Start with a curated set of approved documents for a staff-policy AI assistant. Add sources when unanswered questions expose useful gaps. Give it a whole folder only when that folder already operates as an approved, maintained library. The decision should balance authoritative coverage, stale-document risk and the work needed to keep answers useful.
For a business considering custom AI agents, the practical question is which documents should become evidence for staff answers. A large upload can include more topics while also admitting abandoned drafts, duplicate copies and guidance intended for a different audience.
Compare the evidence behind answers, not the number of files
Whole-folder ingestion means including every eligible document within a defined folder boundary. Curation means deliberately selecting documents against agreed inclusion rules. Neither approach tells you whether a particular answer is supported.
OpenAI documents file search as retrieval over uploaded files using semantic and keyword search. That establishes a way to find relevant material; it does not establish which of your documents represents approved company policy. Source: OpenAI file search
Keep three questions separate: does the library contain an answer, did retrieval find the relevant passage, and does the response accurately represent it? Adding documents addresses the first question only when the added material contains authoritative guidance. A retrieval failure needs a different remedy from a missing policy.
Use one staff-policy library to expose the trade-offs
Consider a hypothetical South African business with a shared staff-policy folder. Every quantity in this example is hypothetical: it contains 36 files, of which 12 are confirmed current policies. The rest include forms, old versions, workshop notes and draft amendments.
A curated starting set includes the current policies and supporting forms whose owners have confirmed their purpose. Staff questions cover leave requests, expense submissions, remote working and finding the correct form. Individual entitlement disputes remain with the appropriate human decision-maker.
The whole folder might contain a useful instruction buried in a workshop note. However, someone still needs to establish whether that instruction was adopted. The curated set might omit a genuinely approved procedure because nobody identified it during selection. That omission should become a maintenance task.
The comparison therefore needs evidence about both coverage and authority. Record questions the curated set cannot answer, then inspect excluded documents for approved answers. Do not count every additional answer from the larger folder as an improvement.
Apply this knowledge-ingestion decision matrix
Use the matrix in a meeting with the policy owner and the person maintaining the assistant. These are proposed selection rules, rather than product defaults or legal requirements.
| Decision factor | Choose a curated set when… | Whole-folder ingestion is reasonable when… | Evidence to record |
|---|---|---|---|
| Approval | Drafts and approved documents are mixed. | Every included document has confirmed approval for this use. | Approver, approval status and document owner. |
| Coverage | A defined set of staff questions drives the assistant. | The maintained folder covers the agreed scope without unrelated material. | Questions mapped to supporting documents; remaining gaps. |
| Current versions | Old copies remain beside replacements. | Superseded copies are reliably removed from the retrieval set. | Effective date, version and replacement relationship. |
| Audience | Documents serve different staff groups. | All included material suits the intended audience and access boundary. | Audience and permitted access, checked separately. |
| Duplicates | Similar copies have uncertain authority. | One authoritative source is identified for each policy. | Canonical source and duplicate handling. |
| Maintenance | An owner can review a manageable selection. | Folder admission and removal already follow a dependable process. | Named owner, update trigger and verification steps. |
Decision record: Chosen approach: ____. Library scope: ____. Policy owner: ____. Maintenance owner: ____. Approved sources: ____. Unanswered questions: ____. Excluded or superseded sources: ____. Next review trigger: ____.
Completion rule: Confirm supporting sources for ordinary questions, record expected handling for missing or conflicting guidance, and verify a policy replacement before expanding access.
A mixed result usually supports a curated starting set with staged expansion. If approval, versions and duplicates are unresolved, broader ingestion introduces unresolved evidence into staff answers. If the folder already meets the right-hand conditions, maintaining a separate selection may create unnecessary duplication of effort.
Test ordinary, missing, ambiguous and duplicate cases
Run the same question set against both candidate libraries. Keep the wording and intended staff audience consistent. The following scenarios are hypothetical, and the expected handling is a proposed workflow.
