Nearly 75% of B2B marketing ops leaders estimate at least 10% of their lead data is inaccurate, outdated, or non-compliant, according to LeadScale. That stat gets worse when you hand an LLM a vague prompt and treat whatever comes back as a qualified list. The spreadsheet looks full. Sales looks at it and calls nobody.
The Prompt Problem Nobody Talks About
Most AI-assisted lead lists start with a prompt like "find marketing leaders at SaaS companies." Broad. Fast. And almost useless, because the output lacks the one thing a sales rep needs before picking up the phone: evidence that the account actually fits.
A defensible lead-list workflow treats every row as a reviewable decision, not a data dump. Each account and contact carries three things: why it fits, what evidence supports that claim, and what's still unknown. The difference between a list sales accepts and one they ignore usually lives in that third column.
Separate Fit From Intent (and Both From Match Rate)
Enrichment vendors love quoting match rates. Those numbers can look impressive: practitioner benchmarks suggest single-source enrichment returns a result about 50-70% of the time, while multi-source waterfalls push that to 80-95%, per Common Room. But a match means the provider found a record. It doesn't mean the data is correct, the person still holds that title, or the company fits your ICP.
A low bounce rate (one agency dataset showed 1.71% across 7.5 million cold emails, per Uplift) can coexist with terrible targeting. Deliverability confirms the address works. It says nothing about whether the buyer wants to hear from you.
Keep company fit, contact fit, and timing/intent as separate scores. "Unknown" should never silently become "pass." When a field is missing, mark it unresolved and decide whether buying the data is worth the potential contact. That decision should be visible to whoever reviews the list, not buried inside a confidence score the LLM assigned.
Citations Make Sales Reviews Faster, Not Slower
The instinct is to clean the list down to a tidy spreadsheet: name, title, company, email, score. Reps don't want to read research papers. True. But they do want to know why a lead is in front of them, especially when the average lead acceptance rate sits around 42%, per Refolk, against a target north of 90%.
A source URL and date-checked next to each material fit claim closes the gap between "marketing says it's qualified" and "I can see why." A rep who can click a link, confirm the company raised a Series B, and verify the prospect's current title will accept or reject in seconds. Without that link, the same rep spends five minutes Googling, gets annoyed, and starts ignoring the list entirely.
This is an alignment fix disguised as a formatting choice. Define the handoff in writing: what constitutes accept, reject, or return-for-review. Attach reason codes. Track them. The feedback loop from sales rejection reasons back into your ICP rules and enrichment sources is where the real compound improvement lives.
Run It This Week
Setup: Write a Claude prompt that specifies your ICP criteria (industry, headcount range, revenue band, geography) and requests a table with: company name, domain, source URL for each fit claim, contact name, current title, and a column for unresolved fields. Instruct the prompt to flag missing evidence rather than filling gaps with inferences.
Before enrichment: Review the company-fit column. Reject accounts that clearly don't match. Hold accounts with missing data. Only authorize paid enrichment on accounts that pass the fit check. This alone can cut your enrichment spend by 30-50% on a typical run.
After enrichment: Deduplicate against your CRM. Check suppression lists (current customers, open deals, unsubscribes). Verify emails. Preserve the verifier's exact status label and check date rather than collapsing everything into pass/fail.
Handoff: Export only reviewed, accepted rows. Include the source URLs, the fit rationale, and any CRM history. Start with a batch of 20-30 and resolve unclear rows before scaling.
Hypothesis: If we add source citations and an explicit fit rationale to every lead record, then sales acceptance rate will increase from our current baseline because reps can verify fit in seconds instead of re-researching each account.
Success metric: Sales acceptance rate on AI-sourced lists. Guardrail: Bounce rate stays below 3%. Stop-loss: If acceptance rate doesn't improve after two batches (40-60 leads reviewed), revisit ICP criteria alignment with sales before running another batch.
Where This Breaks
An audit of a UK B2B SaaS sample found 37% of leads were unserviceable when outreach was attempted. Adding citations and fit checks won't fix that entirely. Web-sourced evidence goes stale. Titles on LinkedIn lag behind org changes by months. The citations make the problem visible, which is the point: a confident-looking spreadsheet with no sources hides rot better than any other format.
List volume drops. Review time per batch goes up. But if your current process generates lists that sales ignores, speed was never the bottleneck. Trust was.