How to Start Using AI at Work When You're Overwhelmed by the Tools

Every week, there’s a new agent (an AI that works through a multi-step task on its own), a new model, a new benchmark chart, and tooling that keeps getting better.

Every week I also catch myself asking the same questions.

Do I use Cursor, Claude Code, or Codex? Do I use Astra, Gemini, or Fable for this?

Should I try local AI, and if I do, how the heck does that even work?

I build software for a living, and those questions still leave me overwhelmed by AI some weeks.

So when a friend of mine got stuck on a work task and couldn’t figure out where to begin, I understood her a lot better than I wanted to. Watching her get stuck made me notice how I get unstuck myself.

These days, when AI feels like too much, I skip the tool question and start with one task I actually need to finish. I write down what the finished result should look like, then list what I have and any limits I’m working under, like sensitive data or the tools my workplace approves.

Then, before the AI does any work, I ask it what information it’s missing and what steps it would take. I try the first small step, and when it’s done, I compare the result to what I wrote at the start.

I’ll walk you through what happened in a group chat of friends, what I think I got wrong while trying to help, and how you can use the same starting point when AI feels like too much.

Why Getting Started With AI Feels Overwhelming, Even for Developers

Look back at those earlier questions. Each one is about a tool, and none says what I’m trying to get done.

The tools keep multiplying, and most of the advice out there assumes you already know which one you want, so you end up comparing options before you’ve decided what you need 😬

Say your manager asks for a summary of this month’s customer feedback.

Asking which AI tool should read it feels like the first question, but the first question is what your manager needs from that summary: a list of complaints, a decision, or a one-paragraph update.

Once you know that, most tool choices shrink on their own.

I think that’s why so many people barely use these tools in their everyday work, even though the tools are right there.

They don’t know how to fit them into what they already do.

So it’s unsurprising that a common question I run into when browsing online is: what do I learn first if I’m not technical?

I’m not the only one seeing it either, since one writer who covers AI says it’s the question she’s always asked online.

And honestly, it sounds extremely familiar to what happened in my group chat.

Overwhelmed by AI at Work: A Friend, Sensitive Healthcare Data, and an Excel Sheet

Said friend works in healthcare and needed to put together an Excel sheet for work. She doesn’t know how to use Excel that well, and the data she’s working with is sensitive.

Before anyone points to the elephant in the room, she is very well aware that AI exists. The problem is that she had to find a way to use whatever tools were available to her within the constraints of her task.

And she didn’t know where to start or how to carry any of it out.

She asked in our group chat for general advice, and everyone chipped in. The first suggestions were the obvious ones: upload the files to ChatGPT or Claude (aka “cloud models” that run on a company’s servers instead of your computer) and ask it to make the spreadsheet.

She pointed out that she can’t upload the data (serious props to her for being conscious of privacy and data sharing).

So someone shared a full prompt with placeholders that asked her to describe the layout of the sheet she wanted so the cloud model could build it from that description instead of the data itself.

I suggested she check whether her employer pays for Copilot, in case it shows up as that little chat bubble at the bottom of the Excel window. It didn’t, so that route was out.

She said she’d try the prompt and came back overwhelmed again, which made it twice that she’d felt too overwhelmed to keep going.

That’s when I noticed the hard part wasn’t only Excel 💁‍♀️

The prompt asked her to describe how the finished sheet should look, but her data wasn’t structured in any particular way yet.

Local AI vs Cloud AI for Sensitive Data: The Hybrid Approach I Suggested

My follow-up suggestion to what I perceived as the underlying real issue was Ollama, a free tool that lets you run an AI model on your own computer. That’s what people mean by local AI.

A local model lives on your machine, not on a company’s server. To help her choose a model, I told her to Google something like this:

“best ollama free local model for understanding and working with data for [your computer memory, like 24GB]”

Then I second-guessed myself.

Ollama isn’t the friendliest place to start if you’re not technical, so I brought up LM Studio, which is considered more beginner-friendly.

