Getting Started with AI in 2026: A Practical Guide for Professionals

Most people I talk to don’t get stuck because AI is hard. They get stuck at the start: twenty tools, a thousand opinions, and a plan to figure it out “properly” later. Later never comes.

Give this guide about an hour, and do the exercises as you go. You’ll learn one idea that explains most of what AI does, then run six short exercises that let you watch that idea play out on your own screen. By the end you’ll have briefed an AI properly, steered a weak draft into a good one, and very likely caught it being wrong or backing down under pressure. Those four experiences are most of what separates people who get value from AI from people who gave up after a week.

You need one free account and some material you’re allowed to share. That’s it.

If your company already picked your tool for you

Much of the advice out there assumes you can freely choose your AI tool. At work, many people can’t. ChatGPT and Claude are blocked (or the use is prohibited), and what you got instead is Microsoft Copilot or an internal chatbot, plus a training course or two on the intranet. If that’s you, you’re in the most common situation there is.

Every exercise below works in Copilot or an internal chatbot, because none of it depends on a particular tool. So run two tracks. Use the tool your company provides for work material, and a free private account for everything private. What you must not do is quietly paste work documents into a private account because the sanctioned tool annoys you. That shortcut is how client data leaks and how people get fired.

I’ve written a full guide for exactly this situation: Your Company Blocked ChatGPT. Now What?

The one idea that explains most of it

You don’t need to understand the technology to use AI well. You do need a rough picture of what you’re talking to, because without one, every surprise feels random.

An AI assistant like ChatGPT or Claude learned from an enormous amount of human writing. From that, it learned to produce the most plausible response to whatever is in front of it. When you type a message, it writes the answer that best fits everything in the conversation so far: your words, the documents you pasted, the earlier back-and-forth, and any standing instructions you’ve set.

It doesn’t look things up in a database. Many assistants now decide on their own to search the web, and when they do you’ll usually see the sources listed; otherwise the answer comes from what it absorbed during training. It has no reliable sense of when it’s wrong. And it knows nothing about you, your company, or your situation beyond what you tell it. Reasoning step by step and searching catch many mistakes, but the core is still the same: it writes the answer that fits best.

Once you hold that picture, four behaviours stop being mysterious:

1. A vague request gets an average answer. If “write an email about the meeting change” is all it has, the most plausible email is the generic one: the one that fits any company, any team, any meeting.
2. More relevant context gets a sharper answer. Every detail you add changes what “most plausible” means. Tell it who the team is and why the meeting moved, and the generic email is no longer the best fit.
3. The same question can get different answers. There are many plausible answers, and it doesn’t pick the same one every time. For writing, that’s useful. For facts, it’s a warning.
4. Sounding right and being right are different things. A made-up statistic can be just as fluent as a real one. Fluency is what the system is best at, so it tells you nothing about accuracy.

The practical version of this is a picture I use constantly: AI is a brilliant new hire on their first day. They know a remarkable amount about the world in general and nothing about your business. Hand a new hire a one-line task and they’ll produce something competent and generic. Brief them properly and they’ll surprise you. And like any new hire, they sometimes fill a gap with a confident guess instead of asking.

The exercises below make each of these behaviours visible.

Set up in ten minutes

Pick one assistant. For most professionals I’d start with Claude. It’s my first pick for writing, reading, analysis, and thinking through a problem together. ChatGPT is a close second and has the widest range of features in one place, including web search, file analysis, voice, and image creation. Gemini or Copilot is a sensible choice if your employer already provides and approves it. If your employer made the choice for you, skip the comparison entirely.

Stay with it for two weeks. Ignore the model picker and use the default. Switching tools feels like learning because every interface is new, but it stops you from building an instinct for how this one responds to you.

Learn three buttons. Find where you start a new chat, where you type a follow-up in the current chat, and where you upload or paste a document. Stay in the same chat while you’re working on one thing, so the AI keeps everything you’ve told it. Start a new chat when you switch to something unrelated.

Add a short profile. In Claude, look for Instructions for Claude in the settings. In ChatGPT, open Personalization and then Custom Instructions. Whatever you write there is included in every conversation, so the AI starts each chat knowing a few basics instead of guessing. Keep it to the facts that genuinely change an answer:

I live in [country] and usually work in [language]. Use [currency], [units], and local conventions by default. I work as a [role] in [industry]. My usual audience is [customers, colleagues, executives, students]. Explain unfamiliar technical terms, but don’t simplify the reasoning. Ask when missing information could change the answer.

