I build AI workflows pretty much every week. I'm the Content Engineer at Delight.ai, and I run our content and organic search program on my own. Workflows are how I keep up with all of it, so my time goes to strategy and writing instead of busywork.
I've built workflows on a lot of different AI platforms by now, and the thing I keep coming back to is that the platform matters a lot less than the process. The buttons are in different places, and the names are different (skills, GPTs, Gems, agents, automations...). But the thinking that goes into a workflow that sticks is the same every time.
The research backs that up. McKinsey's 2026 State of AI survey found that nearly 3 in 4 AI high performers have fundamentally redesigned their workflows around AI, compared with about 1 in 4 other organizations.
This is my personal process for how to build an AI workflow from scratch, plus my opinions on where people tend to go wrong. It works on whatever platform you use.
I'll use one example the whole way through, which is automating my to-do list. It sounds simple, but it touches every step, and it's the workflow I use most.
1. Pick a task to automate with AI (start with the boring stuff)
The best way to start automating tasks with AI is to pick one that's repetitive or eats up more time than it should. It shouldn't be your most complex project, and it definitely shouldn't be "automate my job." Look for something you already do on a regular cadence and could explain to a coworker in five minutes.
A few examples of good starting points:
- Building the deck for a weekly or monthly presentation, where you pull the same data and write the same kind of takeaways every time.
- Keeping your to-do list up to date, because tasks come in from everywhere and keeping the list accurate is its own job.
- Writing a recurring Slack update, like the two I send every Friday (a weekly content wrap-up and our organic wins) that follow the same structure every week.
I always start with repetitive work because you've already done the hardest part of designing the workflow. You know the steps and what a good result looks like, so you can tell right away if the AI got it wrong. And because the task repeats, the time you save compounds every week.
Before you move on, write down the goal in one sentence. For my to-do list, the goal was to make sure every task I'm responsible for ends up on my list and to tell me what to focus on each day. If you can't write that sentence, the workflow isn't ready to build yet.
2. Map out the data and context the task needs
This is the step most people skip, and it's the one that decides whether your workflow works at all. An AI agent can only be as good as what it can see.
For my to-do list, I had to think about every moment where something gets added to my plate. For me, most of my assignments come out of meetings, which made meeting transcripts the foundation of the whole workflow. Without them, the workflow would miss the majority of my tasks.
After meetings, the next biggest sources were:
- Slack messages, where someone asks me for something in a thread or a DM
- Jira tickets that get assigned to me or need a follow-up from me
- My calendar, since every meeting I'm presenting in or running needs prep time
Why does this matter so much? Because an incomplete to-do list is worse than no list at all. If the workflow confidently hands me a list that's missing half my action items, I stop trusting it, and then I'm back to keeping track of everything in my head.
Same goes for any workflow. If your weekly deck is missing a data source, the numbers are wrong, and nobody trusts the deck.
My advice is to list every input you use when you do the task by hand. Then check whether your AI tool can actually reach each one, whether that's through a connector, an integration, an upload, or whatever your platform offers. Any gaps you find now are much easier to solve before you build than after.
Gartner sees the same problem at a much bigger scale. It predicts that through 2026, organizations will abandon 60% of AI projects that aren't supported by AI-ready data.
3. Do a practice run with the AI
This is my favorite part of the process. Before I try to write any kind of formal workflow, I just do the task the way I normally would, but with the AI alongside me. I walk the agent through it step by step, like I'm training a new teammate.
The key is setting expectations up front. I tell the agent to take note of every step we take, so we can turn those steps into a workflow at the end. Something like:
"I want to automate how I manage my to-do list. Let's do it together once, the way I'd normally do it. As we go, keep track of each step, what data we used, and any decisions I make. At the end, we'll use those notes as the basis for a repeatable workflow."
Running with the to-do list example, this forced me to think critically about when and how I actually add tasks to my list today:
- At the end of the day, I check things off and add anything that's high priority for tomorrow.
- After meetings, I transfer my action items from my notes to my list.
Then I asked myself how my list was organized and what was and wasn't working. The biggest problem was that it was one long list with a simple checkbox next to each task. All the prioritization was manual, and there was no sense of how long anything would take.
Creating a data-driven deck had the same weight on the page as filing a Jira ticket for a subheading copy update on a landing page. Those are not the same amount of work, and my list had no way of showing that.
That's the real value of the practice run. It surfaces the steps you do on autopilot, and it makes you look hard at whether the process is worth automating as-is. If your current process is broken, automating it just gets you a broken process that runs faster.
McKinsey's 2025 State of AI report found the same thing at the company level. Redesigning workflows had the biggest effect on whether organizations saw an impact on earnings before interest and taxes (EBIT) from generative AI.
