Demandbase Built a Second AI Model Just to Check the First One's Explanations

Demandbase built a dual-model system called LLMEx specifically to solve a problem most AI sales tools don’t even try to answer: not just what the model predicted, but whether its stated reason for the prediction is actually true. One model generates a plain-English explanation of a prediction; a second “Judge” model independently reconstructs feature-importance rankings from that explanation’s text and checks them against what the underlying model actually weighted — catching explanations that sound reasonable but don’t match the real decision logic.

The engineering approach is notable for what it didn’t require: Demandbase built LLMEx using DSPy for prompt optimization rather than standing up new GPU infrastructure, running the explanation and verification loop as a prompting and evaluation problem instead of a training problem. Tested against roughly 1,900 real Demandbase accounts, explanation accuracy climbed 14 points — from 73% to 87% — as the team scaled training samples from 200 to 1,600, evidence that the faithfulness gap narrows with more labeled examples rather than needing a fundamentally different architecture.

The stakes for getting this right are already visible elsewhere in account-intelligence tooling. Salesforce recently reported Batteries Plus built 10 Agentforce agents to activate a base of roughly 100,000 commercial sales prospects — deployments at that scale depend on sellers trusting what the agent says about why an account is worth a call, which is exactly the faithfulness problem LLMEx is built to catch before it reaches a rep’s queue.

For sales-ops and RevOps teams evaluating AI account-intelligence vendors, LLMEx is a useful marker of where the bar is moving: an explanation attached to a prediction isn’t proof of anything on its own anymore. The question worth asking a vendor now is whether they’ve verified their model’s stated reasoning actually matches its real decision path — or whether the explanation is just a plausible-sounding accessory bolted on after the fact.