Ordinary question: “Where do I submit an expense claim?” The current expense policy names a submission channel and links to an approved form. The assistant should identify both, cite the supporting passage and avoid adding an approval promise. A policy owner checks that the answer matches the source. Both libraries may perform equally here.
Missing guidance: “Can I claim equipment bought before joining?” Neither approved library contains this rule. An excluded draft suggests an answer, but it has no confirmed approval. Expected handling is to state that the approved material does not establish the answer and refer the question to the policy owner. The owner decides whether guidance should be added; the draft is not promoted simply because it fills a gap.
Ambiguous scope: “How many remote-working days can I take?” Two approved documents apply to different staff groups. The assistant should request the relevant context or explain that the applicable policy remains unresolved. The human reviewer confirms the audience distinction and checks whether source labels make it clear. Do not infer entitlement from whichever passage appears first.
Duplicate documents: The folder contains a current policy and a copy with the same title but an older submission channel. Expected handling is to use the confirmed authoritative version. If that authority is unclear, flag the conflict for the owner. The maintenance owner then removes or excludes the superseded copy and repeats the question.
For each result, record the answer, supporting source, applicability and required human follow-up. A fluent response with a citation still needs checking against the cited passage.
Let unanswered questions drive controlled expansion
Keep a short gap register containing the question, intended audience, missing topic, owner and resolution. Classify each gap before adding files: missing approved guidance, an approved source omitted from ingestion, retrieval failure, or ambiguous policy wording.
For an omitted source, confirm approval and add it. For a retrieval failure, investigate why the existing source was not found. For unclear wording, ask the owner to clarify the policy. Where no approved guidance exists, record the gap and leave the substantive decision with the responsible person.
A proposed assistant could return consistent fields such as answer status, supporting documents and unresolved question. Structured Outputs provides schema-constrained responses, but a consistent format is not proof that the policy interpretation is correct. Source checking remains a separate step. Source: OpenAI Structured Outputs
This expansion method makes the coverage trade-off visible. It also gives the owner a reason for each proposed addition, instead of treating every unused document as something the assistant should absorb.
Budget for maintenance after the first upload
Compare the actual tasks each option creates. Curation requires selecting sources and keeping that selection aligned with policy changes. Whole-folder ingestion requires dependable folder admission, removal and version control. A folder full of unmanaged documents shifts work into resolving questionable answers.
Give policy approval and ingestion maintenance named owners, even if one person holds both roles. On replacement, confirm the new document's approval, update the retrieval set, remove the superseded source and repeat affected questions. Include one unrelated question to check that a narrow change has not disturbed other answers.
Any proposed connection that checks a document register or records a gap needs application logic. OpenAI's function-calling guide describes the model requesting a tool call and the application executing it; it does not supply your policy approval process. Source: OpenAI function calling
Before deployment, check current account, plan and region eligibility for the selected products. Keep this knowledge-selection decision within the wider AI automation design. The custom AI agent workflow guide and AI agents versus automation comparison help explain the surrounding implementation choices.
FAQ: choosing documents for a staff-policy assistant
Should we include old policies for historical questions?
Only if historical questions form part of the agreed scope. Identify historical material clearly and require the relevant date before answering. Otherwise, exclude it from the current-policy retrieval set and retain it in the organisation's appropriate archive.
Can we start curated and move to a whole folder later?
Yes, as a proposed rollout approach. Use the gap register to expand coverage, then reassess the matrix. Move when the folder itself meets the approval and maintenance conditions, rather than when the assistant reaches a particular file count.
Who should decide whether an unanswered question needs another document?
The policy owner should confirm whether approved guidance exists and applies. The maintenance owner handles inclusion and verification. The custom AI agent glossary provides a plain-language reference for the assistant's role.
If your business needs help choosing between a managed folder and a curated policy library, get in touch with your question set and document inventory. Those provide a concrete starting point for deciding what the assistant should use.