November 11, 2026 · Jess Pischel
Most business leaders I talk with have never built an AI agent. Every one of them has explained a task to a new hire.
That is the same skill. An AI agent is an AI assistant that has been set up once for one specific job. You are not writing code. You are writing down how the job is done.
Here is the plan I use with teams to build a first one. It works in whatever AI tool your company already uses, because the names change from tool to tool but the parts do not.
One note before you start: follow your company's AI policy, use only the tools you are allowed to use, and only share information those tools are approved to handle.
Every agent I build with a client has three parts:
Depending on your tool, this might be called an agent, a custom assistant, a skill, or a Gem. Do not get stuck on the name. If you can fill in those three parts, you can build it.
The most common mistake I see is starting with the tool. Someone asks me, "Which AI should we use?" and my answer is always another question: what is the job?
Before you hand any task to AI, ask three questions. It takes about a minute.
Three yes answers make a good first job. Good examples: answering routine policy questions from the employee handbook, drafting the weekly team update, or turning meeting notes into follow-ups.
Pick one. Not your whole business, and not a whole department. One job.
AI can only work with what you give it. When people tell me AI "doesn't sound like us" or "gets our business wrong," it is almost always because nobody told it who they are.
For a policy helper, the knowledge is the handbook. For a weekly update, it is two or three past updates you were proud of. Less is safer, and it usually gives better answers too. Your whole shared drive is not background. It is noise.
If the knowledge for this job lives in your head and not in a document, do not sit down to write it all out. Have AI ask you the questions and write it for you. Paste this into your AI tool:
Answer the way you actually talk. Save what it writes in a folder you own, so you can use it in any tool, now or later.
This is the heart of it. Tell the agent who it helps, what it does, the tone to use, and what to do when it does not know. That last part matters most. Giving it permission to say "I could not find that" is what keeps it from guessing.
Here is a starting point you can copy:
Then add three or four starter prompts, such as "How many vacation days do new employees get?" or "Draft this week's update from these notes." They show your team where to begin.
When the work comes back wrong, do not argue with it in the chat. Go back and fix the instructions. They get better every week, the same way a good training document does.
Ask it three kinds of questions:
Check every answer against the real document. If it guesses on the third one, tighten the instructions and test again. Do not skip this step. It is the difference between a helper your team trusts and one they quietly stop using.
At this point your agent works with what you paste in or upload. The next level is connecting it to your real email, calendar, or files, and then having it run on a schedule.
Before you connect anything, ask two questions. Is this tool approved for this information? And what is the least access the job needs? If it only needs to read your calendar, it does not need to send email.
When you are ready to schedule it, every automation has the same four parts:
My rule is simple: automate drafts, not sends. Every Monday morning, a draft of the team update is waiting for you. You read it, fix it, and send it yourself.
The question I hear most is, "How do I trust it to just do the work?" You do not start by trusting it. You start by checking it, the same way you would with anyone new on your team.
AI gets the work started. Your people own the result.
I will build it with you. Book a call and we will pick the job, write the instructions, and test it on your real work. If you are not sure which job to start with, our free AI assessment shows you where AI can help first.