AI in practice: start from the problem, not the model
Most AI initiatives start from the tool. The right starting point is older and far less glamorous: the business decision you want to improve.
Most AI initiatives that fail start in the wrong place. From the model, from the trendy tool, from the vendor who showed up with a demo. The right starting point is older and far less glamorous: the business problem.
AI is not a goal. It is a way to improve a decision that already exists in your operation. When the starting point is the technology, you end up looking for a problem to justify the purchase. When the starting point is the decision, the technology becomes a consequence.
Three questions before any pilot
Before you buy GPUs, hire a vendor or open a project, answer three questions. One sentence each.
- Which decision will this AI improve?
- What does getting that decision wrong cost you today?
- Who owns the process once the AI is in place?
If you cannot answer all three clearly and specifically, it is not time for technology. It is time to go back to the table and understand the process better. That step back costs one conversation. Skipping it costs the whole project.
"Use AI in support" is not a goal
Look at the difference between two sentences I hear all the time:
"We want to use AI in customer support."
"We want to cut the average first-response time without dropping customer satisfaction."
The second one is measurable. It says which decision changes (how and when to respond), what the metric is (response time) and what the constraint is (satisfaction). The first one is a LinkedIn post. Almost every good project starts when someone rewrites the first sentence into the second.
Reversible, repetitive, with data and measurable
A simple filter for where to apply AI first. Prioritize processes that are:
- Repetitive, because the gain multiplies by volume.
- Reversible, because the cost of a mistake is low and you learn fast.
- Backed by data, because the model needs something concrete to lean on.
- Measurable, because without a metric you cannot tell if it worked.
Support, document triage, marketing, back office and internal analysis usually hit all four. Irreversible decisions, high regulatory risk, or anything that affects people directly go in the other bucket: caution, human oversight, no rush.
The intuition behind the models
You do not need to know how to train a neural network to make good decisions about AI. But understanding the intuition helps you separate promise from reality. For the visual explanation of how a network learns, 3Blue1Brown does it better than any slide:
The takeaway
Good technology applied to a poorly defined problem is still waste. Start from the decision, write down the metric before turning on the model, and pick a first case that is small and reversible. The rest gets easier.