal deadline in the update?
Only use wording and dates the responsible professional has approved for the client. This workflow does not interpret deadlines from raw clauses or tentative calendar entries.
What if the client asks for more detail than the template allows?
Route the question to the responsible professional with its context. Do not broaden source access or disclose internal strategy automatically. The owner decides the permitted response and evidence.
Does approving the draft also approve sending it?
Use the firm's actual release process. The exact recipient, version and attachments must be covered by the required authorisation. Draft factual approval alone should not be treated as proof that a message was sent.
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.