How to Actually Start Using AI at Work

Leadership says "use more AI" and nobody tells you how. The easiest way to start is embarrassingly simple: pick a task you already do, upload the data, and ask AI what it can do for you. A low-stakes file and one question are enough to start. Everything else in this piece is that loop, run live on my own YouTube analytics, including the part where PowerPoint failed.

What should your first AI use case be?

Something safe. If your company has no strategy for what data may and may not go into AI tools, pick data with no personal information and no company secrets. Mine was analytics exports for my own videos: views, impressions, click-through rate, retention. Low stakes by construction.

Then one of two openers. Either the specific one, "here is the data, do you notice any patterns?", or the fully open one: upload any reasonable file and ask "what can we do with this?" Providing context and a starting point replaces the plan; you brainstorm the plan together.

Which tool, and why not the cheap one

I ran this in Cowork, inside the Claude desktop app. If your company gives you nothing, it sucks, but buy your own access, and do not start with a weak tool; in my experience Copilot mostly teaches people that AI does not work. Use whatever gives you the latest, best models: on the order of $20 to $100 a month.

The reason to start with the smartest model is calibration. You want to know what is possible before you work backwards to cheaper models and check whether the result is still good enough. Yes, many tasks can be decomposed so a dumb model handles each substep, but that requires an understanding of the process you do not have yet. Start smart, optimize cost later. A year ago I spent two or three months of evenings just playing with this; at some point extracting what you need from AI becomes second nature, and every month you wait makes the catch-up longer.

What the live run actually looked like

I uploaded five analytics exports and asked for patterns. It offered dashboards, reports, spreadsheets, insights, trend plots. I asked for PowerPoint first, and it struggled; the slides it eventually produced looked bad. That counts as a finding: slide design is currently a weak use case, while Excel, which barely worked three months earlier, now works well. Building this sense of which use cases are easy, which are reliable, and which tolerate error is most of what the experimentation buys you.

So: pivot to Excel, add the other four videos, and the multi-video analysis lands real takeaways. Which video pulled the most views, the best CTR, the best retention. Concrete instructions back to me: double down on the product-management topic, reuse the best-performing thumbnail style, fix the hooks, and note that four of five videos sit under 20% retention. Aggregation and data analysis is a prime AI use case; this took one upload and a few iterations.

The general principle underneath: AI is best at getting you to a draft fast. The first draft is the hardest part of anything; iterating on a mediocre draft is far easier than producing one from an empty page.

Where AI needs a human in the loop

Part of the skill is knowing which processes need checkpoints. Some decisions need 100% accuracy because the legal or business consequences of an error are unaffordable; those need a human in the loop or no AI at all. Between the extremes there are levels: AI as copilot, or AI doing all the work with you reviewing. Every one of those setups starts the same way, with experimentation on something harmless.

Turn the one-off into a repeatable skill

Once a workflow proves useful, say the magic sentence in the same conversation: "turn this into something repeatable." All the context is already there, including what failed; mine knew to make the skill Excel-only because it watched PowerPoint flop. The result is a skill, a markdown file with the repeatable steps, that you run in any fresh session. Hand it data, get the report, like passing the number-crunching to an intern whose output you review. From there, a scheduled task can run it for you every morning.

Nobody can build your intuition for you, and an afternoon of trying beats a month of LinkedIn takes. When a workflow sticks, turn it into a skill and let it compound.