Why AI gets your clients' marketing data wrong.

Point a language model at a client's marketing data and it will answer every question you ask. That is the problem. It has no way to tell you that the question could not be answered from the data it was given, so it produces a number in the same tone as a correct one.

A wrong dashboard looks wrong. Somebody notices a total that cannot be right. A wrong AI answer arrives as a sentence, in a chat window, with no total to check it against, and a sentence is the easiest thing to paste into a note to a client before anybody has checked it.

None of these are model problems. All four are properties of the data underneath, and each one produces a confident wrong answer rather than an error.

• One metric, three names

  Each platform names spend its own way. Asked for total spend, a model either picks one column or sums columns that are not comparable.

• Different types

  A number stored as text does not add up. It gets counted instead, and a count looks like a plausible figure.

• Different grains

  One table is a row per day, another a row per campaign per day. Joined without reconciling that, every figure is multiplied by however many rows matched.

• Mixed currencies

  Two platforms reporting in different currencies produce a total that is arithmetically correct and means nothing.

The instinct is to give the model more: the schema, a longer prompt, a description of each table. That helps a model write better queries against data that is already consistent. It cannot make two incompatible columns compatible, because the information needed to reconcile them is not in the data at all. It is in somebody's head.

This is why 'connect your data to AI' is the easy half and the half that gets sold. Making the data answerable is the work.

One name per metric, one type, one grain, one currency, across every source. That is the whole list, and it is what makes the difference between a model that can be trusted with a question and one that cannot.

SightShark does that modelling inside a warehouse you own, for every client, and then points the chatbot, the client dashboards and the alerts at the same tables. The chatbot runs on your own LLM key. The reason all three agree is that there is only one set of numbers for them to read.