Set up your first AI workflow

What is an AI workflow, and how is it different from a chat?

A chat is one question, one answer, and you are the memory. You bring the context, you paste the data, you judge the output, and when you close the tab it is gone.

A workflow is the same job written down once, so it runs the same way every time. Same context, same input shape, same prompt, same output shape. You still press the button and you still check the result. What you stop doing is re-explaining your business every Monday.

An agent sits further along the same line. It picks which steps to take, in what order, and keeps going without you in between. That is a different level of trust and a different set of failure modes, so I would not start there. When you want that shape, Build an agent that does real work is the next one to read. If you are earlier than this and still choosing a tool, start at Where to start with AI.

The line I use: a chat answers, a workflow repeats, an agent decides.

Which task should be your first AI workflow?

Three criteria. It repeats, so the setup cost gets paid back. It has a clear input and a checkable output, so you can tell whether it worked. And it is annoying but not dangerous to get wrong, so week one costs you a correction instead of a client.

Most people fail the second one. They pick something with no checkable output, like help me decide what to focus on, and then there is no way to tell whether the thing is working. You find out in ninety days, by which point you have changed four other things too.

The pattern in that table: the good first picks are the ones where a mistake is visible and cheap.

Why does the impressive AI answer not help?

Run the lazy version first. Export last week's deals, paste them into any assistant, and ask what it thinks.

You get something like this. Pipeline health remains broadly stable, with continued momentum in the mid market segment. Consider prioritising high intent leads and tightening follow up cadence.

Fine. What do we do tomorrow morning? Which deal, which person, which message. The answer is confident and grammatical and contains nothing that came out of your file. It would read the same if you had pasted a different company's numbers, which is the tell.

Ask for a summary of a genuinely bad week and you can still get the word robust back. The file will be sitting right there, saying two deals died.

How do you give a model your actual context?

Context is the step almost everyone skips, and it is most of the difference. With no context a model can only give you the average answer for your industry, because the average is all it has. It will sound sure of itself, because that is how it sounds.

Four things are worth writing down once.

• A one page brief of your business. What you sell, to whom, in what price range, what a good customer looks like, what you do not do. Plain sentences, no positioning language.
• Two or three examples of good output. The actual summary you would have written by hand. This does more work than any adjective in a prompt.
• The vocabulary your customers use. Their words, not yours. If buyers say service charge and your website says annual maintenance fee, the output should say service charge.
• The constraints. What must never appear: prices you have not confirmed, promises about timelines, names nobody has verified.

Then put it somewhere the model can reach without you pasting it every time. Anthropic's help pages describe projects as a place to upload files and set project instructions, so every chat inside that project starts from that background. The other major assistants have their own version of the same idea. The mechanism matters less than the habit: the brief lives in a file, not in your head.

Fact and reading, kept apart. The vendor documentation says you can store files and instructions per project. My reading is that most complaints about the model not being good enough are really the model having never seen the business.

How do you connect the tools you already pay for?

There is a ladder here. The honest version is that rung one is fine, and plenty of useful workflows never leave it.

1. 1Paste and check. You export the file, paste it in under your brief, read the output. No setup, works today. This is where I would start, and I would stay here for at least three runs.
2. 2Built in connectors. The major assistants can read from apps you already use once you connect them. Anthropic's documentation describes connectors as letting Claude reach your apps, retrieve your data and take actions in those services. Google's help page for the Workspace app in Gemini says you pull in Gmail, Docs or Drive by typing @ and naming the service. Most of it now runs on MCP, an open standard for connecting assistants to outside systems.
3. 3A no code automation platform. Zapier's documentation lays out the shape: a Zap is a trigger (the event that starts it) plus one or more actions, and you map fields from earlier steps into later ones so real data flows through. You test each step before publishing. This rung earns its place when the input arrives on a schedule or from a form and you are tired of exporting by hand.
4. 4Scripted. You call the API yourself, or push data to a webhook. Zapier supports catching a webhook as a trigger, which is the usual bridge between something custom and something no code. Worth it when the shape is stable and you are running it daily.

