Data first, context second, AI third.

Take the screens away from an enterprise system, the menus, the forms and the reports, and what remains is data. Every order, payment, shipment and complaint the company has handled is in there. It is the most valuable thing most companies own and the thing they look at least.

Starting at the wrong end

Many AI projects start with the model. Someone chooses a language model, connects it to a few documents and builds a chat window. It demonstrates well. Then a manager asks it a real question, about a real customer or a real shipment, and the answer is confident and wrong.

The model is rarely the problem. It can read the word account. What it cannot know is that in your business an account belongs to a customer, is opened at a branch, has a status and may be frozen. Nobody has told it.

The order I work in

My position is deliberately narrow, and it has an order.

1. Data. What the ERP or core system already records. This is the anchor.
2. Context. Entities, relationships and attributes: the things the business deals in, how they connect, and what is known about each.
3. Language model. It reads and reasons over the context instead of raw tables.
4. Agents. They carry out defined tasks using what the model understands.

The context layer is where the business is written down.

It is the step that gets skipped, because it is slow and unglamorous. It also decides whether the last two are worth anything.

Effectiveness before efficiency

I am often asked how much time AI will save. I would ask a different question first. The first gain is effectiveness: a better decision about which customers to keep, which stock to move, which risk to look at today. Efficiency comes later, and it needs a volume of data that most companies have not yet organised.

For the same reason I do not count routine workflow automation as AI. If a fixed rule can do the job, a rule should do it. It will be cheaper, and it will be right every time.

Where I learned it

In 2018 I was IT & Big Data Director at a durian plantation. We put in a plantation management system, sensors in the field and big data analytics. It was my first lesson in turning operational data into decisions, and what I took from it was the order: the operation had to be recorded properly before any analysis was worth reading.

The tools are far more capable now, and the order has not changed. Data first, context second, AI third. At the end, a person signs off.