AI for small businesses in Iowa City and Cedar Rapids: where to start
How I started with AI in my own Coralville business: write down the manual process, then build a system around it. A starting point for any small business.
Trading cards are a low-margin, high-volume business. I have thousands of single cards to list, and listing them by hand cost more of my time than the cards could pay back.
That’s where I started with AI. Not with a strategy. With a pile.
If you own a business in Iowa City, Cedar Rapids, or anywhere between, I think the way I got going will work for you too. Here it is, start to finish.
The problem, in business terms
Every card I sell online needs four things before it can go up.
Two photos. Front and back. Buyers want to see the actual card they’re getting, so I take every photo myself. That alone takes time.
A way to find it again. When a card sells, I can’t spend an hour digging through boxes for it. So every card needs an address: an internal SKU that says which box, which section, which slot.
The right identity. This is the slow part. There are dozens of variants of the same sports card, and telling them apart one at a time turns the whole thing into a loss leader, assuming I value my time at all.
A fair price. I want my prices to match what the card has actually sold for, so they’re fair to the buyer and responsible for the business. This is a business, and I need to treat it like one.
Add a family, hobbies, and a set of vending machines to keep stocked, and doing all of that by hand wasn’t sustainable.
So I needed a system.
Step one: write down the manual process
Before I opened any AI tool, I documented every action and every decision in the manual process. Photograph the front and the back. Work out which variant it is. Search what it has sold for. Write the title. Pick the category. Record where the card is stored.
That included deciding on my sources of truth: which sites I trust when one says a card sold for a specific amount. For prices, mine is eBay’s own record of sold listings, and nothing else. An AI will happily pull a number from anywhere unless you tell it where the truth lives.
It’s tedious, and it’s the most useful thing I did. That document was my starting point. It showed me which steps were just copying and retyping, which needed judgment, and which one I’d never hand to anyone: the decision to publish.
If you take one thing from this post, take that. Your AI plan is a list of what your team does by hand today.
Step two: build the system around the AI
Then I opened OpenAI’s Codex and started assembling the pipeline, one step at a time.
Today AI helps me identify the cards, organize them, research sold listings, suggest prices, write the eBay listings, and save them as drafts. I approve and post from my phone or my computer. The whole run is written up here, including everything that went wrong.
Very little of the work turned out to be “the AI.” Most of it was building the structure around it. These are the terms that turned out to matter:
| Term | What it means | In my pipeline |
|---|---|---|
| Operating procedure | The manual process, written down step by step | The document from step one. It’s on version 2.10 now. |
| Agent | An AI that takes actions, not one that only chats | Codex opens the photos, runs the checks, and fills in the drafts. |
| Harness | The code around the AI that runs each step in order and checks its work | Small programs, with tests, that count the photos, build the titles, and stop the run when something is off. |
| Guardrails | Rules the AI isn’t able to cross | It can never publish a listing. It can never guess a card. |
| Computer vision | AI reading a picture | It looks at the front and back of each card to work out what it is. |
| Image identification | Matching what it sees to one exact item | Telling one variant from the dozens that look almost the same. |
| Listing requirements | The fixed rules the output has to meet | eBay’s 80-character title, the category, the item details, and the shipping policy. |
| Source of truth | The one place you trust for a fact | eBay’s sold listings, for what a card is worth. |
| Sold comps | What the item really sold for, not what sellers are asking | The basis for every suggested price. |
| Approval gate | A point where a person has to act | Condition, final price, and publishing are mine. |
You don’t need to learn these to get started. You need the first one. The rest show up when you build.
What this looks like in your business
You probably don’t sell cards. The shape is the same anyway.
- Something arrives. An email, an invoice, a form, a photo.
- Someone reads it and retypes it into a quote, the books, or a spreadsheet.
- Someone decides what happens next.
AI is good at the middle step. It reads, pulls out the details, and drafts the next thing. It’s good at research that follows a rule, like “only count sold listings.” It should not be making the decision at the end.
Where I’d keep it out
- Anything a customer reads unchecked. A draft is fine. An email that sends itself is a risk you don’t need yet.
- Decisions about money or people. Let it gather the facts. You decide.
- Health records, card numbers, and anything a regulator cares about. Keep those out until you know exactly where the data goes.
Is my data safe?
It depends on the plan and the settings, which is a dull answer and the true one.
Free consumer accounts and business plans are not the same thing. Anthropic, for one, says of its commercial products that “by default, we will not use your inputs or outputs” to train its models.1 Whatever tool you pick, find that sentence for it before your team pastes in a customer list.
You are not behind
It can feel like every other business has this figured out. They don’t.
In March 2026, a program run by Iowa JPEC and the Iowa Small Business Development Center, with Amazon Web Services and PREDICTif Solutions, offered seven Iowa small businesses help putting their first AI tools in place.2 That’s the stage most of the state is at: early, curious, and looking for the first useful thing.
A first month that works
- Week one: write down the manual process for one job. Every action, every decision, and where each fact comes from.
- Week two: sort the steps into three piles. Copying and retyping, which software can do. Judgment calls, which AI can draft for a person to check. And decisions only you should make.
- Week three: automate one thing from the first pile: the one that happens most often. Use the software you already have, like your email, spreadsheet, or accounting tool, before you buy anything new.
- Week four: for a week, do the job both ways and compare the results. Have a person check what the automation produces before anyone relies on it.
If it saves time, do the next job. If it doesn’t, you spent a month and learned something specific.
What does it cost?
Two things: what it takes to build, and what it takes to run. I wrote up both for my own pipeline, with the math, in what AI automation costs. If you want to know why a run costs what it does, where the tokens go shows it call by call.
If you’d like a second pair of eyes
This is work I take on for small businesses around the Corridor. AI consulting is where it starts: I look at how your work gets done and hand you a short, ranked list of where AI would help. When you’re ready to build, AI and workflow automation takes one task from that list and makes it run.
Or tell me about your pile. I started with one too.
Footnotes
-
Anthropic, “Is my data used for model training?” (another site), on its commercial products. Check the equivalent page for any tool you use. ↩
-
Innovation Iowa, “7 small businesses receive $335,000 in support for implementing AI solutions” (another site), 26 March 2026. ↩