I told her tools like these are great for sensitive data because the data stays on the computer and never goes to an outside server.

If you download and run a model locally in LM Studio, none of your messages, chat histories, or documents leave your computer, according to its privacy policy. That covers running a model locally.

Note 👀
LM Studio also offers cloud features like web search and cloud models, and those do send your requests out, so I’d leave them off for this kind of work. The same thing can be seen on Ollama’s policy, but check to ensure it complies based on any unique needs. Your workplace’s rules about installing software and handling that data still apply too, and I don’t know what hers are.

Explore: Let’s Build An Insanely Useful Local AI Agent With Docker

Why I Thought a Local Model Could Help Her Find the Data Structure

Before anyone can ask ChatGPT or Claude to build a spreadsheet, somebody has to decide what goes in the columns.

Say you have a stack of sign-up forms from an event and want them in a spreadsheet. You first decide what you want from each form: name, email, which session they picked, maybe a note.

Each of those becomes a column, and that layout is the data structure. My friend didn’t have hers yet.

So my plan went like this:

She doesn’t have to give the local model all of her data, because a small sample is enough for it to help her work out the structure.

Then she can ask it to write a prompt for that structure, and maybe generate some dummy data (made-up values shaped like the real ones) to go with it.

She takes the structure and dummy data to the cloud model and asks it to build the layout of the sheet. The real data stays on her computer, and only the dummy version goes to the cloud.

She hasn’t replied since I sent that. I don’t know if she can pull it off, or if she’s even tried.

Honestly, I worry my breakdown made things worse. I was trying to give her a clear path, and I may have handed her local AI, data structures, and a two-tool workflow when she was still stuck on where to begin.

Add the suggestions from the rest of the chat, and it may have felt like a pile of homework 🙈

How to Start Using AI at Work When You Don’t Know Which Tool to Use

She isn’t the only person this happens to, and you don’t need a technical background or my hybrid setup to get started.

I suggested it because she has sensitive data she can’t send to a cloud model.

If your task doesn’t carry that restriction, you can skip the local part entirely.

To start using AI at work when you’re overwhelmed:

Pick one task
Write down what you need at the end and what limits you’re working under
Ask AI what’s missing
Take one small step
Check the result

Each step below shows what it means in practice, using my friend’s situation where it fits.

1. Write Down What You Need at the End of the Task

Start with one sentence a coworker would understand.

For example, “A one-page summary of this month’s customer feedback, grouped by complaint, for my manager” tells an AI far more than “help me with feedback.”

For my friend, this was the step that was missing. She knew she needed a sheet, but she hadn’t decided yet what it had to contain.

If you can’t say what done looks like, how would any tool know? (It can’t read your mind… yet.)

2. List What You Have and the Limits You’re Working Under

Maybe you have a pile of PDFs, or notes in no particular order.

Maybe the sheet has to follow a certain format, the data can’t leave your computer, or your workplace only approves certain tools.

Put all of that in writing before you open a single tool.

3. Ask AI What’s Missing Before It Does Any Work

Though it’s tempting to jump straight to asking for the finished result, ask for the steps, assumptions it’s making, and information it’s missing first.

This is also how you handle not knowing how to prompt, because you don’t need the perfect prompt when you can write one that makes the AI ask you questions.

Try something like this, but treat it as a pattern to adapt and not a script to copy:

“I need to [what you need to produce]. I have [what you have, described in general terms]. I’m not sure how to approach this. Before you suggest a tool or give me a result, ask me whatever you need to know, tell me what information is missing, and lay out the steps you’d take.”

Notice how the prompt doesn’t ask AI to pick a tool or hand you the final result.

It asks for questions first, which also takes the pressure off choosing a tool on day one.

In my friend’s situation, a question like “What columns should this sheet have?” is the kind of gap this prompt is built to surface.

Each part of that prompt has a job, and once you know the jobs, you can write your own version for any task.

