Most people stay at the chat level of AI: four steps to move beyond it
What the chat level is, why people stop there, and what to build at each of the four steps after it.
Most people who use AI at work use it one way. They type a question, read the answer and start again from zero the next day. This is the chat level. It is useful, but nothing you do there is saved. This guide shows what the surveys measured, why people stop at chat, and four steps past it. At each step you make one thing that you keep.
How many people use AI at work, and how often
- 38% of employed U.S. adults use chatbots for tasks at work. Pew Research Center asked 5,119 U.S. adults in February 2026.
- 30% of U.S. employees use AI in their role a few times a week or more. This is Gallup data from May 2026.
- 15% of U.S. employees use AI in their role daily. This is from the same Gallup data.
These surveys measure how often people use AI. No survey that we found measures how people use it: chat only, or more. So "most people stay at the chat level" is our conclusion, not a measured figure. The reasoning is short. When fewer than one in three employees use AI each week, the number who have built skills, agents or workflows is smaller again.
The Gallup figures are on the AI adoption statistics page. The Pew source is in the list at the end of this post.
What the chat level is
At the chat level, you type a request, read the answer and start from zero the next time. The goal, the context and your rules are in your head. The AI gets only the words in the box.
Write a follow-up email to a client about the late report.
Chat is not wrong. It suits a question that you ask one time. It is the wrong tool for work that you repeat, because you pay the setup cost each time and the results change from day to day.
Why people stop at chat
These three reasons are our reading of the studies. The studies did not ask people why they stop.
- Chat needs no setup. You can start in ten seconds. Each later step asks you to write something down first.
- The results vary. In one controlled study, consultants who used AI on a task outside its abilities were 19 percentage points less likely to get the correct answer. One bad result is enough to make a person trust AI only for small tasks. See the AI reliability statistics.
- The average saving is small. A St. Louis Fed survey found that generative AI users saved 5.4% of their work hours. A saving of that size does not push a person to change how they work. See the AI productivity statistics.
The four steps answer each reason. You do the setup one time and keep it. You write the checks down, so a bad result is caught. The saving grows because the work repeats without you.
The four steps past chat
| Level | What you do | What you keep |
|---|---|---|
| Chat | You type a request and read the answer. | Nothing |
| Step 1: Prompts | You state the goal, the context and the "done when" checks. | A prompt that works the first time |
| Step 2: Skills | You save the instructions for a task you repeat. | A skill the AI uses each time |
| Step 3: Agents | You give an AI a goal, tools and a limit. It decides the steps. | A job done from start to end |
| Step 4: Agentic workflows | You put agents on a fixed route with checks. | A repeated job that needs only your review |
Step 1: Write prompts that state the goal
The newest AI models work best when you describe the result, not each step. The prompting guides from OpenAI, Anthropic and Google agree on this point. A goal-first prompt gives the goal with its reason, the context, the "done when" checks, the limits and the output format.
This is the chat request from above, written as a goal-first prompt:
Goal: Write a follow-up email that tells my client the report is five days late and gives the new date. The client is annoyed, so the email must be direct and take responsibility. Context: The report was due November 6. The new date is November 13. The cause was late data from the client's own finance team, but I do not want to blame them. Done when: - The new date is in the first two sentences. - The email is under 120 words. - It contains one apology and no excuses. Limits: Promise nothing except the new date. If you need a fact that I did not give, ask me. Output: The email only, with a subject line.
What you make this week: one goal-first prompt for a task you do often. The AI Prompt Builder asks for each part and writes the prompt.
Step 2: Save your instructions as a skill
When you have pasted the same prompt three times, save it. An agent skill is a saved file of instructions for one kind of task. The AI sees the short description of each skill. When your request matches, it loads the full instructions and follows them.
A skill is different from a saved prompt in three ways. It tells the AI when to use it. It tells the AI what to ask you first. And it contains an example of a good result.
--- name: client-follow-up description: Writes a follow-up email to a client about a late, changed or completed deliverable. Use when I ask for a client follow-up or a status email. --- # Client follow-up email ## Before you write Ask me for any of these that I did not give: the client, the deliverable, the old date, the new date and the cause. ## Rules - Put the new date in the first two sentences. - Keep the email under 120 words. - Use one apology and no excuses. Do not blame the client. - Promise nothing except what I gave you. ## Output A subject line and the email. Under the email, list each fact you were not sure of. ## Example of a good email [Paste one email that you sent and were satisfied with.]
