Practitioner: models ace verifiable tasks but sales-intel workflows hit ~5% accuracy
JayKurtz90 · x · 2026-09-14
A practitioner working across research, writing, sales and knowledge engineering argues that models keep improving on verifiable tasks but still lack domain-specific reliability in dynamic environments. Examples: "sales intelligence" platforms he tested achieve only 5% accuracy on what actually moves deals, and no product like Astra will magically generate research pre-processed for his own follow-up questions. He finds these custom workflows highly valuable but very hard to build, spent his day rebuilding his first attempt at a global business workflow, and asked for a dedicated lab/course on the topic.
More from coding & agent
- Hamel Husain: Codex can already do anything Muse-style agent tools offer — skip the tool sprawl — HamelHusain · 2026-09-14
- DeskRoot Rebuilds AI Assistant Setup as a Folder of Editable Markdown Procedures — KenGuy14 · 2026-09-14
- arscontexta launches early agent-native IDE for typed knowledge bases — blaizedsouza · 2026-09-14
- Reviewing 2,000 Lines of Agent Code Isn't Management, It's a New Programming Interface — srchvrs · 2026-09-14
- Hugging Face agents reproduced 2,226 ICML papers — about a third of the conference — mmitchell_ai · 2026-09-14
- GPT-6 Astra in practice: a stubborn genius best used as advisor, not coder — kevinkern · 2026-09-14