Kevin Livery
Kevin Livery

NexAI

Series

Learn

Lessons on using AI agents, from your first chat to running the work of a business.

Search matches level names, lesson texts and resources.

5 levels of AI agent usage

This series is a path from your first chat with an AI helper to running the work of a business with AI agents. An AI agent is an AI that can use tools and take steps on its own toward a goal you set. Each level adds one new idea and ends with a short exercise you can do today. The order is Kevin Livery's way of arranging the field, built on published guides from AI labs, tool makers, researchers and standards bodies. Start at the level that fits you.

Which level am I?

Pick the line that sounds most like you. It takes you to your level.

  1. I have never used an AI chat helper, or only a few times.Level 1
  2. I chat with AI, but the answers often miss what I meant.Level 2
  3. I write clear requests, but I still copy and paste between the AI and my other apps.Level 3
  4. My AI can use tools, and I want it to finish whole tasks while I do other work.Level 4
  5. I run or build a business and want to decide which work agents do and which work people keep.Level 5

Level 1Talk and check

This level is for you when you have never used an AI helper, or only a few times.

The first idea

An AI helper answers in words. It is not a person, and it can be wrong, so you check what it says.

What you can do

  • Ask an AI chat helper a question and read the answer.
  • Ask it to explain in a simpler way.
  • Check the answer in a book, on a trusted website, or with a grown-up.

How it works

An AI chat helper is a computer program you can talk to by typing. It has learned from a huge amount of writing. When you ask it something, it guesses the next word, then the next, until it has an answer.

It is a bit like the word guesses on a phone keyboard, only much bigger. Most of its guesses are good. Some are wrong. It can sound very sure even when it is wrong. That is why you check.

It is a program, not a person. It has no feelings, and it is not your friend. Keep your address, school name, passwords and secrets to yourself.

Some AI apps are only for teens or adults. Others need a parent to turn them on. If you are a kid, use one with a grown-up who says yes.

Example

Mia wants to know how bees make honey. An AI helper gives her 5 short steps. One step has a word she does not know, so she asks: "Can you say that in a simpler way?"

Then Mia looks up one step in her library book. The book agrees, so she trusts that step. If it had said something else, she would ask her teacher.

Try this today

With a grown-up, take 15 minutes.

  1. Ask an AI helper about something you already know well, like your favorite animal.
  2. Read the answer slowly. Find one thing that is right and one thing you are not sure about.
  3. Check the thing you are not sure about in a book or on a trusted website.
  4. Tell your grown-up what you found.

Common mistakes

  • Believing an answer just because it sounds sure.
  • Typing your name, address, school or passwords.
  • Treating the helper like a friend who has feelings.

You are ready for the next level when

  • You check answers before you trust them.
  • You can ask the helper to try again in a new way.
  • You know what never to type into it.

Resources

Level 2Brief with context

This level is for you when you chat with AI now and then, but the answers often feel generic or miss what you meant.

The new idea

The answer gets better when you tell the AI your goal, the background it cannot know, and the shape you want the answer in.

What you can do

  • Write a request that gives the goal, the background and the format you want.
  • Show an example of what a good result looks like.
  • Improve a weak answer with follow-up messages instead of starting over.
  • Spot a vague answer and ask for something specific.

How it works

The AI only knows what it learned before and what you put in your message. It does not know your job, your customers or your taste unless you tell it. Anthropic's prompting guide compares the model to a brilliant new employee who does not yet know how your team works: the more precisely you explain, the better the result.

Google's guide for Gemini names 4 parts of a good request: a persona (the role the AI should play), the task, the context and the format. It also says to treat the work as a conversation and use follow-up messages to improve the result.

Think of ordering a cake. "A cake, please" gets you some cake. "A chocolate cake for 10 people that says Happy Birthday Sam, ready on Saturday" gets you the cake you wanted.

A clear request makes a useful answer more likely. It does not make the answer true. Google's guide says to review the output for accuracy before you use it, and the US standards body NIST warns that AI can state wrong things with confidence.

Example

A bakery owner first asks: "Write a post about our bakery." The answer is friendly but could be about any bakery.

So the owner asks again: "You write for a small neighborhood bakery. Write a 3-sentence Instagram caption for our new sourdough loaf. It comes out of the oven at 7 every morning. Our customers are busy parents who like short, warm messages. No hashtags. Give me 3 versions."

The second request gives the role, the task, the facts, the reader, the length and the format. The owner picks the best of the 3 versions and changes one word.

Try this today

Take 20 minutes.

  1. Pick a real task from your week, such as an email, a plan or a summary.
  2. Ask for it in one short sentence and save the answer.
  3. Ask again with the goal, who it is for, the facts it needs, the length and the format. Add one example if you have one.
  4. Compare the 2 answers and write down what changed.

Common mistakes

  • Asking in one vague line, then blaming the AI for a vague answer.
  • Leaving out facts the AI cannot know, such as prices, names or dates, so it fills the gaps with guesses.
  • Pasting private customer data or passwords into a chat without checking the tool's rules.

