How to create an AI agent: pick one small job, write down the rules, connect the tools it needs, and test it on copies. Skip building one when the job is rare, the rules keep changing, or one mistake would be costly.
What does it mean to create an AI agent?
Here is how to create an AI agent in one line: you give an AI a job, the information it needs, and the tools to finish it.
Say you run a small furniture shop. Customers email asking whether a table is in stock. An AI agent could read the email, look in your stock sheet, and write a draft answer. Three things made that work: a clear job, the right information and access to the sheet.
A good test of whether you understand the job: can you explain it to a friend in two minutes? If you can, you can probably explain it to an AI. If you can't, the rules need more thought first.
People use the word agent for many things, from a simple chat helper to a program that works through a long list of steps. What matters is the same in each case. The AI needs a goal, rules and somewhere to look things up.
You don't need to know how the AI works inside, just as you don't need to know how a delivery van's engine works to plan a delivery round.
How do you create an AI agent, step by step?
You create one by starting small: choose one job, write the rules, connect the tools, test on copies, then watch it for a while.
Picture a plumber's office that gets the same question again and again: can you come tomorrow? Here's how you could set up a helper for it.
The steps:
- Pick one job that repeats and has a clear end.
- Write the rules in plain words, as if for a new colleague.
- Collect examples of good past answers or past results.
- Connect only the tools the job needs, such as the inbox and a sheet.
- Test on copies or old cases, not live ones.
- Read every result for a week and note the mistakes.
- Fix the rules, then widen the job a little.
What do you need before you start?
Before you start, you need a clear job, good examples and a person who owns the result.
A bakery that wants help with order emails needs its past replies, its price list and someone who will read the drafts. Without those, the helper is guessing.
Do not forget the boring parts. Where will the results be stored? Who gets told when something goes wrong? Answer those two questions early, and the rest is much easier.
The most common gaps are not technical. They are about knowing your own work well enough to write it down. Here's a short list of what to gather:
- A one-sentence description of the job.
- Ten or so real examples of the job done well.
- The rules you follow, including the exceptions.
- Access to the tools the job touches.
- A named person who checks the results.
- A way to switch it off quickly.
How much time does the job take you now?
Type in your own numbers. The result is the time you spend on the job today, by your own count. It is not a promise of what an AI agent would save.
When should you not create an AI agent?
You shouldn't create one when the job is rare, the rules change every week, or a single mistake would be costly.
Think of a restaurant owner who negotiates one big supplier contract a year. That job is rare and personal. An agent would cost more time to set up than it saves, and a slip would matter.
There is no shame in stopping. A small shop that tries an agent for a fortnight and decides the job is too rare has still learned something useful, for almost no money.
Be careful, too, when no one can say what a good result looks like. If two colleagues would do the job two different ways, write that down first. An agent can only follow rules that exist.
Skip the agent, or keep it small, in these cases:
- The job happens a few times a year.
- The rules live only in someone's head.
- A wrong answer could harm a customer or break a law.
- The job needs a human relationship more than a quick answer.
- You can't check the results.
Should you build it yourself or have it built?
Build it yourself for small, low-risk jobs. Have it built for you when the job runs every day or touches your accounts and customers.
Say an online shop wants supplier bills read and entered in the accounts. The first few are easy. The refunds, scans and unknown suppliers are where a quick home-made helper slips.
That's the work we do. We build AI employees for your business and run them on our servers, working inside the tools you already use: Gmail, Google Drive, Google Sheets and Xero. Sales invoicing and supplier bills run today for clients. Drafted customer replies and other jobs are built to order. You check the work and choose when a step goes automatic.
Try the calculator below. It shows how much time the job takes you now, so you can judge whether it's worth doing properly.
What does a good first agent look like?
A good first agent does one boring job well, in plain view, with a person reading what it produces.
Say a small accountancy practice gets many emails that all ask for a copy of a document. A first agent could read each email, find the document in the right folder and write a draft reply with the file named. That's one job, one folder and one kind of email.
Notice what's missing from that example. It doesn't decide anything important, it doesn't spend money, and it doesn't write to anyone without being read first. That's what makes it a good start.
Once it runs well for a few weeks, you can add a second kind of email. Grow it one step at a time, and keep a note of every mistake you find.
You know it's working when the mistakes are rare, easy to spot, and fewer each week.
Keep a simple tally for the first month. Each day, note how many results were right, how many needed a small fix and how many were wrong. A bakery owner can do this on the back of an order pad.
If the wrong ones don't shrink as you improve the rules, the job may not suit an agent. That's a useful thing to learn early and cheaply. Stopping is a fine result.
Questions
How do I create an AI agent from scratch?
Pick one small job, write its rules, collect examples, connect the tools and test on copies. Then watch it for a week before relying on it.
Do I need to code to create an AI agent?
Not for simple jobs. Many tools let you set one up with boxes and blocks. Harder jobs need more care.
How long does it take to create an AI agent?
A simple helper can be ready in an afternoon. A careful one takes longer, mostly because of the testing.
What is the biggest mistake when creating an AI agent?
Starting too big. Begin with one small job and widen it only when it works.
Can you create an AI agent for me?
Yes. We build AI employees to fit your work, run them on our servers and look after them.
Tell us your workflow. We will build it around you.
Describe the job you'd like off your desk. Write to hello@thenewaiemployee.com and we'll tell you honestly if it's worth building.
Talk to us Related service: Services


