One Corpus, Four Audiences, and an AI On Top

Separate what an article is from who may see it.

• Article types describe content shape, not audience. Create a small set (Troubleshooting, How-to/FAQ, Known Issue and Release Note, Policy/Reference). Article types control the template and the fields the six authors must fill, and they determine which data category groups are available. Do not create a type per audience: the same troubleshooting procedure is usually needed by agents and certified partners alike, and duplicating it is exactly how you ended up with three to five near-duplicates.
• Data categories carry topic plus gate. One group for product/module (drives search facets and partner scope), one for audience level (Internal Only, Partner Certified, Customer Public), and one for region or language if localisation is per-article. Channel visibility is then applied to category groups per channel: Internal App sees everything, the partner channel sees only Partner Certified, the customer site sees Customer Public, and the AI agent grounds only on the categories its own channel is permitted to see.
• Certified is a field plus a category, not a naming convention. Use validation status with last-reviewed date and owner to satisfy the contract and legal requirements, and mirror it in the audience category so the platform enforces visibility instead of an author remembering.
• Reject per-article sharing as your gate. Article visibility is category-based; a bespoke shared-with model will drift and will not reliably protect the partner channel.

The test question: can an author publish one troubleshooting article that serves agents and certified partners without a second copy existing?

With six authors and a mostly stale corpus, the only viable strategy is editorial triage plus automation.

• Archive, do not delete. Archiving preserves the record and removes it from agent search and from AI grounding. Rule: unvalidated for 18 months and under a threshold of views in 90 days goes to an archive review queue. Business owners must accept that this queue is a decision list - the six authors should spend their time on the top 200 articles that absorb traffic, not on uniform coverage.
• Push ownership outside Knowledge. Each product area gets a subject-matter reviewer with a quarterly review task, tracked in a report on article version, owner, and last published date. Six people cannot own 2,100 articles; the authoring team should be editors and standards-keepers, not the sole authors.
• Merge before you rewrite. Duplicates poison search ranking and retrieval. Generate a candidate report on title and keyword similarity, and have reviewers merge rather than rewrite - it is faster and it removes the conflict.
• Make using it easier than not using it. Enable suggested articles and article search in the console and surface the top article at case creation. Measure attach rate and handle time for agents who attach versus those who do not. If Slack is still faster, the problem is the search experience, not agent discipline - and no governance model fixes a slower tool.
• Track a small set of health metrics: validation currency, owner coverage, duplicate count, attach rate, self-service success, and repeat contact rate. Do not lead with deflection; it is the easiest number to misread.

An AI agent inherits every defect in the corpus, at machine speed and in the company's voice.

1. Retrieval scope must equal visibility scope. The agent searches with the same category and channel restrictions as the customer site. Test it adversarially: an internal-only article must never surface in a customer-facing answer, and an uncertified article must never reach a partner.
2. Only validated, owned, dated content is groundable. Legal's requirement is effectively an architecture constraint: if every customer-facing answer must carry an owner and last-reviewed date, then only Customer Public plus Validated articles may be retrieved. That also implies the retrieved article must be short and self-contained; long multi-scenario articles retrieve badly and produce hedged, wrong answers.
3. Duplicate removal is a precondition, not a nice-to-have. Two conflicting troubleshooting steps in the index produce a confidently wrong answer, and trust is lost permanently after the first incident.
4. Determinism where the answer is regulated. Retrieve the facts, then assemble the sentence from a controlled template, rather than letting the model compose a commitment about policy or entitlement.
5. A feedback loop with an owner. Every answer needs a thumbs-down path that routes to the article owner; without it, the corpus decays again within six months.
6. Measure on real history before launch: replay several hundred past contacts, have humans score accuracy and completeness, and set the go/no-go threshold on that sample rather than on 'it answered.'

Judgement: if items 1-3 are not done, the correct recommendation is to delay the launch, not to add guardrail prompt text.