The file is always called SKILL.md. In Claude Code, save it at ~/.claude/skills/client-follow-up/SKILL.md. For the Claude apps and Cowork, upload the skill in the skills settings on claude.ai. Then you write only "follow up with Dana about the late report", and the rules apply each time.
What you make this week: one skill for the prompt you paste most.
Step 3: Give one job to an agent
An AI agent gets a goal and tools, and it decides the steps. In chat, you do the work between the answers: you open the files, paste the text and save the result. An agent does those steps.
Two rules make a first agent job safe. Give it a limit: which folders it can read and what it must not do. And check the result yourself. Agents are capable but not reliable enough to work without review. On the OSWorld test of tasks in real computer programs, agents completed 66.3% of the tasks in 2025.
Goal: Prepare the Friday status emails for my five active clients. Do not send them. For each client folder in clients/: 1. Read the project notes and the hours file for this week. 2. Decide which of these applies: on schedule, late, or blocked by the client. 3. Write the email with the client-follow-up skill. 4. Save it in outputs/ with the client and the date in the file title. Limits: Read only the clients/ folder. Send nothing. Change no source file. Done when: outputs/ contains one email for each client, and you gave me a list of each fact you were not sure of.
This job uses the skill from step 2. The agent decides which case applies to each client, and the skill sets how the email is written.
What you make this week: one agent job with a goal, a limit and a "done when" line. Read AI agent vs. chatbot vs. workflow to choose which jobs need an agent.
Step 4: Connect the steps into an agentic workflow
A plain workflow repeats fixed steps and decides nothing. An agentic workflow puts agents on a fixed route. You set the stages and the checks. The agent makes the judgment calls inside each stage.
A first agentic workflow needs four things:
- A start. A day of the week, or the arrival of a file.
- Written steps. The stages in order, with the skill each stage uses.
- A check. A rule or a person reviews the result before it goes to anyone. This is the human-in-the-loop rule.
- A place for the files. One folder with the rules, the inputs, the results and a log.
The Friday status emails from step 3 become a workflow when the steps are in a file, the results go in a dated folder and you review them each Friday before you send them.
What you make this week: one workflow folder. The AI Workflow Folder Builder writes the folder plan and each starter file from your answers. For work that improves a result in rounds, see 20 loop prompts for a solo business.
How to find your level and your next step
Read the five questions in order. Stop at the first one that describes most of your AI use.
- Do you type a new request from zero each time? You are at the chat level. Do step 1 this week: rewrite one request with a goal and a "done when" line.
- Do you have prompts that you paste again and again? You are at step 1. Do step 2: save the prompt you paste most as a skill.
- Do you have saved skills, but you start each task yourself? You are at step 2. Do step 3: give one job with a clear limit to an agent.
- Does an agent do full jobs for you, one at a time? You are at step 3. Do step 4: put one repeated job in a workflow folder with a check step.
- Does a workflow do a repeated job while you check only the result? You are at step 4. Measure the time it saves, then build the next one.
Do not skip a step. An agent with a vague goal does the wrong job faster. A workflow with no checks repeats its errors. To put a number on your progress, use the AI Time Savings Calculator.
Common questions
What is the chat level of AI use?
The chat level is AI use where you type a request, read the answer and start from zero the next time. Nothing is saved. The goal, the context and the rules are in your head, not in a file the AI can use again.
How many workers use AI at work?
In a Pew Research Center survey from February 2026, 38% of employed U.S. adults said they use chatbots for tasks at work. In Gallup data from May 2026, 30% of U.S. employees used AI in their role a few times a week or more, and 15% used it daily.
What is the difference between a prompt, a skill, an agent and an agentic workflow?
A prompt is one written request. A skill is a saved set of instructions that the AI uses each time a task comes up. An agent gets a goal and tools and decides the steps. An agentic workflow puts agents on a fixed route with checks, so a repeated job is done the same way each time.
Do I need to code to move past the chat level?
No. A prompt, a skill and a workflow folder are text files in plain language. An agent tool does the technical steps. Your work is to state the goal, the limits and the checks.
Which step do I start with?
Start with step 1, the goal-first prompt. Each later step uses it: a skill is a saved prompt with rules, an agent needs a goal and limits, and a workflow is a set of steps with checks.
Sources
- Pew Research Center: Americans and AI 2026 (survey of 5,119 U.S. adults, February 2026)
- Gallup: Indicator: Artificial Intelligence (May 2026 data)
- AI productivity statistics, with the St. Louis Fed source
- AI reliability statistics, with the Dell'Acqua study and the Stanford AI Index sources
- AI Prompt Builder, with the prompting guides from OpenAI, Anthropic and Google
- Claude Code documentation: skills, with the file locations