AI CRM Automation: Closed-Lost Deal Revival Agent

Marq's sales team was sitting on hundreds of closed-lost deals with no systematic way to know which were worth a second look. I built an AI agent that evaluates old deals at 60, 180, and 360-day intervals, researches each company for fresh revival signals, and automatically enrolls the ones worth pursuing into a multi-step Salesloft cadence, with the assigned rep notified in Slack. It replaced an entirely manual process, it is now live in production, and along the way I cut the cost of the research step by roughly fiftyfold.

Challenge

Re-engagement decisions at Marq were ad hoc. Reps would occasionally remember to revisit a closed-lost deal, but there was no system, no timing, and no consistency, so revenue sat untouched.

The agent also needed strict guardrails. Not every old deal should be resurfaced: some contacts had opted out, some companies already had an active deal, and certain loss reasons should never trigger outreach. I designed a five-layer suppression system, shaped with input from the VP of Sales and VP of Operations, so the agent only surfaces deals genuinely worth pursuing.

Approach

When a deal crosses a closed-lost age threshold, a webhook fires the agent. It runs the suppression checks first, then assembles context: CRM history, past email and call content, and the company's current situation through live deep research, all grounded against Marq's own product knowledge base with retrieval-augmented generation. An LLM weighs the full picture and decides whether the deal is worth reviving.

If it qualifies, the agent creates a tracking deal in HubSpot and enrolls the contact directly into a multi-step Salesloft cadence, then alerts the assigned rep in Slack. An earlier version drafted an email for the rep to send by hand; moving the output to an automated cadence removed that manual step entirely. When a contact is not yet in Salesloft, the agent creates them first, so reach is not limited to people already in the system.

Two engineering decisions mattered most. The first was cost. The original deep-research step was expensive at scale, so I tested and swapped in a leaner research model that returned the same quality of cited, on-point findings for roughly one fiftieth of the price, which is what made running the agent across the full historical backlog affordable. The second was robustness. Putting real volume through the agent surfaced edge cases that clean test deals never hit: deals with no associated company or contact, special characters breaking the data sent to the language model, and stale owner lookups. I hardened each one so the agent skips gracefully instead of crashing, then rolled the backlog out in paced batches rather than all at once, to protect rep workload and the prospect inbox experience.

Results

• Live in production, automatically evaluating closed-lost deals and enrolling the ones worth reviving into a managed Salesloft cadence.
• Removed the manual re-engagement process, so no rep has to remember to revisit old deals or judge which are worth it.
• Five-layer suppression, approved by the VP of Sales and VP of Operations, keeps opted-out, already-active, and disqualified contacts out of outreach.
• Reach extended beyond the existing contact base by auto-creating missing contacts in Salesloft before enrollment.
• Research cost cut by roughly fiftyfold by testing and swapping the research model, which made a full historical backfill viable.
• Hardened against real-world data edge cases and paced into batches so it runs unattended without overloading reps or prospects.

Tools Used

n8n, HubSpot API, Salesloft API, Gemini, Perplexity (Sonar), OpenRouter, Google Drive API (RAG), Slack API