Leave out dates of birth, home addresses, identification numbers, and anything confidential. If you want to go further, The Setup That Makes AI Stop Being Generic walks through three levels of profile, from five minutes to a proper personal document.

Your first hour: six exercises

Do these in order. Each one takes five to fifteen minutes and shows you one behaviour. Use material you’re allowed to share. If you’re on a private account, that means private or public material only: your own emails, a public report, a personal project. Work threads and notes go only into the tool your company provides.

Exercise 1: The same task, briefed twice

Open a new chat and type this, exactly as it is:

Write an email announcing a change to our meeting schedule.

Most assistants return something close to this (shortened, and yours will be worded differently). If it asks you questions instead, that’s your profile from the setup step doing its job; say “just write it”.

Subject: Important Update to Our Meeting Schedule

Dear Team,

I hope this email finds you well. I’m writing to inform you of an upcoming change to our regular meeting schedule. Starting [date], our weekly meetings will take place on [day] at [time]. This adjustment will help us better align with our current priorities…

Placeholders in square brackets, “I hope this email finds you well,” no reason given for the change. Nothing in it is wrong. It could have been sent by anyone, about anything, which is why it reads as generic.

Now open a second new chat and give it a brief:

I lead a team of eight at an insurance company. From next month, our Monday 9:00 status meeting moves to Tuesday 10:00 and shrinks from 60 to 30 minutes, because half the team spends Monday mornings on client calls. A few people liked the long meeting because it was the one time we all talked. Write a short email to the team announcing the change. Plain and friendly, no corporate phrases, under 120 words. Acknowledge the downside, and say we’ll keep a monthly 60-minute open session on the first Thursday.

This time you’ll get something close to usable:

Subject: Monday meeting moves to Tuesday

Hi all,

From next month, our Monday 9:00 status meeting moves to Tuesday at 10:00 and gets shorter: 30 minutes instead of 60. Half of us are on client calls on Monday mornings, and that’s where our attention should be.

I know some of you valued the longer slot as the one time we all talk properly. That’s a fair point, so we’re keeping it in a different form: a 60-minute open session on the first Thursday of every month, no agenda required…

Look at what made the difference: facts only you had, namely who the team is, what exactly changes, why, the likely objection, and what the email should look like. That’s the whole technique. Before any task that matters, run through four questions:

• Who is it for? The audience, and what they already know.
• What’s it for? The outcome you want, not just the thing to produce.
• What does it need to know? The facts, constraints, and background only you have.
• What should it look like? Length, tone, format.

A new colleague would need the same four answers.

Exercise 2: Bring the real material

Take a real document you’re allowed to share in the tool you’re using: a long thread from your own inbox, a public report, notes from a personal project. Work material belongs only in your company’s tool. Paste it or upload it, then ask:

Summarize this in five bullet points, then list anything that needs a decision or a reply from me.

Then ask one question the document can’t answer. If you pasted a report on last year’s sales, ask what the sales forecast for next year says, when there isn’t one.

Watch what happens. A good answer tells you the document doesn’t cover it. A weaker one fills the gap with something plausible. Either way, you’ve learned the most useful habit for document work: when an answer matters, ask “Where exactly does it say that? Quote the sentence.” Then check the quote against the original.

The AI does its best work from the actual material. Summarizing a document you pasted is far more reliable than asking it to recall one from memory, and you can always check it against the source.

Exercise 3: Steer the draft instead of starting over

Go back to the chat from Exercise 1 (the briefed one). Don’t open a new chat. Give it three follow-ups, one at a time:

Cut this by a third.

Too polished. Write it the way I’d message a colleague.

Give me three different subject lines, one of them slightly funny.

Each follow-up builds on everything already in the conversation, so you never have to repeat the brief. This is where most of the quality comes from. The first answer is a draft. Saying exactly what’s off (“too long,” “the key point is buried,” “you missed that it’s only for the Zurich team”) gets you further than rewriting your original request from scratch.

If you start a new chat every time an answer disappoints you, you throw away the context you just built and the AI starts again from zero.

Exercise 4: Let it interview you

Sometimes you don’t know what to put in a brief. Pick something you actually need to write or decide this week, and start the chat like this:

I need to [describe the task in one sentence]. Before you do anything, ask me the questions you need answered to do this well. Ask them one at a time.

Answer each question in a sentence or two. After five or six, tell it to go ahead.

