Show Your Working

James Pursey, founder of Replicate Labs, and Greg McCallum of PeerLab ask whether AI can score a sales call as well as a trained expert, and give the same answer twice. James ran more than 5,000 scoring tests across 15 AI models to find out. Greg's team at PeerLab supplied the expert reviews the AI was checked against.

Day one looked like a win. Then James ran it again and the scores kept moving. He and Greg get into why newer models mark reps harder and why nearly every fix made things worse. The change that helped accuracy most was giving the model each expert's written reason for a score. James then lays out the setup that held up: Jev and Luna score each call separately and the two scores are averaged. It matched the expert ranking at 0.89, only 1.7% of scores change on a re-run, and it runs up to 28x cheaper than sending each call to the latest model (list prices checked 21 Oct 2026). He also walks through the templates, and they cover building this yourself versus buying it.

Replicate Labs: replicatelabs.ai

PeerLab: peerlab.ai

Free ebook with the full working: replicatelabs.ai/show-your-working

Next episode: Game Winning Kickoffs with Celeste from ASG, Wednesday 11 November.