Knowing how is not knowing what

After the release of GPT‑6 Astra, developers observed that large language models can reliably navigate familiar interfaces such as Microsoft Word and Google Workspace, essentially turning software operation into a retrieval problem. Hanchung Lee illustrated the gap: an agent may draft an email requesting extra headcount, yet it cannot navigate the budget constraints, stakeholder negotiations, and approval hierarchies that actually secure the position. The article argues that training agents only on surface‑level actions rewards superficial success—sending an email—while ignoring the deeper, undocumented processes that determine whether a task truly succeeds.

This criticism fits into the broader push to move AI assistants from “click‑and‑type” bots toward autonomous business partners. Companies like Microsoft, Google, and Anthropic are investing heavily in tool‑use capabilities, but most current reinforcement‑learning (RL) environments are stripped of the messy, informal knowledge that seasoned employees rely on—approval thresholds, out‑of‑office substitutions, legacy file naming conventions, and informal promises. The piece also notes a nascent market for real corporate data dumps, which, while rich in context, raise privacy concerns and lack the mutable, scenario‑based structure needed for RL training. The author suggests that instead of crowdsourcing thousands of annotations, a handful of domain experts should codify the critical unwritten rules, then feed them into simulation tools.

If the industry follows this roadmap, future agents will need to reason about dynamic organizational states—who knows what, who can sign off, current budget balances, and pending commitments—rather than merely executing pre‑programmed clicks. Risks include the difficulty of accurately modeling human discretion, the potential for agents to manipulate approval processes, and the challenge of sanitizing real company data without stripping essential context. Watch for startups that offer “synthetic org worlds” and for enterprise AI teams that begin to embed policy engines and audit trails into their RL pipelines, as these will signal the next phase of AI‑driven office automation.