Where to actually start with AI in your company
Most companies stall on AI because they start with the tool, not the work. Here's a calmer, more useful place to begin.
Most companies don't struggle with AI because the technology is too hard. They struggle because they start in the wrong place — with a tool, a demo, or a mandate to "use AI" — instead of with the work their business actually does every day.
The result is familiar: a few enthusiastic experiments, a subscription or two, and no real change to how the company operates. The tools work. The adoption doesn't.
There is a calmer place to begin.
Start from the work, not the tool
Before evaluating any AI product, write down the tasks your team repeats most often. Not the strategic, once-a-quarter decisions — the small, frequent, unglamorous work: sorting incoming requests, drafting the same kinds of replies, moving data between systems, checking documents against rules, answering the questions customers always ask.
These are where AI creates value first, because they are frequent enough that small improvements compound, and structured enough that a system can handle them reliably.
A generic chatbot rarely touches this work, which is exactly why generic tools so often stall. The value lives in your processes, not in the model.
Look for the three signals
A task is a good first candidate for AI when it shows three signals at once:
- It repeats. The same shape of work happens many times a week.
- It has a pattern. A person could write down roughly how they do it.
- A mistake is recoverable. Getting it wrong occasionally costs time, not trust or safety.
Start where all three are true. Avoid, for now, the high-stakes decisions where a wrong answer is expensive or hard to undo — those come later, once you trust the system and have the checks in place.
Keep a human in the loop, on purpose
Early on, the goal isn't to remove people. It's to remove the repetitive part of their work so they can spend attention where judgment matters.
The most reliable pattern we see: the system does the first pass — drafts, sorts, flags, prepares — and a person approves, edits, or overrides. That single design choice buys you speed without betting the business on a model being right every time. As trust grows, you widen what runs on its own.
Measure one thing before, and the same thing after
Adoption fails quietly when nobody can say whether it worked. Before you change anything, pick one honest number for the task: hours spent per week, response time, error rate, backlog size. Write it down.
Then, after the system is in place, measure the same number. If it didn't move, you learned something cheaply. If it did, you now have a concrete result to build the next step on — and to justify it to the rest of the company.
Build for your business, not around a template
Off-the-shelf tools assume every company works the same way. Yours doesn't. The workflows, the data, the exceptions, the way your people actually get things done — that's the context a system has to fit into, and it's the part a generic product can't see.
This is the heavy, unglamorous work of adaptation: connecting AI to real processes, handling the edge cases, and deploying something that holds up on a Tuesday afternoon when the inputs are messy. It's also the work that separates a lasting result from a promising demo.
You don't have to carry that part alone — but you do have to start it in the right place.
The first step, in one sentence
Pick one task that repeats, has a pattern, and is safe to get wrong occasionally; measure it honestly; put a system on the first pass with a person in the loop; and check whether the number moved. That's a real beginning — and everything larger is built on it.