4. Note what you liked and what you didn't
Once the practice run is done, take a few minutes to review it before you build anything. These are the questions I ask myself:
- Was there a point where more data would've helped, or where I should have pulled data or context in earlier?
- Did the steps flow naturally and feel like they were in the right order?
- Did I find myself backtracking and going "oh oops, I actually want to add this step at the beginning"?
That last one is the biggest signal. Every time you backtrack, it means the order is wrong or something important came in too late. If it happens in a practice run with you guiding the AI, it'll happen every time the workflow runs on its own, and the agent won't know to go back and fix it.
I also pay attention to the parts I liked. If the AI summarized something in a format I loved, or caught something I would've missed, I want that baked into the workflow on purpose so it happens every time.
5. Turn it into an official workflow
Now it's time to make it real. I ask the agent to take everything from the practice run (the steps, the data sources, the decisions, and my notes from step 4) and turn it into a saved, reusable process. Depending on the platform, that might be a skill, a custom GPT, a Gem, a project with instructions, or a saved agent.
A prompt I'd use here looks something like:
"Turn what we just did into a reusable skill. Include each step in order, where the data for each step comes from, the output format I want, and the changes we talked about after the practice run."
Don't just accept the first version. Read it like you're the agent seeing it for the first time. The first things I check are whether it explains why each step exists and whether it says where the data lives.
The more of your own judgment you write down here, the less the agent has to guess later.
6. Do a dry run and tweak
Run the workflow from scratch, the way it'll run when you're not watching. Then watch closely for the questions the agent asks you and the places it gets stuck.
Each question is a piece of context that's missing from the workflow. Each place it gets stuck is a step that isn't clear enough. Both are easy fixes once you see them, so add the answers into the workflow and run it again.
My to-do list went through a few rounds of this. For example, I originally had a separate section of due dates at the top of my list, but after a couple of runs, I realized it was just clutter. I dropped it and had the due dates come to me as reminders in Slack instead.
I also learned that if the list isn't simple, I won't use it. So I trimmed each task down to the task itself, a priority level, and an optional due date. That's it.
This step is where a workflow goes from "technically works" to "something I'd use every day." Don't skip it because the first run looked fine.
7. Schedule it
A workflow you have to remember to run is only half automated. Once it's working, set it up as a scheduled task, and put real thought into when it should run.
The right time is whenever the output is most useful to you. For my to-do list, that's two times a day:
- Each morning, it suggests my priorities based on what's due today and what's due by Friday.
- Each night before I leave, it sends me tasks I may have finished but not checked off, plus anything from my meetings, Slack, or Jira that it thinks I missed adding.
I approve the evening suggestions before anything changes on my list. I like keeping that step, because it keeps me in control and it helps the workflow get better at figuring out which action items are mine vs. someone else's.
We follow the same rule for customer-facing AI agents at Delight.ai. The step-by-step playbooks an AI agent follows never go live until a person approves them, and we call those playbooks Actionbooks.
Think about where you'll see the output, too. Mine comes to me in Slack because that's where I already am all day. If I had to open a separate tool to check it, I wouldn't.
8. Update, update, update
This is the step I'd underline three times if I could. AI platforms add new features and connectors constantly, and my own priorities shift too. A workflow that was perfect two months ago can quietly drift out of date.
My to-do list workflow has already changed a lot. I moved the whole list from Apple Notes to Notion, and I retired an older project-tracking workflow once it stopped being useful. Neither of those would have happened if I'd treated the first version as finished.
My rule of thumb is that if I have to correct the AI on the same thing twice, it's time to update the workflow. I also keep an eye out for new sources it could be pulling from. Every time a new connector or data source becomes available, I ask myself whether one of my workflows would get better with it.
Customer-facing AI agents need the same upkeep. Delight.ai's Zero-Touch Improvement feature looks at where an AI agent fell short in real conversations and suggests fixes, and a person approves each one before it ships.
How to build an AI workflow on any platform (the short version)
The AI tools will keep changing, probably faster than any of us can keep up with. But the steps for building a good workflow haven't changed for me across any platform I've used:
- Pick a repetitive task with a clear goal.
- Map every piece of data and context the task depends on.
- Do a practice run with the AI and have it document the steps.
- Review what worked and what didn't.
- Turn it into a saved workflow.
- Dry run it and fix where it gets stuck.
- Schedule it for when it's most useful.
- Keep updating it.
At Delight.ai, this is just how we work. We build AI agents for a living, so it'd be strange if we weren't using them to run our own day-to-day. Start with one boring task this week, and see how much time you get back.
