> ## Content Index
> Fetch the complete content index at: https://www.slowin.ski/llms.txt
> Use this file to discover other available public pages before exploring further.

# AI reads the number, but who tells it the unit?
- URL: https://www.slowin.ski/data-contracts-ai/
- Published: 2026-11-03T09:00:00.000Z
- Updated: 2026-11-03T09:00:00.000Z
- Description: When a number looks wrong, people notice and complain, and that complaint has quietly enforced every data contract. An AI agent can only check what is written down, so the contract has to say what the data means: units, empty values, valid ranges.
- Author: Krzysztof Słowiński
- Tags: Fractional CTO, Data Contracts, AI Readiness, Data Quality, Data Architecture, Data Engineering

One of the data platforms I worked on processed laboratory results for a pharma R&D team, delivered by an external data vendor. The vendor had only ever sent one concentration unit, so the first version of our internal schema had no unit field; the unit was hardcoded in the processing. Then the unit changed (think nM to µM, a factor of 1,000), and the vendor published a new schema version with the unit as an explicit field. The change was not caught immediately, and adapting to it took a focused effort across every team responsible for the processing, end to end. Had our internal schema carried a unit field from version 1, the fix would have been mapping one new column.

Nothing in that dataset said what its numbers meant. The meaning lived in the processing code and in the experience of the people reading the results, and that usually works, because someone looking at a value 1,000 times too large knows it is wrong and says so. **Data contracts have always been enforced by complaints, not by their terms.**

An AI agent reading the same data does not complain. It ranks, summarises and answers in the same confident tone whether the unit is right or not. When the data consumer cannot complain, the contract has to do the complaining.

## The contract describes shape, not meaning

Most data contracts describe shape: column names, types, which fields may be empty. That is what tooling can check, so that is what gets written down. The meaning sits elsewhere: which unit a value is in, whether an empty field means unknown, zero or not applicable, which values are plausible, and which rows should never be counted. A person fills those gaps from experience. An agent can only check what is written down.

## Silence is not health

An AI feature usually joins an existing dataset as one more consumer, and the team keeps trusting the same signals: tests green, no tickets. But silence only means the reader who used to complain is no longer the one reading.

When a wrong answer does surface, the usual fix is a line in the prompt: "concentrations are in µM". That moves part of the data contract to a place the data producing team never sees and nobody versions, so the next change breaks it again, quietly. A semantic layer, the layer that defines business metrics, does not close the gap either: it says how revenue is calculated, not what an empty field means in a raw dataset.

## Someone has to speak for the agent

In a data contract, the consumer speaks for itself: it agrees the terms, hears about changes, and objects when something breaks. An agent cannot. The natural representative is the person accountable for the AI feature, but they are rarely in the conversation with the producing team, so the contract is agreed without its most literal reader.

## The leadership mandate

The CTO's job is to make meaning part of the data contract rather than something each reader brings. In practice, that is two moves: naming the AI feature owner as the consumer of record, so changes are negotiated with someone, and asking subject-matter experts and engineering teams which context future readers will need, not only what the data source sends today. The first version of a raw dataset is a starting point, not the target. A dataset that explains itself gets a new team member productive faster, and it is the only kind an agent can use well.

## The path forward

Good data quality tests would have caught the unit change, and they should. But a test can only enforce what someone has written down. For every dataset an AI feature reads, write down four things:

| What to write down                                          | What it prevents                                 |
| ----------------------------------------------------------- | ------------------------------------------------ |
| The unit of every measured value, as its own field          | A 1,000× error that passes every check           |
| What an empty value means: unknown, zero, or not applicable | An average that silently counts unknowns as zero |
| The valid range                                             | An implausible value reaching an answer          |
| Rows that must never be counted, such as test records       | Totals that include data nobody meant to count   |

Enforce them with tests in the data pipeline, so a change stops the data before it reaches an answer, and give the agent the same unit and range as context. A range check can still miss a tenfold shift that stays inside the plausible range. That is where expert experience belongs in the tests: compare each load with the previous ones and flag when typical values jump, the way a scientist notices a batch that looks unlike the last.

The earlier posts in this series covered the agreement between internal teams and the data vendor who never signed one. This one is about a reader that will never object. Write down what the data means when it first arrives, while the people who know it still remember.

[Get in touch](https://www.linkedin.com/in/krzysztofslowinski/?ref=slowin.ski)

## ****The data foundation briefing**

****Strategic insights on architecting AI-ready data foundations for engineering leaders.**

Subscribe 

Email sent! Check your inbox to complete your signup. 

No spam. Unsubscribe anytime.