I’ve built websites, offers, ads, and email systems. The useful question isn’t how much technology I can fit into them. It’s whether the work gets done properly.
That’s how I’d approach a first AI workflow. Learn the job, write down what repeats, and remove one piece of unnecessary effort.
Do the job manually first.
Follow one delivery from beginning to end. What information arrives? What is missing? What do you check? Which steps always happen? Where do you need to make a judgment?
If you don’t understand those questions yet, automation will preserve your guesses. A written checklist is a better starting point.
Separate three kinds of work.
| Work | Good starting method | Example |
|---|---|---|
| Predictable steps | A checklist, template, or ordinary automation | Create a delivery folder with standard sections |
| Language-heavy work | AI draft with a human check | Turn approved notes into a first caption draft |
| Promises and decisions | A person with responsibility | Approve claims, agree a price, or resolve an unusual complaint |
Here’s a small content workflow.
- Collect approved facts. A consistent brief makes missing information obvious.
- Create the draft. Give AI the facts, audience, format, and boundaries.
- Check the work. Compare every factual statement with the source.
- Get approval. The owner confirms the message and any business details.
- Deliver the agreed files. Make clear which version is final.
- Record the outcome. Note time spent, revisions, questions, and any errors.
The draft step can be faster without removing approval. That’s often enough to make a small workflow useful.
Keep a human responsible for these.
- The promise you make to the customer.
- Prices, spending, refunds, and commitments.
- Whether the source information is correct.
- Publishing claims under the business’s name.
- Unexpected situations that the workflow doesn’t cover.
Only use customer information in tools you’re permitted to use for that work. When a fictional or anonymized example is enough to test the process, start there.
Test the awkward case, too.
A happy-path test tells you what happens when all the inputs are present. Also try a missing booking link, contradictory opening hours, or an instruction that asks for a claim the brief doesn’t support.
The right result may be a stopped draft with a question for the owner. Silently filling in a gap isn’t successful automation.
Watch a small scoreboard.
Before and after each change
- Total time to deliver one finished job.
- Number of corrections and revisions.
- Tool cost for that delivery.
- Whether the buyer received the agreed result.
If a workflow saves drafting time but doubles editing, it may not be helping. If a plain template works, keep it. The business doesn’t earn extra credit for having more agents.
Your next move
Pick one repeated task. Write the inputs, the steps, the output, and the final check on one page. Improve that piece before adding another tool.
That’s how I’d build something small enough to understand—and reliable enough to keep using.