From Chatbot to Agent: What Changes When AI Can Act
Ask, assist, act — how to think about AI as a delegate, what it does well and badly, and how to find the tasks worth handing over.
A chatbot answers a question. An agent actually does a job for you, step by step. Think of it as a fast helper that is new to your work: you explain the job clearly, then check what it did. Pick tasks that repeat, have a few steps and are easy to check.
- Chatbot
- AI you talk to. It replies with text.
- Agent
- AI that can take actions and finish a task, not just reply.
- Ask, assist, act
- Three levels: it answers, it helps you use tools, or it does the work.
Good first jobs repeat, have a few steps and are easy to check.
- You ask a question
- It answers in words
- You do the work yourself afterwards
- You give it a job
- It takes steps and uses tools for you
- You check what it did
A chatbot answers a question. An agent finishes a job. That one difference changes how you should use AI. Most people have not adjusted. They still treat every AI tool like a search box that talks back. This path teaches you to treat it like a quick, capable colleague who is new to your work. You brief them well, you check what they do, and you give them work that repeats.
Three levels: ask, assist, act
- Ask: you paste something in and it answers. It knows only what the model learned plus what you pasted. Examples: explain this clause, rewrite this paragraph, think of names.
- Assist: it can use tools you connect. It can search the web, read your files, look in a spreadsheet or check your calendar. You still decide what to do with the result.
- Act: it does several steps for you. It researches a few sources, fills in a form, writes and files a document, or updates a record. You review it, approve it and fix it.
The model underneath may be the same at all three levels. What changes is what it can reach and how much it does between your instructions. Each step up saves more of your time. It also costs more in setup, checking and risk. That is why this path is about doing it on purpose, and not about switching on every feature.
Think of it as handing work to a new colleague
Picture a new colleague who is very fast, has read a lot and never gets tired. But they have never worked at your company. They do not remember yesterday. And they would sooner do something wrong with confidence than ask you a question. You would not give this person a vague one-liner and walk away. You would explain the goal and who it is for. You would show an example and say what to avoid. And you would check the work before it went to a client. That is exactly the habit that gets good results from AI.
What agents do well, and badly, today
- Well: collecting and summarizing information from many sources. First drafts of almost anything. Changing the format of content. Pulling facts out of messy text. Routine admin with several steps. Giving you options to choose from.
- Not reliable: exact sums without a calculator tool. Exact quotes and references. News that is not in its sources. Anything about your private situation that it has not been told. Long chains of steps that depend on each other, with no checkpoints.
- Not for use without supervision: actions that cannot be undone, or that cost a lot, or that are legal, medical or risky for your reputation. Keep a person in charge of those decisions.
How to spot a task worth handing over
Look for work that repeats, has several steps, has clear inputs and a recognizable good result, and where a mistake is cheap to catch. Every Monday I collect numbers from three places and write a status email is a perfect example. Decide whether to drop a supplier is not. You can even ask the AI to help you find good tasks:
I'm going to describe how I spend my working week. Act as a workflow analyst.
My role: {{role}}
What I did last week (rough list, with time spent): {{week_log}}
1. For each activity, say whether AI could Ask-level help, Assist-level help, or Act-level do most of it, and why.
2. Rank the 5 best candidates by (hours saved per month x how easy it is to check the result).
3. For the top 2, list the exact inputs I'd need to provide and what a "good result" looks like.
4. Tell me which activities I should NOT hand to AI and why.
Be skeptical: if something is unlikely to save real time, say so.Be honest about the payoff
You will see big claims about hours saved each week. Treat them as advertising, not as a forecast. Real gains come from specific tasks that repeat, where you have built a good prompt or routine. They also come from skipping the blank page on a first draft. The gains vanish when you spend an hour fixing a careless result. The skills that decide which side you end up on are the ones this path teaches: briefing well, organizing the work and checking the results. Measure the time on your own tasks, with and without AI.
The five skills in this path
- Structured prompting (lesson 2): get a usable, consistent result the first time.
- Lasting context (lesson 3): stop explaining who you are and what you need in every chat.
- Connecting AI to your tools (lesson 4): browser agents and connectors, used safely.
- Reusable routines (lesson 5): turn good prompts into your own playbook and chain them together.
- Checking and safe use (lesson 6): catch mistakes and protect private data.
- ✓A chatbot answers. An agent finishes a job in several steps. Think of ask, then assist, then act.
- ✓Treat AI like a fast, new colleague with no memory. Brief it well and check its work.
- ✓The best tasks to hand over repeat, have several steps, are clearly described and are cheap to check.
- ✓Be doubtful about big time-saving claims. Measure your own repeated tasks.
Look at your own week
- List everything you did last week, with rough minutes. Aim for 15 to 25 items.
- Run the task-audit prompt above with your list and your role.
- Pick the best candidate. Write down how you would know a result is good enough to use.
- Try it once this week, with you in charge. Note the time it took with and without AI, including the time to check.
Q1What is the practical difference between assist and act?show answer
At the assist level, the AI uses tools to fetch or study something, and you decide what to do. At the act level, it does several steps itself, and you review and approve the outcome.
Q2Why is a new colleague with no memory a useful picture?show answer
It reminds you to give the goal, the reader, an example and the limits every time, and to check the result before it goes anywhere important.
Q3Name two things that make a good first task to hand over.show answer
It repeats, with clear inputs and a result you can recognize as good. And a mistake is cheap to catch, because the stakes are low and it is easy to check.