When a team asks me where to start with AI, the conversation usually begins in the wrong place. They want to talk about tools — which model, which platform, which vendor. Those are real questions, but they come too early. The first question is simpler and harder: where does work actually slow down?
A bottleneck is where effort piles up. It’s the task that takes longer than it should, that blocks the next step, that someone has to context-switch into repeatedly throughout the day. Find that, and you have a real candidate for automation. Skip it, and you end up with AI that’s impressive but not useful.
Bottlenecks don’t announce themselves
The tricky thing about bottlenecks is that the people inside them often don’t describe them as bottlenecks. They describe them as just how things work. “We always spend an hour on that.” “That’s just the nature of the job.” “We’ve always done it that way.”
The most useful thing I do when scoping an AI project is spend time with the team before suggesting anything. Watch a few hours of actual work. Ask where people check out — the tasks they put off, the ones they apologize for taking so long. Those are the bottlenecks.
Once you name the specific task, the model question becomes much easier. You’re not picking between GPT-4 and Claude in the abstract. You’re asking: can this specific capability — draft a reply, extract this information, classify this input — be done reliably enough to replace a step in this specific process?
The model question has a short answer
For most operational work inside a Philippine SME, the model choice is not the determining factor. A well-framed prompt with a clear task and a human reviewer at the end will outperform a clever model with an unclear task and no process around it.
The things that matter more than model choice: how well the task is defined, how consistent the inputs are, how easy it is for a person to review and correct the output, and whether the tool lives where the work already happens.
Get those right and you can make almost any current model work. Get them wrong and no model will save you.