Monday, seven in the morning. The coffee isn’t even finished yet, but the sales admin at a small trading company already has three files open: an export from the sales system, order recaps from the field sales team, and last week’s report. Before the ten o’clock meeting, all of it has to become one concise report for the manager. The work isn’t hard. The problem is that it comes back every single week, and somehow it always takes longer than expected.
That’s where a simple AI workflow starts to earn its keep. A simple AI workflow is a series of AI-assisted steps whose order we set ourselves: the output of one step becomes the input for the next, and there’s a checkpoint before the final result is used.
In effective AI prompting techniques, we learned how to give one clear instruction. In tips for creating AI skills for repetitive work, we learned how to save our work standards so we don’t have to retype them. Now it’s time to chain the two into a single flow that runs from raw data to a ready-to-send report, without letting confidential company data slip out.
The easiest way to picture it: if a prompt is today’s order and a skill is the kitchen’s standard recipe, then a workflow is the kitchen line itself. Ingredients come in, get washed, chopped, cooked, plated, and then tasted by the head chef before the plate goes out to the table. Each station has one job, and each station knows exactly what kind of “handoff” it will receive from the station before it.
How Is a Simple AI Workflow Different from a Prompt or a Skill?
A simple AI workflow differs from a prompt or a skill because what it controls is the order of the steps and how results are handed off, not the content of a single request or the standard for a single type of task. Prompts and skills are still used, but they become building blocks inside the flow.
In its guide Building Effective Agents (December 2024), Anthropic distinguishes two things that often get mixed up. A workflow is a system whose steps are defined in advance, while an agent is a system that lets the AI decide for itself which steps and tools to use. The most basic workflow pattern is called prompt chaining: a task is broken into several sequential steps, and checks can be placed between them (Anthropic calls these gates) to make sure the process is still on track.
For everyday tasks, a workflow is usually the more sensible choice because its results are predictable and errors are easy to locate. The same guide also recommends starting with the simplest solution and adding complexity only when it’s needed.
So you’re not left wondering where to begin, here are eight steps to build one. We’ll use the same example from start to finish: a weekly sales report at a small trading company, an imaginary distributor of staple foods with a few salespeople and one admin. The pattern will feel familiar to many small offices, and shop owners or small business owners can use the same steps for their own sales recaps.
Pick One Task Worth Automating
The tasks best suited to a simple AI workflow are repetitive, have clear inputs and outputs, follow fairly fixed rules, and carry little risk if the first draft is still wrong. A weekly sales report meets all four conditions, as long as the final decision stays in human hands, in this case the sales manager’s.
That last condition is often forgotten. In a field experiment by Dell’Acqua et al. published in Organization Science, 758 Boston Consulting Group consultants worked on tasks with and without GPT-4. On tasks within the AI’s capabilities, AI users completed 12.2% more tasks, 25.1% faster, and with higher quality. But on one analytical task deliberately chosen to sit outside those capabilities, their chances of getting the right answer actually dropped by 19 percentage points compared with the group without AI. The researchers called this uneven boundary of capability the jagged technological frontier.
In other words, AI can be a huge help on one task and quietly lead you astray on another that looks similar. In a sales report, cleaning up data and writing a summary are jobs whose results we can easily check. On the other hand, setting selling prices, granting special discounts to customers, or choosing suppliers shouldn’t be handed over to AI; let the AI stick to providing input for the decision.
Write Down How We Do the Work Manually, Step by Step
Before involving AI, write down the steps we’ve been doing manually all along, from input to output. This simple map becomes the skeleton of the simple AI workflow we’re going to build.
For our sales admin, the map looks roughly like this: (1) download data from the sales system and merge it with the order recaps from the sales team, (2) standardize product names that are written in different ways, for example “Cooking Oil 2L”, “cooking oil 2 liter”, and “CO 2L”, (3) add up sales by category, (4) compare them with last week, (5) spot products whose stock runs out quickly, and then (6) write a summary for the manager. (If the list stretches to twelve steps, maybe it’s the bookkeeping that needs tidying first, not the report.)
This exercise also surfaces rules that were never written down, such as the habit of separating cash and credit sales, or not counting returned goods as sales. Rules like these have to be stated explicitly, because AI can’t read habits that live only in our heads.
Protect the Real Data Before It Enters the Flow
The data that goes into a simple AI workflow should not be the real data, but sample (dummy) data, data with identities masked, or summaries that no longer contain business secrets. This step happens before a single prompt is sent.
The reason is simple: sales data holds more secrets than it seems. There are customer store names along with the owners’ phone numbers, salespeople’s names, supplier names, purchase prices, and even profit margins. At the office, that data belongs to the company, not to us personally, so company rules on data and AI tools must be followed. People’s names and phone numbers count as personal data, whose processing in Indonesia is governed by Law No. 27 of 2022 on Personal Data Protection, while purchase prices and margins are trade secrets that could cause real damage if they leaked to competitors. Here are five habits you can apply right away:
- Design and test with dummy data. Create a made-up table whose columns and format match the real data exactly. Build and test the flow with this data until it runs reliably.
