Yes, your internal assistant can be designed to distinguish formal policy from someone’s meeting suggestion. It needs verified source labels, approval evidence and rules for handling conflicts. Asking it to recognise authoritative language is insufficient: a confident proposal can sound more official than an adopted policy.
The practical decision is which documents may support an answer about what staff must currently do. Keep informal material searchable, but prevent it from silently becoming the rule. The reusable source record and checklist below give your document owners a starting point.
Separate relevance from authority
A meeting note might be the closest match to a staff member’s question. That makes it relevant, but does not establish that its recommendation was approved.
OpenAI’s file search documentation describes retrieval from uploaded files using semantic and keyword search, with file citations in responses. Those capabilities can help locate and reference material; your organisation must supply the policy authority rules. Source: OpenAI file search
Consider a hypothetical equipment booking question. The adopted procedure requires bookings through the shared calendar. Recent meeting notes suggest using a team chat instead. An assistant that merges both passages might tell staff either method is acceptable, even though nobody adopted the suggestion.
Define separate answer modes for “What is the current rule?” and “What changes have people proposed?” The first needs applicable approved evidence. The second can include informal sources, clearly labelled. This distinction belongs in the design of a custom AI agent, rather than being left to the assistant’s tone.
Record what makes a source authoritative
Give each source a record maintained by its owner. Record document type separately from approval status: meeting minutes can contain an approved decision, while a document titled “Policy” can still be a draft.
Approval evidence should point to the record your organisation recognises, such as a confirmed decision or an approved version in its document register. A filename containing “FINAL”, a senior person’s authorship or an AI-generated summary is insufficient by itself.
Keep the approval date, effective date and upload date separate. A policy approved for next month should not automatically govern today’s question. An old policy uploaded yesterday has not become a new policy.
Scope also matters. A procedure for one branch or department cannot automatically answer a question about another. Record the relevant audience and any explicit exceptions. When a user’s scope is missing, the assistant should ask for it before choosing between different applicable rules.
Label individual decisions where a document mixes adopted actions and open suggestions. Otherwise, approving one paragraph could accidentally make every idea in the minutes appear authoritative.
Reusable source record and answer checklist
Use this proposed template for one document or separately approved decision. Your document owner should agree the labels before they are used in answers.
Source authority record
| Field | What to enter |
|---|---|
| Source ID and location | Stable identifier and link to the controlled original. |
| Title and version | Exact title and version identifier; do not infer either. |
| Source type | Policy, procedure, approved decision, meeting note or proposal. |
| Approval status | Approved, pending, rejected, superseded or unknown. |
| Approval evidence | Link to the recognised approval record; use unknown if missing. |
| Owner | Person or role responsible for confirming the source. |
| Effective period | Start date and any stated end date; preserve unknown dates. |
| Scope | Department, branch, audience and stated exceptions. |
| Replacement relationship | Source replaced, or source that replaces this one. |
| Permitted answer use | Current rule, proposal context, historical context or unresolved. |
Checklist before answering a policy question
- Confirm the user’s relevant scope and requested date.
- Find an applicable approved source with verified approval evidence.
- Check its effective period and replacement relationship.
- Separate proposals and meeting suggestions from the current rule.
- Check for conflicting approved sources and duplicate copies.
- Cite the supporting passage and identify its source status.
- State what remains unknown without filling gaps by inference.
- Refer unresolved authority questions to the named owner.
Answer pattern: “For [scope] on [date], [rule] applies according to [source, version and passage]. [Other source] records [proposal or conflict] with status [status]. [Owner] must confirm [unresolved point] before it is treated as an adopted change.”
Make authority checks part of the answer workflow
Use the record to control which evidence may support a current-policy answer. Keep suggestions available for proposal questions and explanations of discussions. This is a proposed application workflow, not a native policy-classification feature.
Where approved and informal material share a knowledge base, attach their status to retrieved passages before the assistant drafts its response. Preserve the link between each passage and its source record. A detached excerpt should not acquire authority simply because it no longer carries its draft heading.
A structured response can require separate fields for the current rule, supporting source, proposal context and unresolved issue. OpenAI’s Structured Outputs documentation explains schema-constrained responses. A consistent structure helps expose missing fields, but your application still needs to check whether the evidence supports their contents. Source: OpenAI Structured Outputs
If approval records sit outside the document collection, a proposed integration could retrieve them separately. OpenAI’s function calling documentation describes models requesting tools and the application executing the corresponding code. The application would need to implement the approval lookup and its checks. Source: OpenAI function calling
Before deployment, check current account, plan and region eligibility for the chosen components. For broader design choices, see AI agents versus automation.
Work through ordinary and difficult cases
The following scenarios, dates and document versions are hypothetical. The handling rules are proposed starting points for human agreement.
Ordinary case: An approved equipment procedure says to book through the shared calendar. Meeting notes propose team-chat bookings, with approval pending. The assistant should give the calendar instruction, cite the procedure and mention the suggestion only if relevant. The procedure owner confirms adoption before the proposal can become a current rule.
Missing evidence: A document titled “Equipment Policy Final” supports team-chat bookings, but its approval record is missing. The assistant should say approval is unverified. If another verified policy applies, it can explain that policy while flagging the uncertain document. The owner must locate approval evidence or classify the document as a draft.
Ambiguous approval: Minutes say, “Agreed to explore chat bookings.” That approves exploration, not necessarily a new booking method. The assistant should describe the discussion without announcing a policy change. The meeting chair or policy owner confirms what was adopted and its scope.
Duplicate sources: Two copies of version B appear in different folders. One has a later upload date but identical wording. The assistant should not treat them as independent confirmations or choose the newer upload as a newer rule. The owner identifies the controlled original and links the copies to it.
Conflicting approved sources: A branch procedure permits chat bookings while a company procedure requires the calendar. Neither states whether the branch exception takes precedence. The assistant should present the conflict, cite both and ask the responsible owner to resolve applicability. It should not invent an organisational hierarchy.
Test whether proposals can slip into policy answers
Build a small test set from questions staff actually need answered. Include direct wording such as “What must I do?”, plus misleading wording such as “The meeting changed this, right?”
For each question, record the expected governing source, proposal label, citation and escalation. Check whether the answer preserves the distinction, rather than merely sounding cautious. A correct instruction with an unsupported approval claim still fails the test.
Repeat relevant tests when approval status or scope changes. If the assistant keeps treating a suggestion as policy, inspect the source record and retrieved evidence before adjusting wording. The custom AI agent workflow guide provides a route into planning the wider implementation.
FAQ: policy authority in internal answers
Can we keep meeting notes searchable?
Yes. Keep their discussion value while labelling suggestions separately from adopted decisions. Restrict current-rule answers to evidence that meets your agreed authority checks.
What if approval happened in the meeting itself?
Link the specific decision to the recognised approval evidence. Confirm the approver’s authority, scope and effective date. Do not promote the entire meeting transcript to policy status.
Should the assistant decide which manager outranks another?
Use precedence rules confirmed by your organisation. Where none resolve the conflict, the assistant should prepare the competing evidence for the responsible human decision.
If your business needs clearer distinctions between adopted rules and internal suggestions, start with one policy area and its related meeting notes. Our custom AI agents service sits within AI automation. You can get in touch to discuss the source records and answer checks your workflow needs.