Nooks Spent a Year Getting Sales-Agent Accuracy From 50% to 79% on Adversarial Data — Here's What That Actually Took

Nooks’ first production sales agents hit 97% accuracy on standard fact-resolution benchmarks — and under 50% on adversarial cases with conflicting data, the kind that shows up constantly in real CRM records. That gap, disclosed by Nooks co-founder and CTO Nikhil Cheerla in an engineering write-up on what it actually took to move from AI copilot to autonomous sales agent once frontier models became reliable enough in mid-2025, is the starting point for three specific problems the team had to solve. For fact resolution, they built an internal “SalesBench Verified” benchmark using multi-source conflicts and temporal inconsistencies pulled from real customer data, then used calibrated confidence scoring to lift adversarial accuracy to 72-79% — not perfect, but enough that reps know which flagged answers to trust versus verify themselves. For strategy selection, rather than borrowing coding-agent patterns wholesale, they built a filesystem-based system referencing similar accounts plus a “strategy evaluator” tool estimating success probability against comparable historical deals. For message quality, they curated more than 500 examples of bad AI-written sales copy to set quality floors and built per-user contextual memory that generates AI “skills” from a rep’s own email edits — reaching 96% autonomous email sending with no human revision. Reported outcome: meetings per rep doubled within three months, with comparable gains across SMB and enterprise segments.

The piece is also a useful marker of how far the category has moved in just over a year. Nooks’ own earlier framework for the space, published in July, split AI sales agents into autonomous (optimizing for volume) and human-in-the-loop (optimizing for conversion quality, Nooks’ own prior positioning) models — and cited case-study numbers like Greenhouse’s 4x dial increase and HubSpot’s 67% lift in meetings booked per BDR as evidence for the category broadly. This latest post is the same company narrowing that framework into an engineering account of what reliability at the accuracy level required actually costs: not a bigger model, but three separately-built subsystems for fact-checking, strategy, and tone, each with its own benchmark.