Grounding your internal knowledge pilot

Most internal “AI assistants” fail quietly: staff stop using them because answers feel generic, cite nothing useful, or surface information people were never meant to see together. A grounded pilot starts with sources people already trust, access that mirrors real roles, and behaviour that shows its work — not with a model pick.

Pick a narrow slice of truth

Resist indexing “everything we have ever written.” Choose one operational domain where repeated questions burn time: onboarding packs, standard operating procedures, product FAQs, or support playbooks. Within that slice, list the systems of record (SharePoint, Confluence, Notion, a wiki, ticket macros) and agree which folder or space is authoritative when two versions disagree.

Access and PII before retrieval

Mirror group membership from your identity provider; do not invent parallel permissions inside the pilot. Flag fields that must never leave their source system (payroll notes, medical or legal free text). If you cannot explain who can see an answer and why, you are not ready to widen the pilot.

Citations, freshness, and ownership

Citations — Every non-trivial answer should point to a specific page, section, or ticket template so users can verify context.
Freshness — Name an owner per corpus who receives a weekly digest of queries that returned stale or missing sources.
Escalation — When confidence is low, the product should say so and route to a human path instead of guessing.

What to decide in the first month

Success metric: time saved, deflection of repeat tickets, or faster onboarding — pick one primary measure.
Allowed tone and length for answers (especially for customer-facing snippets).
Whether draft or archived content is excluded by default.
How feedback from frontline staff feeds back into the knowledge base.

Trust comes from provenance and boundaries — not from a larger context window.

How we help

We design knowledge pilots around your real workflows and data boundaries. For solution shapes and budgets, see Knowledge base & internal Q&A and Work; for process questions, FAQ.