John SorianoWhen a team asks 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, which feature list.
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 is the task that takes longer than it should, blocks the next step, or forces someone to context-switch repeatedly throughout the day. Find that, and you have a real candidate for AI assistance. Skip it, and you end up with something impressive but not useful.
Bottlenecks do not announce themselves
The difficult thing about bottlenecks is that people inside them often describe them as normal work.
“We always spend an hour cleaning that report.”
“Finance always waits on sales for the final numbers.”
“Support always has to read through the full history before replying.”
“We always check that manually because the system is not reliable.”
Those sentences are clues. They point to work that has become accepted instead of examined.
The most useful thing I do when scoping an AI project is spend time with the team before suggesting anything. Watch the work. Ask where people lose momentum. Look for the tasks they delay, apologize for, or quietly handle outside the official system.
Those are the bottlenecks.
The model question gets easier
Once the bottleneck is named, the model question becomes more practical.
You are not choosing between AI tools in the abstract. You are asking whether a specific capability can improve a specific step: draft a reply, classify a request, extract information, summarize a document, generate a first-pass report, check a record, or route a lead.
For most operational work in SMEs, the model itself is not the determining factor. A clear task, stable input, human review path, and measurable business outcome matter more than chasing the most impressive benchmark.
The things that matter more than model choice:
- How clearly the task is defined.
- How consistent the inputs are.
- How easy the output is to review.
- What happens when the tool is unsure.
- Whether customer or employee data is handled appropriately.
- Whether the tool lives where the work already happens.
Get those right and many current tools can be useful. Get them wrong and no model will save the project.
The better starting point
Start with the operational pain, not the AI capability.
If sales follow-up is slow, map the follow-up workflow. If support response time is slipping, map intake and routing. If leadership reporting takes too long, map the data sources and approval steps. If compliance review creates delay, map what needs review and why.
The right first AI project is rarely the flashiest. It is the one where the business can say, plainly: this step used to slow us down, now it does not.