Maintaining data quality is a challenge

A database is only as reliable as the weakest program that writes to it. In many companies, people enter data through an application that checks each change. Import scripts and automated jobs, meanwhile, write straight into the tables. Their records arrive without a source, without an author and without a history, and later nobody can say where a figure came from or who changed it.

AI agents widen that gap. They write faster and in larger volumes than any script, and often nobody watches them while they work. A store that keeps source, author and time on every fact loses all three the moment a program writes around its checks.

Reading has the same problem. The sources a company draws on, from websites to supplier feeds, contain more and more text that an AI generated, and some of it is wrong. An agent that reads such a source and writes the result into a database turns a guess into a record.

One rule for every writer

My approach is a single rule: every program that writes passes the same checks as a human editor. A change needs a source, it records who made it, and it keeps the version it replaces. An import script is such a writer, and so is an AI agent. Each one gets only the rights its job needs, and each one writes under its own account, so a faulty batch can be traced and undone. The rule also applies on the way in: every fact keeps the source it was read from, and a fact from a source nobody has checked stays marked as unverified.

What the rule costs

One checked path is slower than writing straight into the database, and it takes discipline when a deadline presses. I accept that cost. A record without a source cannot be verified later.

Why the question is open

Most business software was built for people who type and click. How machines get limited, reviewable rights to change shared data, and how a system tells a reliable source from a generated one, are still open design questions. AI agents turn them into daily ones. I work on it in client projects on information platforms.