When someone decides to become the AI person at their company, they almost always start with the task that annoys them most. The instinct makes sense, but the starting point is wrong. The most annoying task is rarely the one worth starting with: it tends to be irregular, touchy, or the kind where a single slip costs someone right away. The value isn't in the building anyway; it's in what you decide to build. I unpack that separately in the piece on deciding what's worth doing with AI at all.
A better start is quieter. You begin at your own desk, with work you're already responsible for, you gather proof in numbers, and only after the first wins do you reach for a real constraint on the business. It's a gentle on-ramp to the same role I describe from the constraint side in the piece on the in-house AI person: there you walk straight in through the bottleneck; here you build the road up to it from your own work. Why the window for this role is open right now I explain separately in the piece on the overlooked AI career opening for 2026. Here I'll show you the move itself: four steps that earn you the role from the ground up.
Listen
A ~5-minute conversation about this article — why you start with your own tasks, not the company's biggest problem.
Generated by NotebookLM, reviewed by AURA.
Start at Your Own Desk
The first step is an audit of your own work, without looking at the whole company. Sit down and write out what you actually do in a week: not the big projects, but the repeat tasks that come back every Monday. Then run that list through a simple filter with two conditions.
A task is fit to start with if it meets both at once. First, it eats real hours every week. Second, it's low-risk when the AI gets it wrong: you stay in the loop, check the output, fix it and move on, and nobody pays for the mistake. One condition without the other isn't enough. A task that burns time but where a slip is costly, you set aside for later. A low-risk task that takes you fifteen minutes a month gives you no proof at all.
What usually gets through this filter? The weekly status report, meeting notes, tidying your inbox, cleaning up data in a spreadsheet, a bit of research behind a single decision. These are deliberately small annoyances, not the company's strategic bottlenecks. You don't take on the whole company at once: you start at your own desk, because here you have full control. You don't have to ask anyone's permission, and every error stays between you and your task. If you're after the broader key to which kinds of task are worth automating in the first place — the recurring, valuable ones — I lay it out in the piece on what businesses really want from AI today.
Gather Proof in Numbers
Step two: you take the tasks from the top of the list and actually automate them. This isn't about scale yet, only about one thing: proof that it works. And proof means numbers, written down before you forget them. "This report used to take me two hours a week; now it takes ten minutes." The hours you win back this way are your proof; without a record, all that's left is a vague sense that things are "somehow faster."
Log it from day one: the task name, time before, time after, and exactly what the AI does now. It's a dull register, and that's exactly why it works: you have hard data ready for the moment you need to talk about it. There's the same logic here as with a coach who gets into shape himself before he starts training anyone else. Nobody buys advice from someone with no result of their own; you show the effect on yourself first. And measure honestly: count the whole task including the check on whatever the AI returns, not just the moment you click "generate." An inflated number will crack at your manager's first question anyway.
When the first task clicks into place, you don't stop. You go back to the list and take the next one, then another. Every closed task is a new line in the register and more hours won back. Something more valuable than the automations themselves grows alongside them: a concrete, countable trace of what you can do. That trace, not the list of tools on your CV, is what convinces people. The same mechanism — that what you show matters, not how many courses you've passed — I lay out in the piece on what you've actually built: there from the side of work visible on the outside, here from the side of quiet proof inside your own company.
Show What Works
Step three: you make that proof visible. The register alone won't convince anyone while it sits in your folder. Show the results at a team meeting, briefly, on a single example. Offer to solve a colleague's most tiresome task — the one they were complaining about earlier. And document everything along the way.
How you talk about it weighs as much as the result itself. You don't walk into the meeting with "I used an AI chatbot for this." You speak the language of business: "this saved us eight hours before the quarterly report." The first version sounds like a curiosity, the second like something your manager will remember and repeat. The difference is simple: one describes the tool, the other the effect on the business.
Along the way, collect your best prompts and workflows in one internal document the whole team can use. Write them up so someone without your practice can reproduce them in a few minutes, because getting others to adopt these solutions is a separate job, often harder than building them in the first place. Once others start using them, your name is on something their daily work rests on. From the person who "does something with AI," you become the point people come to on their own.
From Small Annoyances to Bottlenecks
Once you have a few wins and people start coming to you for help, you're ready to move up from small annoyances to constraints. That safe, low-risk start was the right move: it gave you a place to experiment, learning and still gathering proof. But automating annoyances doesn't grow the business. What grows it is going after the constraints, and that's what really gets paid for.
Now you repeat the audit from the first step, only across the whole company, and the question changes. It's no longer "what annoys me" or "what eats a few of my hours." It's: what actually holds the business back — concretely, if a flood of new customers showed up tomorrow, what would break first? That question, and the way it pins down the bottleneck, I lay out in the piece on what businesses really want from AI today. Where the answer lands is your first real project. Saving the team a few hours a week makes you helpful; removing a bottleneck the whole company snags on makes it money. That's an entirely different conversation with your manager.
Step four: you turn all of it into a position. You add up all the hours and money your automations have saved and turn them into a single number. Something like: "through these automations I'm giving the team back the equivalent of one full-time hire a year." A number like that gives the company a real reason to carve out a budget for AI projects. As long as your work is described as "helping here and there," it's hard to assign it a budget; one number turns scattered favors into a line item you can plan around. You take it to your manager and you don't ask for a favor: you present a specific role, with a title, built on what your register shows. You don't ask for the role. You show that you're already doing it.
The Takeaway
The whole path fits in one sentence: you don't win an AI role by declaring you're good at it, but by stacking up proof until the role becomes obvious. You start at your own desk, because there a mistake is cheap and the win is yours to count.
If you work in a regulated industry, the same move applies with one caveat: don't point AI at sensitive data and don't automate anything without sign-off. You can still be the AI person, experimenting after hours on stand-in data; more on that situation in the piece on the overlooked AI career opening for 2026.
One move for today: write out your week and mark the two tasks that eat hours and where nothing bad happens if the AI slips once. That's your starting list.