The short answer
Run a weekly AI sweep over Salesforce or HubSpot for three gaps: missing close date, stale next-step, and stage that contradicts the last touch. ChatGPT or Claude proposes a fix. The rep confirms in one batch. Clari or Gong supplies the call evidence. Every accepted change is logged with reviewer and rationale, so the QBR opens on data that already won an argument once.
The five-step hygiene workflow
1. Define the hygiene contract. Pick the fields a deal must satisfy: close date inside the current quarter, next-step within 10 days, stage matching the last logged activity, amount within 15% of opportunity size. Reps cannot improve a moving target, and agents need a contract before they touch live records.
2. Run the AI sweep. A Claude or ChatGPT job queries Salesforce or HubSpot every Monday, joins 14 days of Gong transcripts and Outreach activity, and emits one row per offending opportunity: field, current value, proposed fix, evidence link. No writes yet.
3. Rep confirms in a batch. The rep opens one queue (not 40 Salesforce tabs) and votes accept, edit, or reject. Clari pulls the same queue for the forecast call. Same pattern as ICP scoring queues: agent proposes, human confirms.
4. Write back with attribution. Accepted edits flow to the CRM with a structured note: reviewed_by, reviewed_at, evidence_url. Salesforce field history and HubSpot timeline carry it. When the deal closes or slips, you can replay the decision.
5. Carry the queue into the QBR. Export eight weeks of accepted, edited, and rejected items. Patterns surface fast: one rep always rejects close-date pushes, one segment always lacks next-steps. Fix the process, not the records.
Worked example: an AE working a $180K renewal
Marta runs 22 opportunities in Salesforce. Monday's sweep flags 9. The biggest: a $180K renewal in Stage 4, close date six weeks past, Gong call on May 22 where the buyer said "let's revisit in Q3." Claude proposes Stage 3, close August 15, next-step "send Q3 agenda by June 5." Marta accepts with one edit (Stage 2). Clari drops $180K from Q2 commit before the Tuesday call.
The persistent-state pain
Most cleanup decisions live as ephemeral DMs. "Pushed Acme to Q3, buyer wants a pause" goes into Slack; the Salesforce note says "follow up." Six weeks later, nobody remembers which call drove the slip. The CRM holds the field; the reasoning evaporated.
One way to solve this is a workspace like Dock that holds the cleanup queue, the agent's proposed edits, the rep's accept/edit/reject, and the linked Gong evidence as durable rows. Salesforce and HubSpot stay the system of record for account, contact, and stage. Dock holds what the agent interpreted: which field was wrong, why, what the rep decided. Pointers (salesforce_account_id, hubspot_deal_id) keep rows linked. Same structure as the Dock-for-sales surface and AI Salesforce prospecting.
Why it matters
Salesforce's State of Sales research finds reps spend about 70% of their week on non-selling work, with CRM data entry the largest bucket. A workflow that batches cleanup and attributes each edit replaces "what does this field mean?" with a logged decision you can defend.
CTA
Read the full pillar: AI sales prospecting that actually compounds.
FAQ
Q: Should the agent write directly to Salesforce, or wait for the rep? A: Wait on stage, amount, and close date. Those three fields move the forecast. Agent-only writes are safe on enrichment fields and activity logging. See the agent audit pattern.
Q: How does this interact with Clari or Gong? A: Clari reads the same Salesforce records, so accepted edits flow in within minutes. Gong is the evidence layer: the agent links the call timestamp justifying each fix.
Q: What about HubSpot teams without a separate forecasting tool? A: HubSpot's deal pipeline is the forecast. The workflow is identical. The cleanup queue lives outside HubSpot so a stalled write doesn't block the rep.
Q: Won't reps rubber-stamp accept? A: Some will, the first month. When accepted edits turn out wrong, the rep's name is on them. Three cycles in, accept rates settle around 60-70%.
Sources: Salesforce State of Sales statistics, why reps spend 25% of their time on CRM data entry.