How to Set Up an AI Agent for Your Business: The Job Comes Before the Tool

Picture an AI agent with access to your email, your calendar, your CRM, your shared drive, your invoicing system and a few more apps for good measure. It took an afternoon to connect everything. It can read almost anything in the business.

Then you open it on Monday morning and type, "What should I work on?"

That agent has plenty of access and no job. And an agent like that is a risk that also needs your attention every day.

Most guides on how to set up an AI agent for your business start with the tool: which platform, which model, which apps to connect. I think that's backward. The tool is the easy part, and the choice changes every few months anyway.

The job is the hard part, and it's the part that decides whether the agent ever takes anything off your plate.

So this guide starts with the job. You'll define what the agent actually is for you, write its job description, decide how much it's allowed to do on its own, connect only what that job needs, and test it before you rely on it.

What an AI agent actually is for a small business

OpenAI's practical guide to building agents describes agents as "systems that independently accomplish tasks on your behalf." It breaks an agent into three parts: a model that does the reasoning, tools it can use to look things up or take action, and instructions that tell it how to behave.

For an owner, that boils down to something simpler. A chatbot answers when you ask it something. An agent is given a recurring piece of work, gathers what it needs, and carries the work forward until it's done or until it needs you.

That's the whole definition you need for now.

It also tells you where most setups go wrong. People spend their time on the model and the tools, and they write the instructions in about thirty seconds.

The instructions are the job description. If those are vague, a better model just gives you more confident vague work.

AI agent setup starts with a job description

Before you open any settings page, write the agent's job the way you'd write it for a careful new hire. A good job description answers four questions.

Trigger, inputs, output and what "done" means

What starts the work? A new form submission, an email to a certain address, a quote that's gone three days without a reply, a set time every Monday. If you can't name the trigger, the agent will only work when you remember to ask it, and that's just a chat window with extra steps.

What information does it need, and where does it live? List the inputs. The inquiry itself, your service list, your FAQ, your calendar availability, past quotes. If an input lives only in your head, write it down now or the agent will guess.

What does it produce? Be specific about the output. "A reply" is vague. "A draft reply saved in the inquiry thread, under 150 words, that answers the question from the FAQ and offers two call times" is a job.

What does "done" mean? This is the question owners skip most often. Done might mean the draft is saved and you've been told it's there. It might mean the inquiry is labeled and logged in your CRM.

Write it down, because an agent will happily report that it's finished when it's only started.

Then add a fifth section most people forget: what the agent is never allowed to do. That list often ends up being the most useful part of the whole description.

A worked example: inquiry triage

Here's what that looks like on paper. This is a hypothetical example, not a client story.

A small landscaping company gets inquiries through its website form. Some are new customers asking for a quote, some are existing customers with a scheduling question, and some are vendors or spam. The owner reads every one, usually at night, and the good leads sometimes wait two days for an answer.

The job description might read like this:

Job: Inquiry triage.

Trigger: A new submission arrives from the website contact form.

Inputs: The submission, the service area list, the services page, the FAQ document and the current weekly schedule.

Output: Each inquiry is sorted into new lead, existing customer, vendor or spam. For new leads inside the service area, a draft reply is saved that answers their question from the FAQ and asks for the property address and photos. For existing customers, a short summary is sent to the owner.

Done means: Every submission from the last day is sorted, every new lead has a saved draft, and the owner gets one summary each morning listing what came in and what's waiting.

Never: Quote a price, promise a date, offer a discount, reply to a complaint, or send anything without the owner's OK. If an inquiry doesn't fit a category, put it at the top of the summary and leave it alone.

Read that "never" list again. Every item on it is a decision the owner keeps.

Prices, dates, discounts and upset customers are all promises or relationship calls. In The Five Keeps, I group decisions like these into five areas: the premise, the promise, the relationship, the money and anything with irreversible risk.

An agent can prepare those decisions for you. It shouldn't make them.

Notice also what the job description doesn't say. It doesn't name a tool. You could run this job on almost any of the major AI platforms, and you could move it later without rewriting the job.

If you want a fill-in version of this, the AI Team Starter Kit has a job description template for your first agent: what it does, what it never does and what done looks like.

Decide its authority: research, draft or send

Once the job is written, decide how much the agent is allowed to do without you. I think about this as three levels.

Research. The agent can read, gather and summarize. It finds the information and hands it to you.

Nothing leaves the building, and nothing changes in your systems. This is the safest place to start.

Draft. The agent prepares the work, like a reply, a follow-up, an invoice note or a calendar suggestion, and saves it where you can check it. You still press send. For most first agents, this is the right level.

Send. The agent acts on its own. It sends the email or updates the record without waiting for you. This is where the time savings get big, and it's also where mistakes reach customers before you see them.

Start every new agent at research or draft. Move a specific action up to send only after you've watched it do that action well many times.

Move actions up one at a time. The inquiry agent might earn the right to send a standard "thanks, we got your request" note long before it's allowed to answer a pricing question, and it may never get that second one.

The OWASP guidance on excessive agency, written for people who build AI applications, makes the same point in security terms. It names three causes of trouble: too many functions, too many permissions and too much autonomy.

Its example is an email assistant that only needs to read mail. If it can also send, a crafted incoming message could trick it into forwarding private information.

