Undocumented databases were a cost people absorbed. AI agents make it visible

Old databases that nobody explained were a cost companies learned to live with. Jeff Smith of Oracle argues that AI agents change that. On the AI Engineer show, he said agents pay the price of a poorly documented data model every time they work. That makes documenting it worth doing. Smith has built database tools at Oracle for 25 years.

What was said

Smith asked listeners whether a new hire has ever found a diagram on the break-room wall. It would explain the code words, the inside jokes and the data model. His own example was a company's sales that are not in a sales table. They sit in a table with a cryptic code name, he said, because an engineer chose it 30 years ago.

An agent in that situation does not read a wiki page. It queries the database to work out the layout. Smith said: "It will find the data if you ask it. It'll just take a long time." It also searches repeatedly, he added. Without shared memory across everyone's agents, he called this a huge waste.

His fix is to describe the data in plain English, inside the database. Oracle's tools can turn existing Confluence pages, PDFs and Word files into schema notes. An agent then asks for one compact summary of the layout. It no longer runs queries to list every table and column and sample rows from each. Smith said past staff suffered the same way. "Now that AI is a thing, we can't let the poor bots suffer," he said. He added that the notes help humans too.

Why it matters

Our reading: a missing data dictionary used to be hidden in onboarding time and tribal knowledge. Nobody put a price on it. An agent repeats the same discovery work on every task. It runs on a meter, so the waste shows up as time and usage cost.

For managers, that turns documentation from a chore into a spending decision. If several teams point agents at one old system, each can repeat the same searching. A written schema is the cheaper, one-time fix. Smith's description suggests the work is mostly gathering what the company already knows.

The other side

Smith gave no figures for the waste. He did not say how many tokens, hours or dollars an undocumented schema costs. The case rests on his description of how agents behave, so a buyer should measure it first.

He also works for a database vendor that sells these tools. His remarks came in a talk about Oracle's own product.

Notes do not make results trustworthy. Earlier in the talk Smith said natural-language-to-SQL is "amazing, but I think it's dangerous." He warned that queries can look right and return the wrong data. Annotations help an agent find its way. They do not guarantee that the answers are correct.