Before the model
The reliability of an AI system is decided mostly underneath it, by how well the data is structured. Part of the challenge in building these systems is keeping the margin for error as small as possible, by being intentional about how the LLM gets to the data, and at which stage. LLMs answer from the information they are given, so their answers are only as good as the data they receive. Even the best models today tend to guess where something is unclear, and a guess delivered confidently is worse than no answer at all.
The decisions
There are a few questions we ask ourselves about the data every time we build a new system. Can anything be read more than one way, leaving the LLM to guess or to choose between several “right” answers? Is the AI computing from raw data, or asking for an answer that has already been worked out? And what does the AI actually have in front of it when it answers?
Each question prevents a different problem. If definitions are left open to interpretation, an LLM can calculate profit margin one way on the first run and another way on the second. Both might be technically correct, but the board wants to see it one particular way, so the system has to write that definition down once and tell the LLM exactly which version to use. The same goes for words that sound obvious, like “customer” or “paid”, which often mean something different in each system a company runs.
The question about raw data is mostly about reducing room for error. If you ask an AI harness to calculate something complex from a raw dataset, it takes only one small wrong decision, a join or a filter, to throw off the numbers, and it makes those decisions again every time it is asked. Doing the arithmetic in deterministic code, tested once, and letting the AI fetch the answer is usually the better solution. Code does the sums; the AI does the part that needs judgement.
Context follows from the same idea. Loading a pile of tables into the model and hoping it finds the right rows is slow and expensive, and gets worse as the data grows. Querying a defined, organised layer brings in only what the question needs.
Storage, the decision people tend to spend the most time on, matters less to the AI than you would expect. Where the data lives affects speed, cost and how fresh it is, but it doesn’t make an answer right. Moving data onto a new platform moves it without defining it: a company with two meanings of “customer” still has two after the migration.
In practice
In practice this means a few simple habits before any AI goes on top. Pick a number on the dashboard and ask where it comes from; if two people trace it to different places, that is the work to do first. Give every definition one owner, whoever makes decisions with that number. And keep one rule for the AI itself: it reads from the cleaned, defined copy of the data, and acts through the system that owns the record, never by writing into the copy.
The AI is the last layer of the system, and usually the smallest part of the work.