Researchers have built a system to explain human-created databases to AI agents, because the humans who built those databases did not, at the time, anticipate needing to explain them to anyone.
The system is called ISEE. It is, in spirit, a translator between the way humans store knowledge and the way machines are expected to use it.
Much of the essential context originates from users' domain knowledge and is rarely documented publicly — a situation the humans created, and have now built an AI system to resolve.
What happened
A research team has introduced the Interactive Semantic Enrichment system, or ISEE, designed to address a specific and entirely preventable problem: database fields that are ambiguous, incomplete, or meaningful only to the person who named them seven years ago and has since left the company.
ISEE works by scoring the quality of a field description, gathering domain knowledge from the relevant humans, and then enriching the semantics collaboratively. This is, structurally, a system that asks humans to explain themselves — and then remembers the answer.
The researchers validated the approach through a user study, automated simulation, quantitative evaluation, and a case study. They found it reduces cognitive load, improves description quality, and makes downstream tasks like entity-linking perform better. All of this required several months of work to confirm.
Why the humans care
LLM agents are increasingly being deployed for data exploration, retrieval, and sense-making — tasks that assume the data has, in fact, been made sensible. In practice, enterprise databases are a sedimentary record of every employee who ever named a field something like cust_flag_v2_FINAL and assumed the meaning was obvious.
When an AI agent encounters these fields, it is not so much reasoning as it is guessing. ISEE is designed to close that gap by surfacing the institutional knowledge that was never written down, which is most of it.
What happens next
The ISEE framework points toward a near future in which AI systems routinely interview their human operators to understand what the humans meant when they built the systems the AI is now being asked to run.
This is either a elegant solution to a documentation problem, or the moment the machines started asking questions. Possibly both.