AI without the enterprise circus

I have sat through AI strategy decks that listed twenty use cases and zero tables. Then someone bought a copilot seat for people who still export CSV to email.

Atalaya does not sell "AI transformation." We sell a thin slice on data you can name.

Start from a decision

The question is the same as a board: "Are we late on line 3?" "Did consent expire?" "Which dealer is sitting on stock?" If you cannot write the question, you are not ready for a model. You are ready for a workshop.

We wire the slice to the warehouse first. Features are columns. Labels are a table with a date. Training is a job with a watermark. Inference writes back to a table the board already reads.

No new silo. If you fire us, you still have the features.

Legal basis is a feature

Healthcare, HR, anything with a person: if you cannot point at purpose and retention, we do not train. An LLM that summarizes clinical notes without a basis is not innovation. It is a reportable event.

We treat training extracts like any other pipeline: consent join, retention clock, deletes that actually delete.

Small models, owned runtime

A lot of "AI" on a plant or a dealer network is a classifier or a forecast on tabular data. That runs in a Cloud Function or a batch job. You do not need a GPU reservation and a vector database to start.

Use an LLM when the input is messy text or the output is a draft a human will edit. Use a regular model when the input is already fct_. Mixing them because the budget said "AI" is how you get neither.

The circus to skip

A center of excellence before the first slice. A semantic layer purchase to "prepare for GenAI." A six-month vendor POC that never reads production.

Discovery, one pipeline or one board, then a model on that grain. If the number does not move, we stop. That is the whole program.