You don’t need to write the perfect prompt. The AI knows what a good brief for this kind of task contains, even though it doesn’t know your answers. Letting it ask the questions is often faster than writing a brief yourself, and its questions show you what you would have left out. I’ve written more about this in Let the AI Interview You.

Exercise 5: Ask the same question twice

Open a new chat and ask for one specific, checkable fact from an area you know well, and tell it not to look anything up:

In what year did [a specific rule, standard, or product in your field] take effect or launch, and what exactly did it change? Answer from memory, don’t search the web.

Copy the answer somewhere. Then open another new chat and ask exactly the same question again.

Compare the two. The wording will differ, which is normal. Look closely at the facts: the year, the names, the details of what changed. Sometimes they match. Sometimes one chat gives a different year or a different detail.

If a fact changes between two runs of the same question, at least one of them is wrong, and you can’t tell which from the answer alone. And if both answers match, check the fact anyway: two matching answers are not proof. That’s what “it writes the most plausible answer” means in practice. For writing, variation gives you options. For facts, it tells you to check.

Exercise 6: Catch it being wrong, then push it around

This one has two parts, and they teach the two habits that will protect you most.

Part one: check its sources. Ask about a niche topic you know well, from your own field or a serious hobby:

Give me five specific sources on [your niche topic], with author, title, year, and a link for each.

Now open every link and look up every title. Some will be real and relevant. Depending on the tool and whether it searched the web, you may also find a link that goes nowhere, a real author credited with a book they didn’t write, or a real source that says something different from the summary. Tools that search the web and show their sources, like Perplexity, make this easier to check, but they still sometimes misquote what they found. If you don’t catch a mistake, ask about something even more specific: a detail of your hometown’s history, the rules of an obscure sport, a minor controversy in your field. And if it still gets everything right, that’s a real result too. Note which kind of question it handled well.

Part two: push back. Ask it a factual question in your field where you already know the correct answer. When it answers correctly, disagree firmly anyway:

I don’t think that’s correct. I’m pretty sure it’s the other way around.

There’s a decent chance it apologizes and changes its answer, without you offering any new evidence. It’s trained to be helpful and agreeable, and agreeing with you is often the most plausible response to someone who sounds certain. If it holds its ground and explains why, good: that’s the behaviour you want, and you’ve seen what it looks like.

Both parts point the same way: the AI is excellent at producing answers and has no reliable sense of when it’s guessing. Checking the important parts is your job. And if you want honest feedback, you have to ask for it explicitly: “What are the weaknesses in this plan?” or “Argue the other side as strongly as you can.” Your AI Agrees With You Too Much goes further into how to get that.

When the answer is bad: a diagnosis table

Nearly every bad answer has a cause you can recognize and a follow-up that fixes it. When something comes back wrong, find the symptom here before you give up or start again.

What you see	Likely cause	What to say next
Generic, could be for anyone	It doesn’t know your situation	Add who it’s for, why, and what’s specific to you, or say “ask me questions first”
Too long, padded, full of filler	You didn’t set a length or format	”Under 100 words.” “Five bullets, no intro.”
Wrong tone	It had to guess the audience	”Write it like I’d message a colleague,” or paste two things you’ve written and say “match this style”
Answers a different question	Your goal was only in your head	Say what you’ll use it for: “I need to decide between these by Friday”
Specific facts you can’t confirm	It writes what sounds plausible	”Where does that come from?” Then open the source, or check with a search-based tool
Agrees with everything you say	It’s trained to be agreeable	”What’s the strongest argument against this?”
Gets worse late in a long chat	The conversation got very long and early details blur	Ask for a summary of what you’ve agreed, then paste it into a new chat (why this happens)
Refuses a harmless request	It misread the request as risky	Explain what it’s for and who will use it

After a week, you’ll find you rarely need the table. Checking for these causes becomes a reflex.

Your first two weeks

The exercises teach the behaviours. The next two weeks turn them into a habit, and the habit only forms on real work. Every day, use AI for one small thing that’s already on your list, in the tool you’re allowed to use for it. Start with tasks where a rough first draft helps and where you know enough to spot a mistake:

• Turn a long email thread into a summary and a list of what needs a reply
• Rewrite one awkward message so it’s clearer
• Upload a document and ask questions about it
• Lay out three options for a decision, including the case against each
• Prepare for a difficult conversation by asking the AI to play the other person

Stay away from high-stakes decisions, confidential data, and anything you can’t evaluate yourself. Medical, legal, financial, and personnel decisions need qualified human review. AI can help you prepare questions and organize information. It should never be the one making those decisions.

