There is a version of AI adoption that goes: buy the licences, connect the tools, tell everyone to use it, wait for the productivity. It generates a lot of activity and very little operational change, and the reason is almost never the model.
A language model is very good at reasoning over what you give it, and completely indifferent to whether that thing is right. Point it at a warehouse where “active customer” means four different things and it will answer confidently, four different ways, and none of the answers will arrive flagged as suspect. You have not bought intelligence. You have bought a confident guessing machine — and one your team will believe for a while.
What an agent actually needs
When we put an agent into an operation, it needs three things that have nothing to do with AI:
- Definitions written down as code rather than held in someone's head, so the same question always resolves the same way.
- Events with reliable timestamps, so “what changed this week” is a query and not an argument.
- A boundary around what it can see, so personal data does not end up in a prompt because nobody thought about it.
With those, an agent is reasoning over facts and becomes genuinely useful. Without them, every output needs a human to check it — which is exactly the cost you were trying to remove.
The order matters more than the ambition
We build the data layer first, every time, including when the client came to us for the AI. That is sequence, not caution. The layer is what makes the agents trustworthy, and it is also the part that keeps paying off when you change your mind about which model or which vendor to use.
Models will keep changing. A clean, well-defined event history will not.
Then put your people where it counts
The point of the agents is not headcount. It is that repetitive work stops consuming people who are good at judgement. Someone spending their week pulling reports, chasing exceptions and re-keying between systems is someone not spending it on a customer who needs a decision made properly.
So the sequence is: quality data first, then agents on top of it, then your people deployed where a human genuinely does it better. That last part is not the consolation prize. It is the whole return — nobody ever won a customer because their reporting reconciled.