- Mask identities with codes. Customer names become “CST-017”, salespeople become “SLS-02”, suppliers become “SUP-03”, and territories become “Area A”. The table that maps codes to real names stays on your own computer and is never uploaded.
- Drop columns you don’t need. A sales report by category doesn’t need phone numbers, addresses, national ID numbers, bank account numbers, or purchase prices. Delete them before uploading.
- Summarize before you upload. If the AI’s job is only to write the narrative, send totals by category and by day, not the details of every transaction.
- Keep formulas separate from data. Ask the AI to create formulas or a report template based on the dummy data, then run those formulas yourself on the real data in Excel or Google Sheets. That way, the actual numbers never leave your computer.
To be fair, masking names alone isn’t always enough. Latanya Sweeney’s classic research at the Data Privacy Lab estimated, based on the 1990 census, that 87% of the United States population could potentially be uniquely identified by just the combination of ZIP code, gender, and date of birth. A follow-up study by Philippe Golle using the 2000 census found a lower figure, around 63%, but the message is the same: a few pieces of data that look trivial can point to a specific person when combined. In a company sales report, a combination of territory, date, and an unusually large order value can likewise point to a particular customer. So company data policy remains the main reference, and real data should only go into AI services that have actually been approved.
Break It into Small Steps with One Task per Step
Each step in a simple AI workflow should handle only one task, because a request that piles on several tasks at once is more likely to miss. That’s the heart of prompt chaining: a big task is broken into several small requests that run in sequence.
Claude’s official documentation on prompt chaining lists three benefits: each subtask gets full attention so there are fewer errors, the instructions are clearer, and problems are easier to trace. If a report from one giant prompt is wrong, we’re left guessing where it went wrong; if the flow is staged, we can point right at it: “Oh, it’s the product-name step that’s off.”
The six manual steps above can be slimmed down to three:
- Step A, Clean: the AI standardizes product names and categories, then flags odd rows, such as negative quantities or zero prices.
- Step B, Calculate: the AI does not calculate anything itself; instead, it writes spreadsheet formulas for totals by category, best-selling products, and comparisons with last week. The calculations run in the spreadsheet.
- Step C, Report: using the spreadsheet’s results, the AI writes a three-point summary and a draft report email for the manager, including a note on products whose stock needs topping up soon.
Step B is split that way on purpose. Language models are good at writing and explaining, but they can slip when adding up dozens of rows of numbers while still sounding confident. Spreadsheets are the opposite: no good at storytelling, but their calculations are consistent. So let each one do the part it’s best at.
Set a Standard Handoff Format Between Steps
A handoff format is the standard shape of each step’s output, such as a table that always has the same columns, so the next step receives exactly the input it expects. Without it, the AI workflow can easily “break” halfway through.
If Step A sometimes produces a five-column table and sometimes a seven-column one, the formulas in Step B fall apart immediately. It’s just like a vegetable station that sometimes sends diced carrots and sometimes whole ones; the cook at the next station is bound to get a headache. For our example, the dummy data going into Step A looks like this (dates in day-month-year format, prices in rupiah):
Date | Product Code | Product Name | Quantity | Unit Price
03-11-2026 | PRD-014 | Cooking Oil 2L | 12 | 36,000
03-11-2026 | PRD-022 | premium rice 5kg | 5 | 78,000
04-11-2026 | PRD-014 | cooking oil 2 liter | 9 | 36,000
The instruction for Step A could be written like this:
Clean up the sales table below. Standardize product names based on Product Code, and add a Category column (Oil, Rice, Sugar, Beverages, Others) and a Total column (Quantity × Unit Price). Do not change the Quantity and Unit Price figures. For rows with a negative quantity, a zero price, or an empty product code, write “NEEDS CHECK” in the Notes column. Output a table with these columns: Date, Product Code, Product Name, Category, Quantity, Unit Price, Total, Notes.
Step B then only needs to be told that its input is that eight-column table and asked to write formulas. The result would be something like =SUMIFS(G:G,D:D,"Oil") for total sales in the oil category (the separator may be a comma or a semicolon depending on your spreadsheet’s language settings), plus a percentage-change formula: (this week’s sales − last week’s) ÷ last week’s. If the report format is already saved as a skill, like the monthly report skill example in the previous article, Step C just calls that skill. This is where the three topics meet: clear prompts fill each step, skills store the standards, and the workflow sets the order.
Add Automatic and Human Checkpoints
A checkpoint is a check between steps that decides whether a simple AI workflow may continue or has to stop first. There are two kinds: automatic checks for clear-cut rules, and human checks for things that require judgment.
Automatic checks work for anything that can be counted or matched. In our example, three matter most: the number of rows after cleaning must equal the number of rows in the original data, the total revenue from the formulas must match the total in the sales system, and no row marked “NEEDS CHECK” may slip through. If any of these fails, the flow stops and we check the data first. This matters because AI tends to “fill in” blank or odd spots with guesses that sound convincing.
Now, a human check is a must before results go out to other people or are used to make decisions. The AI can draft the email, but the admin still rereads it, and then the manager, who knows the warehouse situation and upcoming promotions, reviews it before it’s taken to the meeting. Automation platforms offer this kind of station too; Zapier, for example, has a Human in the Loop step that pauses an automated flow until someone reviews, edits, and approves the AI-generated draft. This is the head chef tasting the dish before the plate goes out.
