Apple Built an AI That Reads Your Database Code and Writes the Missing Documentation

Key points

• Apple ML Research published research on Glyph, an AI system designed to automatically document and classify unlabelled columns in large corporate databases.
• Glyph uses two cooperating AI agents: one writes plain-language descriptions, the other assigns data-sensitivity labels used for privacy and compliance.
• The system reads the actual source code that generates each data column, grounding its descriptions in how the data is really produced.
• Undocumented data columns create legal and regulatory risk, particularly under privacy laws that require firms to know exactly what personal data they hold.

Every large company runs a data lake, a vast internal store of tables and columns holding customer records and transaction logs. Those columns accumulate fast, and the people responsible for documenting them can't keep up. The result is a sprawling catalogue where thousands of columns carry no label: no description, no note about whether the field contains personal data, nothing.

That gap is not just untidy. It breaks data discovery, weakens access controls, and creates genuine regulatory exposure under privacy rules that require companies to know what personal data they hold.

Glyph is Apple's answer. The system, described in research from Apple ML Research, frames the problem as a job for two cooperating LLM agents. An LLM agent is software built on a large language model (the technology behind chatbots like ChatGPT) that can carry out multi-step tasks on its own rather than answering a single question.

What do the two agents actually do?

One agent, the Descriptor, writes plain-language descriptions of what each column contains. The other assigns sensitivity labels from a classification scheme, flagging fields that hold personal or regulated data.

What makes Glyph different from simply asking a chatbot to label columns is where it looks. The Descriptor reaches back to the pipeline source code, the actual programming instructions that create each column, retrieved on demand. Descriptions are grounded in what the data genuinely represents, not a guess from the column name alone.

The two agents are arranged as stateful graphs, an architecture that lets each agent remember what it has already worked out and share that knowledge with the other. Their outputs feed each other: a good description helps the classifier, and a preliminary sensitivity tag can sharpen the description. Our 3 September story on [AI agents writing code](/ story/ai-agents-write-the-code-now-engineers-design-the-walls-that-keep-it-honest) explored a related tension: when generated outputs run automatically, the constraints around them matter as much as the outputs themselves. That concern applies here too.

Should anyone outside a data team care?

Yes. Patients and employees are the people whose data sits in those undocumented columns. When a company cannot say what a field contains, it cannot enforce the right access controls, which means sensitive records may be visible to staff who should not see them. Regulators across the EU and US require companies to maintain records of personal data processing; gaps in documentation are a compliance failure waiting to become a fine.

A system that automates the never-finished work of cataloguing data is genuinely useful. The honest caveat is that this research describes a production system Apple built for its own data infrastructure. Independent peer review of the full results, and performance figures from outside Apple's environment, aren't yet publicly available. What a system does inside one company's carefully controlled pipelines doesn't always travel cleanly elsewhere.

Governance teams at large organisations spend enormous time on exactly this task. Automating even a reliable fraction of it would matter. Watch for independent benchmarks: that's what will tell us whether Glyph's approach holds up beyond Apple's own walls.