DealsFlow is a human‑in‑the‑loop agentic system deployed at a UK early‑stage fund. A large language model (LLM) reads startup pitch decks, automatically cites evidence and rates each assessment dimension; backend code then aggregates the ratings, caps them and only ever lowers scores.
Three mechanisms implement this:
- A stage‑aware corroboration ceiling (78, 70 or 60 out of 100) for factual categories lacking external evidence;
- A sceptic call that may lower but never raise a score;
- A pre‑fetch pattern that retrieves and indexes all evidence before a single generation pass.
Evaluation used $N=243$ SEIS/EIS investments, but only the 142 that are at least three years old because newer outcomes are unresolved. The full pipeline’s score correlates with total‑value multiple at Spearman $\rho=0.45$ (95% CI 0.32–0.57) and achieves AUC 0.71 for identifying write‑offs, versus $\rho=0.38$ and AUC 0.68 for an unconstrained LLM. The gain is modest and stems mainly from the sceptic call; we did not test the impact of its lower‑only rule.
Investments that later exceeded $2\times$ had 64% of factual categories corroborated, compared with 38% for write‑offs (omnibus Kruskal‑Wallis test across five outcome groups, $p=0.031$). Retrieval was not date‑restricted, so these figures may overstate screening skill.
In a study of 20 inbound applications, analyst time per application fell from 44.7 to 18.9 minutes including verification and correction, 84% of 240 cited claims were fully supported by sources, yet 70% of reports required at least one correction.
Review