Back to Markets
Stocks● Neutral
How Agentic Systems Depend on Tacit Knowledge
AI agents fail on unwritten company rules. Firms that map tacit decision logic see error rates drop 80%. The investment signal is in the operators who win the knowledge.
Continue with
Companies racing to deploy AI agents are discovering a hard limit. The intelligence that governs day-to-day decisions–the unwritten rules a credit officer applies, the pattern a senior trader recognizes before pulling a bid, the shortcuts a product manager uses to navigate internal approvals–lives outside any process map. It sits in people's heads.
Financial services, software, and consulting firms have found this out through expensive trial and error. They built agents that could parse 500-page compliance manuals and execute documented workflows. The bots still made bad calls because the real rulebook is invisible. A loan officer knows which exceptions the credit committee will accept without being told. A consultant knows that a client's procurement team greenlights anything under $50,000 but flags every expense above that. Those heuristics are never written down.
The gap shows up in agent error rates. One large bank tested a customer-service agent trained only on its formal FAQ. The agent handled 60% of routine queries but generated nearly 25% escalation requests from customers who sensed the answer was wrong. After the bank mapped the tacit escalation rules–who to route a disputed fee to, how to phrase a goodwill credit–the error rate dropped below 5%. That is a typical pattern.
Leaders who understand this spend less time training agents on documents and more time extracting the unwritten logic. They interview senior operators. They shadow decisions. They build taxonomies of exceptions that look like judgment but follow repeatable patterns. The payoff is not just fewer errors. Agents that absorb tacit knowledge can replicate the best performers' judgment at scale, turning a firm's implicit competitive edge into a programmable asset.
The alternative is costly. Companies that skip the tacit layer build agents that fail in edge cases, require constant human overrides, and erode trust. They end up maintaining a human-in-the-loop for every non-standard decision, which defeats the scaling logic of automation. The difference between the two outcomes shows up in operating margins within 18 to 24 months, according to consultants who track implementation results.