Internal data team or implementer: the real cost of each path

To have AI capability in a retail operation, there are three paths. Build an internal data team. Hire someone who implements. Buy a managed service from outside.

The question “which is cheapest” misleads, because it compares the wrong number. The cost of the license or salary is what shows on the budget. The real cost is time to first result and the dependency that remains afterward.

The three paths, as cost shapes

The difference is not in technical capability. It is in the shape of the cost and in who is responsible when the system breaks at eleven at night.

Build an internal team. Cost in senior salary, fixed and recurring, plus ramp time. A good data scientist takes months to learn the operation. The asset stays in house, and so does the risk: the key person leaves with the knowledge in their head.

Hire an implementer. Project cost, fixed by scope, with knowledge transfer at the end. Faster to first result, because it works from inside against your data. The internal team takes over afterward, formed by the project.

Buy a managed service. Recurring cost, predictable, no ramp. It turns on fast. In exchange, the operation stays dependent on outside for every change, and the knowledge never enters the house.

The decision tree

The deciding condition is horizon: who operates and improves this two years from now?

• If the answer is “an outside vendor, forever”, the managed service closes.
• If it is “my team, but I do not have it yet”, the implementer forms that team while delivering.
• If it is “my team, which already exists and has room”, building internally makes sense.

The trap in each path

The internal team fails when it is built before the first criterion, hiring for a target that does not exist yet. The managed service fails when dependency becomes hostage: every adjustment goes through a queue you do not control. The implementer fails when the criterion stays vague and the project stretches with no number that says it is done.

The defense, on all three, is the same: the criterion written before the code, and knowledge transfer as an acceptance condition, not a favor.

What the choice decides

Take the AI capability decision on your desk. Ask the team which path you assumed last time. Almost always nobody chose; it drifted to whatever felt safest, which is usually the most expensive over a two-year horizon.

Tell me the shape of the problem and the team you already have. In one hour I will send back which path fits, with no bias, with the cost of each over a two-year horizon, not on the quarter’s budget. If it is an implementer, the next conversation is the two-week Diagnóstico, which ends with criterion, timeline, and cost signed before the code.