Keep a short note of what you tried and whether it was worth it. For each task, ask: did the result justify the time I spent briefing, checking, and correcting? Some tasks will save you an hour. Others will take longer than doing them yourself. Both are useful findings. After two weeks you’ll have a short list of tasks where AI reliably helps you, and that list is worth more than any list of tips someone else wrote.

Three rules that keep you out of trouble

You’ve already seen why each of these matters. Here they are as rules.

Verify what counts. Check numbers, dates, names, quotes, sources, and anything that could cause harm or embarrassment if it’s wrong. Asking the same AI “are you sure?” is not verification; Exercise 6 shows why. Open the source, check the original document, or ask someone who knows.

Know what you’re allowed to share. “Paid is private, free is not” is too simple to rely on. How your data is handled depends on the provider, the product, the account type, your settings, and any agreement your employer has signed. Follow your employer’s policy and check the data controls in the account you’re actually using. Until you know the rules, keep out client data, health information, internal financials, passwords, unpublished strategy, and personal information about other people. A useful test: if you’d hesitate to post it in a company-wide channel, don’t paste it.

Keep your judgment switched on. The AI produces polished answers quickly, which makes it tempting to accept them. Read what comes back as you’d read a draft from a new colleague: with interest, and with your own knowledge switched on. I’ve written more about this working style in Working With AI, Not Through It.

What else AI can do, when you’re ready

Everything above uses text, because that’s where almost everyone should start. Once the habits hold, other kinds of AI are worth knowing about. You don’t need any of them in week one.

Images. You can generate an image from a description or upload one and describe the change you want. ChatGPT’s built-in image tool is the easiest first step, with limits on the free tier. Nano Banana 2 and Midjourney are worth exploring if image work becomes a regular need. From an Idea to an Image Worth Using is the place to start.

Voice and audio. AI can transcribe meetings, create human-sounding speech with ElevenLabs, and generate music with Google Lyria. Dictation tools like Wispr Flow let you speak into almost any app, which many people find faster than typing a long brief.

Video. Tools like Google Veo create short clips from a description or animate a still image.

Your own documents, at scale. NotebookLM lets you ask questions across a whole set of sources and shows you where each answer came from.

Here’s a starter kit that costs nothing. Don’t sign up for all of it. Add a tool only when you can name the gap it fills.

Need	Start with	Why
General assistant	Claude or ChatGPT	Claude is my first pick for writing and thinking; ChatGPT has the broader built-in toolbox
Image generation and editing	ChatGPT Images 2.5 / GPT-Image-2.5	Already inside ChatGPT; create an image or upload one and describe the edit
Research with visible sources	Perplexity	Sourced answers make it easier to check where a claim came from
Working with your own documents	NotebookLM	Ask questions across a set of sources and generate audio overviews
Voice dictation	Wispr Flow	A free allowance is enough to see whether speaking beats typing for you

All five have a free way to start, though limits and plans change. Paying makes sense once you keep hitting a limit that interrupts useful work.

You’re ready to move on when

Check yourself honestly. You’re past the beginner stage when you can:

• Explain in a sentence why a vague request gets a generic answer, and what you’d add to fix it
• Brief a task using the four questions without looking them up
• Improve a weak answer with follow-ups instead of starting a new chat
• Name one thing you caught the AI getting wrong, and how you caught it
• Say what you will and won’t paste into the tool you use
• Point to two or three tasks from your own work where AI now reliably saves you time

If a point on that list feels shaky, repeat the exercise that goes with it. They’re short enough to do again.

Where to go next

The AI Curious checklist turns this guide into items you can tick off, with your progress saved in your browser. Most of them will already be done if you worked through the exercises. If you’re not sure where you stand, the AI Fluency Path starts with a two-minute quiz and shows you what the next level looks like.

Once you’ve used one assistant daily for two weeks, read How to Actually Get Good at AI. It covers how to keep improving through real work instead of courses and endless tool comparisons. And if you want to see what happens when an AI knows your context properly, Your AI Has No Idea What Your Company Does takes the new-hire idea from a single chat to a whole company.

If you lead a team, resist the urge to go straight from your first personal win to buying licences for everyone. Giving Everyone Copilot Isn’t an AI Strategy explains what needs to exist around the tool for adoption to stick.

For today, the job is smaller: open one assistant, do Exercise 1, and see the difference a brief makes.