You Can Trust Your Digital Workers With Structured Data: The Structured Data Module

TL;DR

A hallucinated word is a typo. A hallucinated number is a decision. Nobody misses a wrong adjective in a summary. Everybody misses one flipped cell in a 40,000 row extract.

LLMs are probabilistic engines. Your general ledger is not. Next token prediction gives you no guarantee that an operation was applied to every row, and no way to prove where a value came from.

The fix is architectural, not a better prompt. Keep the data out of the context window and let the model reason over schemas while auditable tools do the work.

If you cannot show the working, you cannot sign off on the number. Full provenance per value, from source system through every transformation, reconstructed from what actually ran.

Our Argument:

Ask an enterprise team where agentic AI has burned them and they rarely point at prose. Summaries are fine. Drafts are fine. The trouble starts the moment an agent touches a table.

There is a good reason for that. A hallucinated adjective in a two-page summary gets caught on the first read. A single row dropped from a 40,000-row extract, or a 0 quietly flipped to a 1, does not get caught at all. It flows into the model, the forecast, the submission, the board pack. By the time anyone notices, the decision has already been taken.

This is not a prompting problem. LLMs predict the next token. That is a superb mechanism for language and a poor one for arithmetic across 200 columns. They offer no guarantee that a transformation was applied consistently to every record, no guarantee that a join did not silently drop rows, and no way to prove after the fact where a figure came from. Probabilistic engines are the wrong tool for work where a single cell has to be right.

So we took the data out of the model.

The Structured Data Module gives Digital Workers a standard set of tools that operate on data outside the context window entirely. The model sees the schema and a sample of rows, enough to reason about what needs doing. It never holds or retypes the dataset. Every operation runs on files on disk, through code the agent writes and a judge agent reviews before it executes. The agent decides. The tools do the work.

That one architectural choice changes what you can promise. Because tools are the only route to the data, nothing gets silently missed or altered, and every value carries its lineage: the API or database it came from, and each transformation applied on the way to the final number. Explainability comes from the trace of what actually ran, not from a narrative the model writes about itself afterwards.

It also means you do not reshape your estate to suit the agent. Data arrives in whatever format you already have and the agent maps it at the start of the run. Enterprise connectors plug into the major platforms directly, and because the storage location is configurable, the working data can sit inside your existing environment instead of being pulled down for an agent to play with locally.

Accuracy on structured data is not a leaderboard metric. It is an audit question. Regulated teams do not ask whether a number is probably right. They ask who can show the working.

We built the Structured Data Module so the answer is the Digital Worker itself.