You are ready for the next level when

  • Your first request usually gets a useful answer.
  • You give the goal, the context and the format without having to think about it.
  • You can fix a weak answer with 1 or 2 follow-up messages.

Resources

Level 3Connect your tools

This level is for you when you write good requests but keep copying and pasting between the AI and your other apps.

The new idea

Once an AI is connected to tools and apps, it can look things up and take actions for you instead of only writing text.

What you can do

  • Let the AI search the web or your files so its answers use current facts and show their sources.
  • Connect it to everyday apps such as your calendar, email or project tracker, so it can read from them and, when you allow it, act in them.
  • Turn on only the connections a task needs.
  • Check where an answer came from before relying on it.

How it works

On its own, a chat model only has what it learned in training and what you type. Anthropic calls a model with search, tools and memory added an augmented model: it can decide to search, read a file or call an app, then use what it found in its answer.

Connectors are the plugs between an AI and your apps. The Model Context Protocol (MCP) is an open standard for these plugs, and its documentation compares it to a USB-C port, one shape that fits many devices. Anthropic connects Claude to apps such as Jira, Asana and Zapier this way, and OpenAI lets ChatGPT use connectors to apps such as Gmail and GitHub.

Connecting tools is like giving a house sitter keys: you hand over the keys to the rooms they need, not every key you own.

Every connection is also a door. The security group OWASP calls it excessive agency when an AI has more tools, permissions or freedom than a task needs, because a wrong or manipulated answer can then cause real damage. Hidden instructions in a web page or file can try to trick an AI into sharing data or doing harm, which is called prompt injection. Give each connection only the access its task needs, and, as OpenAI advises, turn off connectors a task does not use.

Example

A freelance designer asks: "Which of my clients have not paid this month? Draft a polite reminder for each." The AI is connected only to the invoicing app, with read access only.

It finds 3 unpaid invoices and drafts 3 emails. The designer reads each draft, corrects one amount the AI got wrong and sends the emails personally.

Try this today

Take 25 minutes.

  1. Open the AI tool you already use and find its settings for web search, files or connected apps.
  2. Turn on just one: web search, or one app you use daily.
  3. Ask a question that needs that tool, such as: "What changed in my field this month? Give the source for each point."
  4. Open 2 of the sources and check that they say what the AI claims.
  5. Turn off anything you do not need.

Common mistakes

  • Connecting every app at once with full access.
  • Trusting an answer with sources without opening the sources.
  • Letting the AI read untrusted pages or files while it can also reach your private data.

You are ready for the next level when

  • You know which tools your AI can use and what each one can see.
  • You open the sources before using an answer.
  • Each connection has the least access its task needs.

Resources

Level 4Hand over a task

This level is for you when your AI can use tools and you want it to finish whole jobs on its own.

The new idea

You give an agent a goal instead of instructions, and it plans, acts and checks in a loop until the job is done, pausing to ask you before risky actions.

What you can do

  • Describe a task by its outcome: the goal, the limits and what done looks like.
  • Let an agent plan the steps, use tools and correct itself along the way.
  • Set approval points before it pays, sends, posts or deletes anything.

How it works

Developer Simon Willison's one-line definition: an LLM (large language model) agent runs tools in a loop to achieve a goal. Anthropic describes that loop as planning, acting, checking the result of each step and pausing for human feedback.

Anthropic separates agents from workflows: a workflow follows steps fixed in advance, while an agent chooses its own steps, which is more flexible but costs more, and its mistakes can compound. Anthropic and OpenAI both recommend starting with the simplest setup that works.

Browser agents such as ChatGPT agent from OpenAI and Claude in Chrome from Anthropic can open pages, click and fill in forms. Both makers say their agent asks before high-risk actions such as purchases, and both warn that hidden instructions on web pages can try to trick it into harmful actions.

Delegating to an agent is like delegating to a new team member: explain the goal, show where things are, name the decisions that need your approval and check the result.

In a March 2025 study, the research group METR found that agents completed short tasks far more often than tasks taking a skilled person several hours, and they seemed to struggle more with chaining steps than with any single step. OpenAI says its own agent can still make mistakes, so start with tasks that are short, clear and easy to verify.

Example

A shop owner briefs an agent: "Find 5 suppliers of recycled paper bags. For each, collect the minimum order, the price per 1,000 bags and the delivery time from their own website, in a table with links. Do not contact anyone or sign up for anything."

The agent searches the supplier sites and builds the table, stopping once to ask whether a price in another currency is fine. The owner checks 2 rows against the websites, finds one outdated price and asks the agent to recheck it.

Try this today

Take 30 minutes.

  1. Pick a research task with a clear finish, such as comparing 5 options for a purchase.
  2. Write the brief: the goal, where to look, what to collect, the format and what it must not do (no purchases, sign-ups or messages).
  3. Run it in the agent or research mode of a tool you already use, and watch its first steps.
  4. Check at least 2 results against their sources and note what it got wrong.

Common mistakes

  • Giving instructions instead of a goal, or a goal without limits.
  • Letting an agent pay, send, post or delete without your approval.
  • Delegating long, ambiguous tasks before short, clear ones go well.

