Production patterns for AI in go-to-market. Research, qualification, and outbound systems built to survive contact with a real pipeline.
Most GTM AI fails in one of two directions.
Direction one: volume. A team wires a model to a sequencer and 10x's output. Reply rates fall by more than 10x, the domain gets flagged, and six months later nobody will approve another AI project. The system worked exactly as specified. The specification was wrong.
Direction two: abandonment. A thoughtful research agent produces genuinely good account briefs. Reps don't read them, because the brief arrives in a tab they don't have open at the moment they need it. Usage decays to zero by week five and the pilot is quietly not renewed.
Both failures are architectural, not model failures. This repo is the patterns that avoid them.
| File | What it is |
|---|---|
agents/account-research.md |
Account research agent spec — source hierarchy, output schema, and the conditions under which it must return nothing |
prompts/qualification-rubric.md |
Fit and timing scoring that a rep will actually argue with, which is the point |
evals/outbound-quality.md |
The eval that gates every message. Written before the generator was |
Score the account, don't just describe it. A research agent that returns prose gets skimmed. One that returns a score with the three facts that drove it gets acted on. The prose is a supporting detail, not the deliverable.
The refusal path matters more than the happy path. An account brief built on a thin or ambiguous source set is worse than no brief, because a rep will cite it on a call. Every agent here has explicit conditions for returning insufficient_evidence and a rule that it never fills a gap with inference.
Never let the generator grade itself. Message quality evals run on a separate model with a separate rubric and a held-out sample that the generator has never seen. Self-grading produces a number that goes up while performance goes down.