Enterprise AI work was mostly data cleanup, not model choice

A 3-month pilot built a conversational on top of internal company data. The goal was to let and sales teams ask complex questions and get accurate answers. A working demo was ready in the first week, but turning it into something reliable for real work took the next 80-plus days.

The main lesson was that the model was not the hard part; the and data quality mattered more. An LLM can only answer from the context it receives, so fragmented or outdated text leads to bad answers even with a strong model. About 5% of the work went into connecting the LLM, while about 95% went into data engineering.

Company documents looked clean at first, but embedding them exposed several versions of the same client contract across different drives, including draft versions.

Key points

  • A first demo was built quickly, but took most of the 3-month pilot.
  • The assistant was meant to answer complex questions from and sales teams.
  • Model took little time compared with data engineering work.
  • Bad context made even strong models produce poor answers.
  • Duplicate and draft contract versions across drives created problems.
Read original