Tacit Intelligence

Companies that rely on off-the-shelf foundation models are missing the real competitive edge: the judgment held by their best employees, such as an underwriter spotting a fraudulent claim in thirty seconds. This expertise is not written down in documents; it lives in decisions made under pressure and exceptions caught in the moment. When an expert leaves, that judgment leaves with them unless it has been captured. Without a way to encode this tacit knowledge into a model the organization controls, that value leaks to generic AI systems that anyone can access.

Tacit Intelligence connects to a company's existing knowledge sources (Confluence, Slack, SharePoint, Salesforce, Jira, LMS systems, and more) through a Tacit MCP server, which pulls this data into a pipeline that generates real-world scenarios with expert-validated definitions of a good answer. These scenarios are used to train specialized small language models, evaluated with statistically rigorous methods including Markov chain Monte Carlo sampling for coverage. Models are trained on open-weight base models like Gemma, deployed with hot-swappable LoRA adapters, one-click rollback, and sub-3-second cold starts, inside the customer's own VPC across AWS, Azure, Databricks, or Snowflake. A continuous learning loop captures out-of-distribution scenarios so the model keeps improving, with every training run auditable and every data source and update subject to customer approval.

The site quotes Microsoft CEO Satya Nadella at Davos (Jan 2026) arguing that firms without their tacit knowledge embedded in weights they control are leaking enterprise value and lack sovereignty over their AI.