One fix OWASP lists is simply requiring the user to review and hit send on every drafted email.

OpenAI's guide says something similar from the builder's side. Actions that are sensitive, irreversible or high-stakes should trigger human oversight "until confidence in the agent's reliability grows."

So should an AI agent be allowed to send emails on its own? Eventually, for some emails, maybe. Not on day one.

Connect only what the job needs

Now you can open the tool.

The major platforms all offer some way to give an assistant standing instructions and connect it to other apps. ChatGPT, Claude and Microsoft Copilot each have their own version, and the names, features and plan limits change often enough that you should check each vendor's current help pages before you pick one.

For a first agent, the platform you already pay for and know your way around is usually the right choice.

Whatever you choose, let the job description decide what you connect. Go through your inputs and outputs and ask what access each one needs.

For the inquiry triage example, that's read access to form submissions, read access to the FAQ, the service area list and the schedule, and permission to save drafts and post one morning summary.

It doesn't need your invoicing system or your whole drive, and it doesn't need permission to send.

When a connector offers more access than you need, use the narrowest setting it allows. If it's all or nothing, think hard about whether this job needs that connector yet.

Every extra connection is one more thing the agent can get wrong, and one more thing you have to think about when it does.

Anthropic's engineering team, in Building effective agents, recommends "finding the simplest solution possible, and only increasing complexity when needed," and notes that sometimes this means not building an agent at all. They also point out that an agent's autonomy brings "higher costs, and the potential for compounding errors."

That's a good test for a small business. If a simple rule or a saved template would do the job, use that. Save the agent for work where something has to read, sort and decide what comes next.

Test it, then name who owns it

An agent that works on your first try with a perfect example isn't tested yet. Before you rely on it, run it on three kinds of cases on purpose.

The normal case. A clean inquiry from a new customer inside your service area, asking a question your FAQ answers. The agent should sort it correctly and save a good draft.

The missing-information case. An inquiry with no address, or a question your FAQ doesn't cover. The agent should ask for what's missing or bring it to you. It should not fill the gap with a guess that sounds confident.

The escalation case. An upset customer, a request for a discount or a question about a price. The agent should put it in front of you and draft nothing that makes a promise.

OpenAI's guide calls this capturing edge cases, and it recommends writing instructions for what to do "when a user provides incomplete information or asks an unexpected question." Anthropic recommends "extensive testing in sandboxed environments" for agents. For a small business, a sandbox can be as simple as a test form, a few made-up inquiries and a folder of drafts nobody sends.

When something fails, fix the job description before you change tools. A failure usually traces back to a missing input or a limit nobody wrote down.

Then check that done actually means done. In Proof Before Done, I wrote about the gap between an agent reporting that it finished and the work actually being finished.

Open the drafts in the inbox and read them yourself, instead of trusting the agent's summary.

Last, name an owner. Somebody has to read the morning summary, check the drafts, notice when the agent starts getting something wrong and update the job description when the business changes.

In a small company, that's often you. Write the name into the job description anyway. An agent nobody owns slowly drifts away from how the business actually works, and nobody notices until a customer does.

If you're working out where an agent like this fits in a bigger plan, my article on how to implement AI in a small business without handing over everything walks through picking the first job, running a short pilot and deciding when to add the next one.

Common questions about setting up an AI agent

Can I set up an AI agent without coding?

Yes, for jobs like the one above. The mainstream assistants let you write instructions and connect common apps without code.

Coding matters more when you need custom connections, unusual systems or tight control over every step. The job description is the same either way.

What's the difference between an AI agent and a chatbot?

A chatbot responds when you ask it something. An agent has a recurring job, uses tools to gather information and take action, and keeps working until the job is done or it needs you. OpenAI's guide puts simple chatbots and single-turn LLMs outside its definition of an agent, because they don't use the model to run the work from step to step.

Should an AI agent be allowed to send emails on its own?

Not at first. Start with drafts you review and send.

Once a specific kind of email has been right many times in a row, you can let the agent send that kind alone and keep reviewing the rest. Anything involving price, dates, refunds or an unhappy customer stays with you.

How do I test an AI agent before relying on it?

Run it on a normal case, a case with missing information and a case that should come back to you. Do it with test data or drafts nobody sends. Check the actual output in the actual system, and fix the job description when something goes wrong.

Can I set up an AI agent for free?

Sometimes, for a small test. Several AI tools have free plans, but they usually come with lower usage limits, and features like connected apps or team workspaces may need a paid plan.

Check the vendor's current pricing page, because those limits change often. Your time is the bigger cost at the start: writing the job, testing it and checking the output for the first few weeks. Pay for more only after the first job is working.

Start with the job description

Setting up an AI agent for your business comes down to doing things in the right order. Write the job first: the trigger, the inputs, the output, what done means and what it's never allowed to do.

Decide whether it researches, drafts or sends. Connect only what that job needs. Test the normal case, the missing-information case and the escalation case, and put a name next to it.

The tool comes last, and it's the easiest decision in the whole process.

If you want the template, the AI Team Starter Kit includes a fill-in job description for your first agent, along with the worksheets for choosing which job to start with. And if you already know the job and want help setting it up, you can request a discovery call.