rulebase-co Internal vs external knowledge
Two libraries that answer the same question differently is worse than one wrong answer. Customers get the external article; agents read the internal wiki; the AI agent grounds on whichever was indexed last — and nobody notices until a fee, timeline or eligibility mismatch surfaces in complaints.
The prize is a clean boundary: customers see outcomes and steps they can take; agents see authority, escalation and exceptions — with deliberate links, not copy-paste duplication.
What leaks when internal is wrong
Internal docs fail quietly because customers never see them — until something bridges the gap:
| Leak path | What goes wrong |
|---|---|
| Agent paraphrases internal SOP in reply | Customer receives procedure meant for staff |
| Macro copied from internal wiki | Wrong tone, internal links, or pre-release info |
| AI agent indexed internal + external | Confident answer from deprecated internal draft |
| Agent cites "what we usually do" | Tribal knowledge overrides published policy |
| Escalation note pasted to customer | Internal codes, competitor mentions, frustration |
| Search spans both libraries | Customer finds internal page via misconfigured portal |
Wrong internal content is an agent efficiency problem. Wrong external content is a customer harm problem. Content that exists in both and diverges is both.
Policy vs procedure — the split that holds
| Layer | Audience | Contains | Example |
|---|---|---|---|
| Policy (external) | Customer | What is true — rights, fees, eligibility, timelines | "Refunds process in 5–10 business days" |
| Procedure (internal) | Agent | How we execute — systems, codes, escalation, exceptions | "Refund queue in admin → type RFD → attach screenshot" |
| Guidance (internal) | Agent | Judgement — when to bend, tone, de-escalation | "If >£500, offer supervisor callback before processing" |
Customers need policy. Agents need procedure. Never publish procedure as policy — it exposes internal tooling, creates security risk, and goes stale on every UI change.
When agents need to explain why a policy exists, external articles get a plain -language rationale — not internal decision history.
What belongs where
External (help centre):
- Questions customers can fully resolve alone
- Stable product behaviour and customer-visible policy
- Prerequisites written in customer language ("You'll need your order number")
- Links to customer-facing forms and status pages
Internal (wiki / agent KB):
- System click paths, admin tools, refund codes
- Escalation matrices and ownership rosters
- Exception handling and compensation bands
- Pre-release, beta, or embargoed information
- Competitive talk, legal strategy, staffing plans
- Draft policy under review
Both — but linked, not duplicated:
- Topic exists externally; internal adds "how we handle edge cases"
- External says what; internal says how with a single link up, not two copies
If you maintain two full copies, they will diverge. Pick one source of truth per fact.
Duplication audit
For each high-volume contact driver:
- List external article(s) on the topic.
- List internal page(s) on the topic.
- Compare factual claims — fees, timelines, steps customers must take.
- Flag contradictions and stale pair (one updated, one not).
| Finding | Action |
|---|---|
| Identical content in both | Keep external; internal becomes link + agent-only addendum |
| Internal newer | Update external or deliberately mark internal as ahead of published policy |
| External newer | Update internal; notify agents |
| Only internal exists | Decide: publish customer version or accept agent-only driver |
| Only external exists | Ensure internal has escalation path, not re-copy of external |
Prioritise drivers by contact volume and regulated risk.
AI and search boundaries
Before grounding any AI on knowledge:
- Index separation — external corpus only for customer-facing bots; internal for agent assist if at all.
- Access control — internal wiki must not be web-crawlable or SSO-leakable into public search.
- Citation rules — agent assist may cite internal; customer bot must cite external URLs customers can open.
- Re-index on publish — policy change updates external first; internal follows same day.
An agent assist tool that retrieves internal docs and drafts customer replies is high-leverage and high-risk. Human send or explicit "customer-safe excerpt" mode is mandatory.
Governance
| Role | Responsibility |
|---|---|
| Policy owner | External factual truth |
| Support ops | Internal procedure currency |
| KB lead | Duplication audits, deprecation, cross-links |
| Localisation | External only unless internal is translated for agents in that market |
When policy changes: external publishes first (or simultaneously with agent comms), internal updated same release, macros and AI corpora refreshed — same change ticket.
Traps
- "Agents need everything in one search." Unified search without labeling internal vs external causes leaks. Label and filter.
- Pasting internal into customer drafts "to save time." Short-term speed; complaint risk.
- Internal wiki as dumping ground. Unowned pages become the de facto policy agents trust.
- Customer help centre with agent-only sections. Sections leak via URL sharing and search; use auth-gated internal instead.
- Assuming SSO equals safe. Agents forward links; screenshots escape.
Present results to the user
- Boundary definition — policy / procedure / guidance with examples from their catalogue.
- Duplication map — topic pairs (external ↔ internal) with contradiction flags.
- Leak-risk paths — macros, AI indexing, search config, macros.
- Remediations — merge to single source, link-only internal, deprecate duplicate.
- AI/search guardrails — what to index where, citation rules.
- Ownership table — who updates which side when policy changes.