How to set up an AI workflow in one afternoon

1. 1Pick one task from your own version of the table above. One task, not a system.
2. 2Do it by hand once, slowly, and write down what you looked at and what you decided. That is your prompt in rough form.
3. 3Write the one page context brief and save it as a file.
4. 4Write the prompt with placeholders for the parts that change each week.
5. 5Run it on last week's real data, not a sample. Compare the output to what you actually did.
6. 6Fix the prompt where the output was wrong, and add whatever you had to correct into the brief.
7. 7Run it three more times before you automate anything.
8. 8Only then decide whether a connector or an automation platform saves enough steps to be worth maintaining.

Step seven is the one people skip, and step seven is where you find out whether the thing works.

What does the check step look like?

Every workflow needs an output you can inspect in ten seconds. Not a summary you have to trust. A table you can audit.

So reject the paragraph and ask for rows. One row per item, with the source attached to the row. If the model says a deal stalled, that row carries the line from the file that says so. If it cannot attach the source, the workflow has just told you it invented the row.

Google's own help page for the Gemini Workspace app is useful here, because it lists what the connection cannot see (comments, images inside documents) and tells you to check the sources on what you get back. A vendor telling you to verify its output is a vendor telling you to build a check step. I would take the hint.

A worked example you can copy

Say you run a small services business, and every Monday you open the CRM export and write a short update for yourself. Here is the prompt I would save and reuse. The bracketed parts are the ones you replace each week.

You are helping me run my Monday review.

CONTEXT
[paste your one page business brief here]

INPUT
Below is my deal export for the week of [date].
The columns are: [list your actual column names].

[paste the export]

WHAT TO PRODUCE
A table with one row per deal that changed this week. Columns:
1. Deal name
2. What changed, in one short sentence
3. The exact line from the export that shows it
4. The single next action, naming the person and the channel
5. Confidence: high if column 3 is copied text, low if you inferred it

RULES
- Column 3 must be text copied from the input. Never paraphrase there.
- If you cannot find evidence that a deal changed, leave it out.
- No summary paragraph and no recommendations section. The table is the output.
- If fewer than three deals changed, say so and return the short table.

The check takes a minute. Pick three rows, open the file, confirm the quoted line exists and says what the row claims. If all three hold, trust the rest for this week. If one fails, the prompt needs work, not the model.

How does this break?

It breaks in boring, predictable ways, which is good news, because you can watch for all of them.

• The input format changes. Someone adds a column, renames a field, or exports dates differently. The prompt still runs, the table still looks confident, and it is quietly reading the wrong column. Name your columns in the prompt, and check the row count against the file.
• The model changes underneath you. Assistants get updated, and a prompt tuned against one version can drift on the next. Nobody sends you a note about your particular use. Keep an old input and re-run it now and then to see whether you get roughly the same output.
• It is subtly wrong for three weeks. This is the real one. The output looks right, you stop checking after week two, and in week five you notice that deals with no activity have been dropped the whole time. The check step is a weekly habit, not a launch ritual. That is why the output has to be a table.
• The connection stops mattering without telling you. Zapier's documentation notes that a published Zap only processes data created after it was published, and that you test each step before turning it on. Automations rarely announce that they have gone quiet. Put the run count somewhere you will see it.
• You automated something you should have deleted. This happens more than people admit. Before you climb to rung three, ask whether the weekly report actually has a reader.

When should you stop tuning and build an agent instead?

When the deciding is the work. If the hard part is choosing which of six things to do next, and each step depends on what the last step found, one prompt will keep disappointing you no matter how well you write it. That is the point where the extra machinery of an agent buys something real. Before that, it is cost.

Two other signals: your check step has passed unchanged for a month, and your context brief has stopped growing. If you are still editing the brief every week, the workflow is still teaching you things about your own business. Stay there.

If the first workflow you want is about finding and qualifying leads, Find leads with AI goes into that case specifically.

What to do this week

Find the thing you did last Monday and will do again next Monday. Write the one page brief for it. Run the prompt above against your own columns, then compare what comes back to what you wrote yourself.

Then ask the question that decides whether any of this is worth continuing: if this table were wrong, how long would it take me to notice? If the answer is longer than a week, the workflow is not the problem yet. The check step is.