Run It at the Level That Fits Your Needs
A simple AI workflow can run at three levels, from a manual relay in a chat window to a flow that runs on its own through an automation platform. Start at the lowest level, and move up only once the flow has proven stable.
The table below compares three levels of running a simple AI workflow in terms of how they work and which tasks suit them best.
| Level | How it works | Best for |
|---|---|---|
| Level 1: Manual relay | Send a prompt per step, copy the output into the next step | Trying a new flow, weekly tasks |
| Level 2: Skill per step | Call saved skills or templates in order | Routine tasks with fixed standards |
| Level 3: Automation platform | The flow is triggered automatically, e.g. by a new email or form | Frequent tasks in large volumes |
So most office workers and small business owners can already get the benefits of a simple AI workflow at Levels 1 and 2, without extra tools or connecting any accounts. A weekly sales report, for example, can simply run as a three-prompt relay every Monday morning.
Level 3 uses platforms such as Zapier, Make, or n8n to connect email, forms, spreadsheets, and AI services. The risk goes up as well, because those platforms hold access to our accounts, including the ones that store real sales data. Grant as few permissions as possible, strengthen account security as discussed in ways to secure your digital accounts, and make sure the tools you use are approved by your company.
Measure the Results, Then Maintain the Flow
The success of a simple AI workflow is measured by the time genuinely saved after subtracting the time spent preparing data and checking, not by how sophisticated the flow is. So record how long the work takes before and after using the flow.
As a rough illustration (example figures, not actual measurements): if the manual sales report takes two hours every week and the workflow brings it down to 45 minutes, including masking data and matching totals, we save 75 minutes a week. Over a year of about 48 working weeks, that adds up to 60 hours, or roughly seven and a half working days. Not bad for just one report, right?
That figure is in line with field findings. An analysis by the Federal Reserve Bank of St. Louis by Bick, Blandin, and Deming (2025) found that workers using generative AI reported saving an average of 5.4% of their work hours in a week. For someone working 40 hours a week, that’s about 2.2 hours. The survey was conducted in the United States and is based on respondents’ own reports, so it’s natural if results where you work are different.
Finally, maintain the flow: save versions of the prompt for each step, update the category list when new products are added, and occasionally test it with dummy data that is deliberately messy, not just clean data. A flow that never gets checked can quietly drift off course over time.
Frequently Asked Questions about AI Workflows
Do I need to know how to code to build a simple AI workflow?
Building a simple AI workflow doesn’t require coding. A manual relay in a chat window and calling skills in sequence only need everyday language, while automation platforms generally provide visual flow builders. Coding is only needed for highly specialized flows.
What is the difference between an AI workflow and an AI agent?
An AI workflow runs steps in an order that humans have defined, while an AI agent decides for itself which steps and tools to use to complete a task. Workflows are easier to predict and supervise, which makes them better suited to everyday tasks with a fixed pattern.
How many steps should a simple AI workflow have?
There is no standard number of steps for a simple AI workflow; what matters is that each step handles only one task. For a start, three to five steps are usually still easy to supervise. If a flow feels too long, it may actually be two flows that should be separated.
Can an AI workflow run on the free versions of ChatGPT, Gemini, or Claude?
An AI workflow at the manual relay level can run on any AI chat service, including free versions, because it simply involves sending prompts in sequence. Features such as saved skills, large file uploads, or integrations with other apps may depend on your subscription plan, so check the terms of the service you use.
How do I create dummy data that resembles real data?
Good dummy data uses the same column names, date format, and number ranges as the real data, but its contents are made up. Copy the structure of the real table, replace every name with a code, fill in random numbers that still make sense, and add a few deliberately wrong rows so the checkpoints get tested too.
So, Which Task Will You Chain First?
Back to that Monday morning. The dummy sales data has been used to test the flow, the real data is processed with formulas in the office spreadsheet, and the AI only cleans up, writes formulas, and drafts the summary. The admin just matches the totals, and the report is on the manager’s desk well before the ten o’clock meeting. The boring part has been handled by the kitchen stations we set up ourselves.
In short, there are eight steps: pick a safe task, map the manual work, protect the real data, break it into steps, set a handoff format, add checkpoints, run it at the right level, then measure and maintain. The first step is simple: pick one task this week, create ten rows of dummy data, and run it as a manual relay in a chat. That’s your first simple AI workflow.
The time you save deserves looking after, too. A workflow doesn’t guarantee that free time automatically fills up with good things; it only opens up space, whether for more strategic work in the spirit of the habits of an effective and reliable worker in the AI era, learning something new, or heading home to family a little earlier.
If you had to pick one report or routine task to chain into a workflow this week, which one would you most like to hand over to the AI “kitchen”? Share it in the comments, including how you mask your data before uploading it; tips and corrections from readers always make posts like this more useful.
In the end, a good workflow isn’t the most sophisticated one, but the one that keeps data safe while giving us more time for the things that truly need a human mind.
This post is also available in:
