Point retrieval at the revenue corpus (L4)

The steps

1. 01

   Count the re-asks in the channels where they already happen

   Tool: Slack

   Search 90 days and tally repeated questions. Expect four clusters: pricing and discount exceptions, competitor comparisons, security and compliance answers, and integration or technical limits. Rank by frequency and take the top 20. That ranking is your corpus scope, and it will be far smaller than "all of our content," which is the point. • Owner: Enablement • Tool options: Slack search or Microsoft Teams search across the sales channels • Pitfall: trying to build a corpus of "all of our content" instead of the top 20 real re-asks • Definition of done: a top-20 question list exists with a frequency count on each

2. 02

   Define the corpus tightly and name an owner per source

   Tool: Google Sheets

   List every source to be indexed, with three fields on each: named owner, review cadence, and last-reviewed date. Include only current approved decks, live pricing, current battlecards, security documentation, and won-deal call summaries. Explicitly exclude drafts, personal drives, anything without an owner, and any folder called "archive" or "old." Unowned content is how the system starts lying, and it lies in the confident voice of an approved source. • Owner: Enablement plus RevOps • Tool options: a source registry in Sheets, Notion or Confluence • Pitfall: indexing drafts, personal drives or unowned folders like "archive" • Definition of done: every source has a named owner and a review date, and unowned sources are excluded in writing

3. 03

   Index with metadata that makes staleness enforceable

   Tool: Pinecone

   Every chunk carries four metadata fields: source, last-reviewed date, owner, and intended audience. Then configure retrieval to filter out anything past its review date rather than surfacing it with a warning label. The gate is a filter, not a caveat. A stale answer with a disclaimer is still the answer a rep repeats to a customer. Test it deliberately: take a document, backdate its review date past the threshold, and confirm it disappears from results. • Owner: GTM engineer, or your vendor's admin if buying • Tool options: Pinecone, Weaviate, Qdrant or pgvector with LangChain or LlamaIndex; or Glean, Guru or your enablement platform • Pitfall: surfacing stale content with a warning label instead of filtering it out entirely • Definition of done: a deliberately stale document is proven excluded from live results

4. 04

   Require a citation, an owner and a date in every answer

   Every response returns four things: the answer, the source document link, the last-reviewed date, and the owner's name. Then define the failure behavior explicitly: when retrieval finds no current source, the answer is "no current source for this, ask [named owner]." A refusal with a routing instruction beats a fluent guess every single time, and reps forgive refusals while they never forgive one confident wrong answer about pricing. • Owner: GTM engineer • Tool options: the prompt template or the vendor's answer configuration • Pitfall: letting the assistant produce a fluent guess instead of a routed refusal when no source exists • Definition of done: 20 test questions return full citations, and questions with no current source return the routed refusal rather than an answer

5. 05

   Deploy where the questions already get asked

   Tool: Slack

   Do not build a portal. A portal is a place people have to remember to go, and they will not go. Put the assistant in the channel where step 1 found the questions, so the behavior change is zero. Announce it with three of the top-20 questions answered live in the channel, so people see the format and the citations. Then say plainly what is out of scope. • Owner: Enablement • Tool options: the Slack or Teams app for your chosen platform, in the existing sales channel • Pitfall: building a standalone portal instead of deploying inside the existing channel • Definition of done: 50% of the top-20 questions are being answered by the assistant in the existing channel within 30 days

6. 06

   Run a weekly gap log and let it drive the content roadmap

   Every week take the top three failures and do one of two things: fix the source, or write the missing document. Log which you did. The gap log becomes the enablement content roadmap, which makes the roadmap demand-driven rather than guessed at the start of the quarter. That is quietly the biggest return in this playbook, you stop producing content nobody asked for. • Owner: Enablement • Tool options: a weekly log of unanswered queries and thumbs-down responses • Pitfall: letting the content roadmap stay guessed at the start of the quarter instead of demand-driven • Definition of done: four consecutive weeks of gap-driven content updates have shipped

7. 07

   Instrument the result

   Tool: Looker

   Instrument: percentage of repeat questions answered without a human, plus median time-to-answer versus the human baseline. Capture the baseline in step 1 while you are already counting. • Where it breaks: the corpus includes everything anybody ever wrote. Precision collapses, a rep gets one bad answer on pricing, and they never come back, trust is lost per-incident and regained per-quarter. Second failure: no freshness gate, so last year's pricing gets quoted with full confidence. Third: it ships as a portal, adoption sits near zero, and everyone concludes reps do not want it. • Visual guidance: a retrieval flow with the freshness gate drawn as an actual gate between index and answer, the rejected-stale path in gray, and a citation card at the end showing source, date and owner.

Tools in this playbook

• Slack
• Google Sheets
• Pinecone
• Looker

Next playbooks

• Notion-Native Account Plan Agent, "Yes Chef" (L3)
• AI call highlight reels for coaching (L1)
• Onboarding ramp assistant built on your own content (L2)
• Certification with AI roleplay scoring (L4)