Chat versus a process

If you have used ChatGPT, Claude, or Gemini, you know the rhythm: you type, it answers, you type again. Somewhere in the last year or so a second kind of button showed up in those same apps. It has names like Deep Research, Work, agent, or auto browse. You tap it, and instead of an answer you get a progress log that runs for ten minutes. This module is about that second button. This first lesson is about what is actually different under the hood, because once you see it, the rest of the module is just details.

What a chat really is

Back in AI Basics you learned that a model predicts the next word, over and over, until it has produced something that looks like a good response to your exact words. A chat reply is one pass of that prediction. The model reads everything in the conversation so far and writes one answer, start to finish, in a few seconds.

Here is the part people miss: during that one pass, the model does not stop to look anything up, does not re-read what it just wrote, and does not go back to fix a mistake it made in sentence two. It writes the way you would answer a question from a friend across the table. From memory, fluently, and with whatever confidence comes naturally. Web search, when it is turned on, adds one quick lookup before the answer. It does not change the one-pass shape.

That is fine for most things. Rewriting an email, brainstorming names for a food truck, explaining what an HOA lien is. The information is either in your message or it is common knowledge, and one pass gets you there.

What a process is

A process, in this module, means the AI does not answer once. It works in a loop: it makes a plan, takes one action, looks at the result, decides what to do next, and repeats until the job is done. Each trip around the loop is small. Run one search. Open one page. Compare two numbers. Write one section. The power comes from stacking dozens of those trips.

The three parts of every loop are worth naming, because you will see them in every feature this module covers:

• Plan. Break the request into smaller questions or tasks, and pick the first one.
• Act. Do one thing that produces new information: a search, opening a page, running a calculation, clicking a button.
• Check. Look at what came back. Did it answer the sub-question? Does it contradict something found earlier? What is still missing? Then go back to Plan.

The word you will hear for this is agentic, which just means the AI is choosing its own next step instead of waiting for you to type one. Nothing mystical. It is a loop with a decision in it.

Who runs the loop

The model itself cannot loop. Remember lesson 3 of AI Basics: the model is a brain in a jar, text in and text out, and everything else comes from the harness, the app wrapped around it. The harness is what makes a process possible. It takes the model's plan, actually runs the search, actually fetches the page, hands the result back to the model, and asks "what next?" Then it does that again. The model decides; the harness does.

This is why the same company can offer both a chat and a process using the same model. Deep Research in ChatGPT and a normal ChatGPT conversation may be powered by similar brains. The difference is the harness around them: one is built to answer once, the other is built to keep going until a checklist is done. It is also why these features cost more and come with monthly limits, which lesson 6 gets into. Every trip around the loop is another call to the model.

You can see the harness at work in the progress log. Lines like "searching for," "reading," or "opening" are the harness reporting the action it just took on the model's behalf. Lines that read like thinking, "these two sources disagree, checking a third," are the model deciding what comes next. Once you can tell the two apart, the log stops looking like noise and starts looking like a coworker narrating their work.

Why it takes minutes, and what the wait is buying you

People new to these features get nervous during the wait. Nothing is broken. A research run that takes twelve minutes is usually doing somewhere between a few dozen and a couple hundred small actions, each one a separate call to the model plus a real web request. Search, wait for results, read them, decide, search again. Ten minutes of that is not slow. Ten minutes for a human doing the same clicking would be a good afternoon.

What the wait buys you is three things a one-pass chat cannot give:

• Evidence instead of memory. Facts in the result were read from a page during the run, not recalled from training. They can still be wrong, but they can also be checked, because the page is cited.
• Self-correction. The Check step catches contradictions between sources and goes looking for a tiebreaker. A chat has no Check step at all.
• Coverage. A loop can visit twenty sources. A chat reply is limited to what fits in one pass.

The trade is time, cost, and a new kind of review job for you. A long result with citations feels finished in a way a chat reply does not, and that feeling is dangerous. Lesson 5 is entirely about that.

How this changes what you type

A chat wants a question. A process wants a brief: the goal, what a good result looks like, what to include and exclude, and how you want it delivered. The same four ingredients from AI Basics lesson 5 (goal, context, format, example) apply, but they matter more here, because a vague brief sends the loop off in a vague direction for ten minutes and you get a long, confident answer to a question you did not quite ask.

One more mindset shift. In a chat, you correct as you go. In a process, you correct at the start, because you are not in the loop while it runs. Most research features will ask you two or three clarifying questions before they begin. Answer them carefully. That thirty seconds is the highest-leverage moment in the whole run.

Try this yourself

Pick one question you would normally just ask in chat and run it both ways. First as a plain chat message, then using whichever research feature your app has (lesson 2 covers where to find them). Use a brief like this for the process version:

I run a small gutter company in Washington County, Utah. Research which three local business bank accounts have the lowest monthly fees for a business with under $20,000 average balance.

Include: the bank, the account name, the monthly fee and how to waive it, and any transaction limits.
Exclude: online-only banks with no branch within 30 miles of St. George.
Deliver: a short table, then one paragraph on which you would pick and why. Cite the page you got each fee from.

Then compare. Notice which one gave you numbers you could click on, which one made you feel more confident, and whether those were the same answer. That gap is what the rest of this module is about.