“I need to [what you need to produce]. I have [what you have, described in general terms].”

Saying what you need to produce gives the AI a destination, and describing what you have gives it a starting point.

Tip: The “in general terms” part keeps sensitive details out of the chat!

“I’m not sure how to approach this.”

Admitting you’re not sure how to approach it gives the AI permission to ask questions instead of guessing.

I cannot stress how important this is. The quality of the response that I get when AI is able to ask those clarifying questions that it has, whether or not it prompts you for them, is considerably better than what I would get had I not answered.

Tip 💯
If you constantly hear mentions of “context”, but you find yourself confused on what that context is supposed to be, or you find yourself rambling too much instead of providing the structure, direction, and information that the context should really provide, then a way to bypass that is to have AI ask you questions because, in asking you questions, it is building that context.

“Before you suggest a tool…”

Asking it to hold off on suggesting a tool or giving you a result stops it from sprinting to an answer before it understands the task.

And asking what’s missing, plus the steps it would take, turns the chat into an interview, which is the actual skill here.

When the reply comes back, look for questions about the format you need, who the result is for, and the information you have.

If it jumps straight to a finished result anyway, tell it to stop and ask you questions first. For all you know, you might actually need to interrogate and question whatever approach or steps it comes up with.

Nothing about the wording is magic. If you can get an AI to question you before it starts working, you can rewrite that prompt in your own words for any task, and your version will probably fit your situation better than mine.

Tip 👇
Describe your data in general terms instead of pasting sensitive details into a tool your workplace hasn’t approved. If your work involves sensitive data, check what your employer allows before you put any of it into any tool.

4. Take the Next Small Step With a Tool You Can Already Use

Pick one step and do only that. If the first step is “decide what columns the sheet needs,” do that and stop there.

You don’t need an agent, local AI, or three tools chained together for it.

Those can be great for situations like my friend’s, but they’re options and not requirements.

If you have more than one tool available and can’t tell which to open, three questions usually narrow it down for me:

Is any of the data sensitive? If it is, stick to tools your workplace approves, and describe the structure of your data instead of pasting the data itself (more on that below).
Which tools does your workplace actually approve? If it approves only one, that’s your tool, and the decision is already made.
Does the AI need to read a file, or can you talk the problem through? If it has to read a file, you need a tool that accepts uploads, and only if the data allows it. If you’re working out steps or a layout, a plain chat is usually enough.

5. Check the Result Against What You Wrote in Step 1

Don’t take an AI-built sheet as final. Spot-check a few rows against your source (this is your so-called human-in-the-loop).

Anything that affects a decision, a person, or money deserves a closer look before it goes anywhere. Automate repetitive, non-critical tasks only (AI or not).

Tip: You can correct a vague first answer by adding more detail about the task and asking again. This is called iteration and is what makes a tool useful and accurate (to a degree). Expecting a perfect result on first try is called “one-shot” and often results to more bugs and inaccuracies than you’d ever have imagined.

Remember to treat this as a flexible starting point. Some tasks will need more steps, and some will need a person who knows the subject to look at the result.

What to Do When You Can’t Upload Your Work Data to AI

Most advice about using AI at work quietly assumes you can paste whatever you want into a chat box.

My friend couldn’t, and anyone handling sensitive data may be in the same spot.

The first place to look is whatever your employer already provides. If Copilot in Excel is on the list, that’s the easiest route, and it’s why I asked my friend to check for that little chat bubble.

When there’s nothing like that, you can describe the structure of your data instead of pasting the data itself. That’s what the prompt someone shared in the group chat did by asking my friend to describe the layout she wanted, so the cloud model could build the sheet without ever seeing the real data.

The catch in her case was that she hadn’t decided on the layout yet, which is why working out the structure comes first.

If you need to show an AI what something looks like, dummy data does the job. Made-up values shaped like the real ones give the cloud model enough to work with, and nothing sensitive goes along with them.

