ChatGPT and company knowledge: implementing document access

With access to company documents, ChatGPT can draft an answer to an employee's question and identify the supporting material. The benefit comes when the employee reaches the current instruction sooner and can check whether it applies. Connecting a folder does not resolve contradictory versions of a procedure.

Before implementation, define the questions the tool should answer, who will ask them and where reliable sources live. The company's entire document collection is too broad for a first trial.

Start with one team's questions

Choose an area with an accountable document owner. Installation and product-support instructions used by a service team are one example. Collect questions from ordinary work, including some whose answers are absent from the documentation.

For each question, record the reference document and what the answer needs to include. Add a discontinued product, a newer instruction and a document the test user cannot access. This set checks content and access boundaries together.

If a document cannot be found, the answer should explain that gap. A search returning no result does not prove that the company has no applicable policy. The employee needs a named route for clarification.

Choose files or a persistent app connection

Approved document copies can support a small trial. The team can see exactly what it supplied, but someone must update those copies when the source changes.

For a larger collection, assess a supported connection to the source system, such as SharePoint or Google Drive. ChatGPT Work can use files, projects and approved apps through plugins. Capabilities depend on the plan, configuration and connection. See OpenAI documentation.

Test formats individually. Include tables, scans and instructions that depend on a diagram. A correct answer from a simple text file does not establish that the workflow can handle the whole archive.

Establish whose permissions the connection uses

An individual connection uses that person's source-system access. A shared or agent-owned connection can expose its resources to people using the agent, even when their own accounts have different permissions. OpenAI describes these access models in the ChatGPT Work admin FAQ.

Check the source account scope, available actions and audience of the shared tool together. For an instruction-search workflow, enforce read-only access where the connection supports it. An answer or file shared subsequently with another person creates a further access decision.

Test with a restricted user account. Separately test access removal and a change of group membership. Account for the integration's update behaviour and timing. Removing a source permission is insufficient evidence that all previously shared copies have been withdrawn.

Check freshness and supporting sources

Approved documents need an owner, an effective date and a version identifier. Distinguish archived material from current instructions. If two documents disagree, the result should surface the conflict for resolution.

Require a source reference specific enough for an employee to open the relevant passage. During acceptance checks, verify that the passage supports the claim. A plausible-looking link does not establish that the procedure contains the stated rule.

Procedural knowledge may not contain current operational status. How to request a spare part can be answered from an instruction; whether the part is available today requires inventory data. The latter may need a separate integration.

Budget for access, cleanup and ownership

Eight Business Standard users billed monthly cost 8 × $25 = $200 a month, before taxes and additional usage. We checked the rate on 30 September 2026 in OpenAI pricing. Seats do not cover preparing source material or implementation work.

An illustrative internal effort estimate: reviewing 60 documents for 10 minutes each requires 60 × 10 / 60 = 10 hours. At an assumed fully loaded hourly cost of €25, that is €250 of labour. Correcting content, scans and permissions can add further work.

Syntalith audits start at €600 excluding VAT, and training starts at €600 per day excluding VAT. Configuring sources and testing access need a scoped quote. See service pricing. Allow recurring time for document ownership and review of incorrect answers.

Decide whether a custom knowledge system is necessary

Assess a supported connection first when it covers the sources and requirements. A custom solution may be necessary for unusual access rules, a customer-facing interface, specialized retrieval or a required approval workflow. We compare the options in ready-made platforms versus a custom knowledge base.

At the end of the pilot, measure time to a verified answer, the correctness of supporting sources and cases handed back to document owners. Send us a description of your sources and some example questions. That lets us plan a trial before connecting a larger share of company material.