How to Build a Small AI Team: From Chatbot Prompts to Repeatable Systems, by Paul Kalra, Schulich alum and founder of Part n Parcel. Second in a series on building your way into a career.
The build that changed how I work did not start at work. It started on an ordinary weeknight in my kitchen, when my wife put a name to something she had been carrying: the mental load. Two kids, two careers, and one running list in her head of everything that could be forgotten. Spirit days. Camp forms. What was for dinner. Which school email went to her inbox instead of mine.
I knew exactly what she meant, because my mornings were six apps. Weather, then email, then the meal planner, then two newsletters, then a website for event updates, then trying to get her attention to ask what the school had sent her that I never saw.
So we built Nora.
Nora is our family’s AI assistant. She reads both of our calendars and both of our inboxes, and every morning there is one message waiting: the weather and what it means for dressing the kids, what each kid needs at school or camp that day so no spirit day gets missed, tonight’s dinner from the meal plan we agreed on over the weekend, anything happening locally worth knowing about, and a rundown of the week ahead. Six apps became one message.
She acts, too. When we needed an airport transfer for a family trip to Portugal, Nora found three companies matching our requirements, emailed all three, collected the options, recommended one, chased the details we were missing, and split the cost with the rest of the group. Every email went out as a draft I approved first. The only thing I did myself was pay.
Nora is one assistant with wide access and many jobs. Call it god mode. For a family, god mode is exactly right.
Then I tried the same thing at my company, and it broke.
God mode works at home. It failed at work.
At work, different people need different things from different agents. The person handling our marketing should not be talking to the same agent that watches our finances, and neither agent should hold the other’s permissions. One all-powerful assistant either becomes a bottleneck or holds more access than any single job should. So work got a team instead: several agents, each with one job, its own skills, and only the access that job needs.
God mode vs. a small AI team: god mode is one AI agent with wide access and many jobs; a team is several agents, each with one narrow job and only the access that job needs. God mode fits a household or a solo project, where one trusted assistant serves one audience. A team fits work, where different people need different agents, and each agent needs different skills and permissions.
My company, Part n Parcel, is a managed shipping economics company for Canadian e-commerce merchants: two people serving hundreds of brands, now a few months into running it with a small AI team. A chief of staff triages my email and hands work to the others. One agent works on winning new business. One watches the clients we have and flags anything at risk. One helps us keep customers happy. One checker audits them all.
We are still testing. Some weeks the agents save us days. Some weeks we spend an evening fixing how one of them talks. The direction is right, and the goal is clear: the repeatable work moves to the team, so two people can spend their time on judgment and relationships.
In the first article of this series, I made the case for the Build Before You Apply method: create visible proof of work before you ever submit an application. This article is the promised next step, because managing AI like a team is the most valuable proof you can build right now, and you do not need a company to practice it. You need a problem, a laptop, and a willingness to just start.
The AI Team Ladder: from prompts to a team
Here is the whole idea in one sentence: stop asking AI for answers and start giving it jobs.
The AI Team Ladder: the progression from prompting a chatbot for one-off answers, to saving fixed instructions that run the same way every time (a skill), to giving AI one narrow job it performs on its own with memory and rules (an agent), to running several agents with defined roles, shared knowledge, and approval rules in a workspace where you act as the manager (a small AI team).
Rung 1 is prompts, where almost everyone is. You ask Claude, ChatGPT, or Grok for help with one thing, once. Useful, but nothing accumulates.
Rung 2 is a skill: a saved set of instructions for one job, written once and run the same way every time. In the AI world these are called skills, and you do not have to write them all yourself. Before you build one, ask your AI whether someone has already published a skill or playbook for that job on GitHub. Often they have. Our own developer agent runs on a playbook a stranger wrote for non-coders and shared for free.
Rung 3 is an agent: a skill with a schedule and a memory. It does not wait for you to open a tab. It runs when something happens or a time arrives, remembers yesterday, and brings you results instead of waiting for questions. Keep it narrow. One job, a fixed output format, and permission to say “not enough data” instead of guessing. Give your first agent a boring job on purpose. A morning brief is perfect: daily, useful, and nothing breaks if it gets a detail wrong.
Rung 4 is a team. Different jobs get different specialists, plus the two things specialists need: shared knowledge and a manager. The manager is you.
Do this now: pick one task you have done at least three times and ask your AI to turn it into a saved skill. That is rung 2, and it takes twenty minutes.
Just start: what we used and how we set it up
Setup is where most people stall, so here is the smallest version of it.
You do not need to understand how any of this works before you begin. Open Claude, ChatGPT, or Grok and say: I want to build a small AI team to help with this problem. Ask me what you need to know, then help me build the first piece. That is how we started. I opened Claude Code, an AI coding assistant, described the team I wanted, and it built the pieces while I approved each step. If you can describe the team clearly, AI can build it with you. Describing it clearly is the skill.
