Your First AI Project at Work: Where to Start
The biggest mistake businesses make on their first attempt with AI is starting with the most complex or most critical task — fully automating accounting, handing customer service entirely over to AI. The result is usually disappointment and a verdict of “it didn’t work for us.” The right starting point is the opposite: something small, frequent, error-tolerant, and clearly measurable. In this piece I walk through how to pick that task, with a real example.
3 Criteria for the First Project
What makes a task a good candidate for your first AI project isn’t its complexity — it’s three simple criteria. As I explain in the AI automation guide, businesses that follow the right order get to results much faster.
1. Does it happen often?
Pick something done at least a few times a week or day. Automating a report you generate once a year has low payoff; a task you repeat every day pays off fast and keeps compounding.
2. Is the error tolerance high?
AI sometimes gets things wrong in early attempts. The question to ask is: what happens if it goes wrong? If you review and send an email draft yourself, the risk is low. If you let AI approve an invoice directly, the risk is high. Your first project should be chosen from tasks where the cost of a mistake is low.
3. Is the outcome easy to measure?
Pick something where you can compare before and after: minutes spent, error count, response time. You can’t prove an improvement to your team or yourself if you can’t measure it — and you need that proof before you can move to the next step.
Good and Bad First-Project Examples
Concrete examples are clearer than abstract criteria. Here are good and bad candidates for a first project:
| Good First-Project Examples | Bad First-Project Examples |
|---|---|
| Drafting first-response emails to customers | Automating all of accounting |
| Pulling a weekly sales/performance summary | Handing customer service entirely to AI |
| Building an FAQ bot | Delegating critical decisions (pricing, hiring, legal approval) |
| Data entry / cleanup (invoices, stock lists, CRM records) | Letting unsupervised AI write brand voice or legal copy |
The tasks in the right column aren’t impossible — they’re just the wrong choice for a first project. Think of them not as things AI “can’t do,” but as steps you shouldn’t skip before you’ve built trust and process. As I explain in Making Money with AI: Real or a Trap, most of the overhyped promises come from businesses that skipped this ordering.
The technical side should stay simple too: the first project is usually solved with a single prompt, no need to build a complex agent.
A Real Example: What a First Project Looks Like
A small e-commerce business I worked with launched its first AI project on drafting replies to customer questions. Before that, the business owner personally wrote a reply to every order/return question — around 25-30 emails a day, 3-4 minutes each.
What we built wasn’t complex: a prompt template compiled from past replies, a simple flow that reads the incoming email and generates a draft response. AI wrote the reply, the business owner reviewed and approved or edited it before sending — no email ever went out unsupervised.
The first two weeks went slower than expected because the drafts’ tone didn’t quite match the brand voice yet — we revised the prompt three times. After the third week, most drafts could go out with a minor edit, and time spent per email dropped from 3-4 minutes to about 1 minute. A small but measurable win — and the business owner said, for the first time, “this is actually working.”
What Happens After the First Project
Once you’ve landed a small win, your team now trusts AI and you know where to expand next. The simple template I follow:
- Write down the result. How many minutes/errors/steps were saved? This proof makes it much easier to get approval for the next project.
- Apply the same three criteria to a second process. Pick the next task that’s frequent, error-tolerant, and measurable — it can come from a different department.
- Upgrade the tool as complexity grows. If a single prompt isn’t enough anymore, move to a more structured solution like an AI agent — but only once the need actually shows up, not from the start.
- Keep the measurement loop going. Make before/after comparison a habit for every new project.
This kind of expansion is much less risky than one big “AI transformation” project, and it leaves a concrete win at every step. I go into this expansion sequence in more detail in the AI automation guide.
You Can Ask Me to Do This
You can ask me to do this: it’s usually a remote, hourly-billed job that takes between 2 hours and 2 weeks. We pick and set up your first project together; the AI starter package is built exactly for this need. If you need something broader in scope, take a look at AI consulting.
Frequently Asked Questions
What should the budget be for a first AI project?
Most first projects start on a small budget; the real cost is the consulting/setup time. It ranges from a simple 2-hour prompt setup to a 1-2 week job that includes data cleanup and testing.
I have no technical background — can I still start?
Yes. You don’t need to write code for a first project; you just need to ask the right questions and describe the process. I handle the technical setup, you evaluate the process and the result.
What happens if the first project fails?
Since a first project chosen by these criteria is already low-risk, “failure” is usually a small revision need, not a disaster. Revising the prompt a few times, as in the example above, is normal.
Which AI tool should I use (ChatGPT, Claude, etc.)?
Tool choice is secondary for a first project — picking the right task matters more. I usually recommend ChatGPT, Claude, or a custom setup depending on the nature of the task; the decision becomes clear once we’ve defined the project together.