You are ready for the next level when

  • Your briefs state the goal, the limits and what done looks like.
  • You set an approval point before every action that cannot be undone.
  • You verify finished work against sources, not against its confident tone.

Resources

Level 5Run work with agents

This level is for you when you run or build the work of a business and want to decide, task by task, what agents do and what people keep.

The new idea

You stop handing agents single tasks and start designing the work itself: every task is treated as potential agent work, while people set goals, supply context, check results and own outcomes.

What you can do

  • Map a business's work and choose where agents fit first.
  • Write agent briefs, controls and checks that a whole team can reuse.
  • Keep a named person accountable for every outcome.

How it works

Microsoft's 2025 Work Trend Index describes a progression from AI as an assistant, to agents as digital colleagues handling tasks at human direction, to people directing agents that run entire business processes. It calls the person in charge an agent boss, someone who builds, delegates to and manages agents. This series turns that idea into a method.

AWS and a research paper published by the Knight First Amendment Institute treat autonomy as a design choice, not a fixed property of the tool.

Think of a head chef, who plans the menu, assigns stations, tastes every plate before it leaves the kitchen and answers for the result, while the cooks do most of the cooking.

The method

  1. List the work

    List every kind of work the business does, such as sales, marketing, delivery, support, finance, hiring and operations, and the recurring tasks under each, like answering pricing questions, issuing invoices or preparing weekly reports.

  2. Sort each task

    Ask 4 questions per task: is the goal clear, can the result be verified, what happens if it fails, and does it depend on human judgment, empathy or relationships? OpenAI's guide points agents at complex judgment, hard-to-maintain rules and unstructured information, noting that simple deterministic automation may suffice elsewhere.

  3. Choose a setting

    Assign each task one of 3 settings: the agent suggests and a person acts, the agent acts after a person approves, or the agent acts independently and a person audits the results. Start low and promote a task only when its record justifies it, as AWS also advises.

  4. Brief the agent

    Each brief covers 5 elements: the goal and a definition of done; the context, such as price lists, tone of voice and past examples; the permitted tools; the limits, such as a refund ceiling or no customer emails without approval; and the checks the work must pass. OpenAI recommends building agent instructions from existing procedures and support scripts.

  5. Check the work

    Decide before the agent starts how its output will be verified: everything at first, then a regular sample. Judge the result, not the effort. Anthropic reports that human testers of its research agents caught fabricated answers and weak sources that automated evaluations missed.

  6. Keep people responsible

    Name one owner for every outcome; an agent can do the work but cannot carry the accountability. The AI Fluency framework by Rick Dakan, Joseph Feller and Anthropic calls this diligence: taking responsibility for what we do with AI.

  7. Start small and grow

    Begin with one process, one agent and a few tools, adding agents only when one is clearly insufficient, as OpenAI and Anthropic both advise. In Anthropic's June 2025 data, agents used about 4 times the tokens of a chat and multi-agent systems about 15 times (tokens are the units AI processing is measured in).

Risks and safeguards

RiskConfident but wrong output, which NIST calls confabulation.
SafeguardVerify facts against sources and require human review of anything leaving the business.
RiskExposure of private data.
SafeguardGrant each agent only the data and permissions its task requires, and follow each tool's data rules.
RiskIrreversible actions such as payments, deletions and customer messages.
SafeguardRequire human approval, as OpenAI's guide does for high-risk actions until confidence in the agent grows.
RiskPrompt injection, meaning hidden instructions in web pages, emails or files.
SafeguardLimit what each agent can reach, disable unneeded connectors and keep approvals on risky actions.
RiskCost.
SafeguardCap the steps and the budget, such as a maximum number of iterations, and confirm each task is worth its cost.

What agents still do badly

Agents still perform some work poorly. They appear to struggle with long chains of steps (METR, March 2025), were not yet good at coordinating with each other in real time (Anthropic, June 2025), and can fabricate facts, choose weak sources or follow hidden instructions (NIST, Anthropic, OWASP). Design the system around these limits.

Example

A small design agency maps its work. In sales, an agent drafts replies to inquiries from the price list, and a person approves each one. In support, an agent answers project-status questions from the tracker alone, and a person reviews a weekly sample.

In finance, an agent prepares invoices and a person approves every payment; hiring stays with people because it depends on judgment and trust. Each owner reads the agent log weekly and adjusts settings from the record.

Try this today

Take 30 minutes.

  1. List 10 recurring tasks from one area of your business, such as sales or support.
  2. Give each task a setting and an owner.
  3. Choose the task with the clearest goal and the easiest check, and write its 5-part brief.
  4. Write down how you will verify its first 10 results.

Common mistakes

  • Automating a task that nobody owns.
  • Granting broad data access and permissions for convenience.
  • Adding agents before one agent works reliably on one process.

You are doing this level well when

  • Every agent task has a written brief, a setting and a named owner.
  • You can show how each type of agent output is verified.
  • You promote or demote tasks based on their record, not on hope.

Resources