Local AI is the last option I’d consider, and only if your workplace allows it. This way, you’ll have a local model on your computer to help work out the structure from a small sample, with only the structure and dummy data going to the cloud.

Note: Software installs and data handling fall under your workplace’s rules. If, like me, you’re on VPN with tight security, even an app like Ollama may not be in the allowed list. For those with more freedom, device restrictions may also limit your access to the “best” local AI models (which will affect your result). However, local AI is a rather phenomenal area with lots of possibilities so I recommend you tinker with it either way.

If none of these fit, the answer may be a conversation with your employer about which tools are approved, not a workaround (especially my techies 😏).

What to Learn First When You Want to Use AI: Breaking an Unfamiliar Problem Into Parts

What I do for a living is put things together from whatever’s available. Sometimes the starting point is a pile of data and a set of requirements. Sometimes it’s something blurrier, like “I want to do X.”

Either way, I break the problem down and work out which approaches could work.

I often combine a few of them, based on what the tools can do and what restrictions I’m working within.

My HOA Connect project is a good example. I was building a portal and dashboard for my HOA, which meant working with sensitive financial data, including things like account numbers.

I didn’t want any of that reaching a model it shouldn’t. I was also limited by my machine, which only supports local models that aren’t great at pulling data out of PDFs with complicated layouts because of the size of the model I could load.

Instead of fighting with that model, I split the work into two parts:

A GitHub Actions workflow (an automated job) runs a Python script that extracts the data from the PDFs and structures it. The script is deterministic, which means the same PDF produces the same output every time, with no guessing involved. This is plain old code; no AI.
Once the data is structured (excluding account numbers and other sensitive, irrelevant details), I pass those to the local model, which acts as an assistant that can read the data and answer questions about it.

This way, the model never has to deal with the complicated PDF layouts because the script has already turned them into a format it can handle and expects. (This is data normalization, and it came out of trying and failing to have a rather small local model handle the PDFs 😬)

My friend’s problem has the same shape but with different details. She has data in no particular layout; she can’t send it to a cloud model, and she has to figure out which tool handles which part.

For a problem like that, I don’t have to start by picking an AI model.

Note: I used to think of this as plain old work and not a skill. Watching my friend stop on a problem I would’ve started breaking down without thinking revealed to me that it’s worth more than I’d given it credit for.

Use AI to Understand the Problem, Not to Skip Understanding It

The tools will keep changing, and the ability to work through a problem you haven’t seen before carries over to whichever tool comes next.

It’s a smaller version of an argument economist Daniel Susskind makes about education, stating that people should learn to use AI effectively and critically in every profession instead of trying to predict which jobs will exist.

Note how none of this means you should do everything by hand.

AI can help you break a problem down, point out what’s missing, and teach you the steps along the way, and that’s a good way to approach any unfamiliar task with AI.

If you’re not using it to some extent, you’re actually wasting time.

The line I care about lies between using AI to understand a problem and using it to skip understanding the problem altogether.

Guess what? If you can’t tell whether the spreadsheet is right, you have no way to catch it when it’s wrong.

AI can give you access to capabilities you don’t already have. But how far those capabilities take you depends on whether you can define the problem, work out what you need, ask useful questions, and judge the result.

It’s a Wrap

If you take anything away from this post, let it be to pick one task and start there.

I don’t know how my friend’s story ends since she’s been quiet, but I’m hoping she gets the sheet done.

What I took away from this particular instance was that a complete plan isn’t automatically helpful for someone who has already stopped twice.

Much of the overwhelm surrounding AI tools, part of what makes it feel like too much right now, is easily solved by skipping the two comparisons.

Pick one task you actually need to finish, write down what the finished result looks like, and ask an AI to help you figure out what’s missing. That’s enough for a first step.

Keep learning and exploring 👋

AI Career Local AI Workflow

How to Start Using AI at Work When You’re Overwhelmed by the Tools