Then you need somewhere for the team to live. Two options.
Buzz is what we use. It looks like a group chat, except some of the members are your agents, each with its own name. It is free, and it runs on the AI subscription you already pay for rather than on pay-per-use credits that quietly add up. It lets different agents run on different AI providers, which matters once you add a checker. And each agent’s job and rules live in a plain text file it reads, so the management layer is something you can open and edit. The cost is setup: it takes an evening or two more, and you are the IT department afterwards. AI walks you through it.
Grok Bot, launched by xAI in August 2026, is the buy-it option: always-on agents that come hosted and maintained. Faster to start, nothing to maintain, but you pay monthly and you are tied to one provider’s models.
Whichever you choose, put the team on a small rented server in the cloud rather than your laptop, so it keeps working when the lid is closed. AI can walk you through that too.
The decision rule: if your goal is learning and your budget is student-sized, build it, because maintaining the system is part of the education. If you only want the output, buy it.
The manager’s rules
Anyone can spin up five agents. The skill employers will pay for is running them without a mess. Start with these.
Gate money and reputation. An agent can research, draft, track, and summarize on its own, because all of that can be undone. Anything involving your money or your name gets your approval. In between those extremes, autonomy is earned, not assumed:
The Ten Approvals Rule: an AI agent earns the right to do a task on its own only after its manager approves that same task ten times in a row with no corrections. Until then, the agent drafts and the human decides. Money and reputation are never delegated, whatever the streak.
That is why Nora’s vendor emails were drafts until she proved herself, and why paying was mine either way.
Make hiring hard. Every agent you add means more accounts to secure and more training to do, so the team stays small by default:
The Hiring Bar: a new AI agent is justified only when a job needs a different audience, a different brain, or a different schedule than the existing team provides. Everything else is a new skill taught to an agent you already have.
At my company, the chief of staff enforces this on me. She challenges every hire I propose, and most of my ideas fail the bar and become a skill instead.
Your checker runs on a different brain. One of our agents exists to challenge the others: check their work, push our thinking, ask why we are not aiming ten times higher. He deliberately runs on a different AI provider than the rest of the team, because a checker with the same brain has the same blind spots. Two models disagreeing is information. Five copies of one model agreeing is an echo. This is why it matters that your workspace allows more than one provider.
Give the team a memory and a direction. Our agents share one knowledge base that they keep updated and we humans feed. When I watch a useful video or read a good paper, it goes in, and every agent gets smarter from it. Yours can be a shared folder of notes every agent is told to read. Then, each quarter, I sit down with the chief of staff and set goals. She asks me the questions she cannot answer from data, then tracks how we are doing, and mid-quarter the checker pushes the thinking. That is ordinary management. It works on agents too.
Do this now: describe your problem to your AI in three sentences, then say: ask me what you need to know before we build anything. Let it do the planning with you. The plan comes out of the conversation.
The tenth grader doing it right now
My niece is in grade 10. She and her mom run a small jewelry business, selling bracelets and necklaces at trade shows. It started as a bonding project, and then it ran into the question every business eventually asks: how do we grow online, so we are not spending every weekend behind a table?
So she is building her own small AI team to find out, and the build itself is her story. A tenth grader who can show a university admissions reader a working team, the rules she wrote for it, and what it did for a real business has evidence, where most applicants have only potential. If she can do it before grade 11, you can do it before graduation.
What it actually costs
A week of evenings gets version one working. A pro subscription to one AI assistant for the first month, roughly the cost of a textbook, is worth it while you learn, and a small cloud server is a few dollars a month. Budget for the unglamorous part too: the first few weeks are noisy. You will train the agents, answer their questions, fix how they talk, and tune down how much they send you to review. That period is the onboarding you would give any new hire, and describing it is what makes you credible in an interview.
The team you graduate with
Most students will graduate having used AI. A few will graduate having managed it. When an interviewer asks about AI experience, one group will describe a chatbot. You will present an org chart and its results.
Full transparency, as before: this series is produced the same way, with AI drafting alongside me and a human making every final call. Climb one rung at a time. Save the skill this week. Give it a schedule next month. Add a teammate when the first one has earned its place. When you get stuck, ask the AI. That is how we got here.
Built a team using this method? Publish the story, tag Schulich, and submit it to the Centre for Career Design with your answers to four questions:
- The problem you noticed.
- The team you built and the rules you wrote for it.
- What it did for a real project.
- What you would tell the student one year behind you.
The annual follow-up I proposed in the first article will feature students, what they built, and what happened next. First team or fifth, the series grows as students add theirs.
Next in the series: The Lifestyle Business, why building a company around the life you want is a real career path, not a consolation prize.
