How to Build AI Assistants and Agents: From Custom GPTs to Autonomous Workflows (Lesson 5)

Most people use AI to save a few minutes. The teams pulling ahead are using it to eliminate entire categories of work — not just speed them up.

There’s a name for the gap between those two groups: it’s the gap between prompting AI and building AI. Lesson 5 is about crossing it.

In this session, Ian Arden walks through how to build AI assistants that work inside your business systems — and how to take that one step further into autonomous agents that take action, send messages, and execute workflows on their own, without anyone pressing a button.

What Is a Custom GPT?

A Custom GPT is a pre-configured AI assistant you build once and use repeatedly. Instead of writing a new prompt every time you need something done, you encode your rules, context, and output requirements into a system instruction — and the assistant follows them automatically, every session.

Here’s the simplest way to think about it: a system instruction is a job description for your AI. Just like you’d brief a new hire on their role, responsibilities, and how you expect them to communicate, you brief your AI assistant the same way. The more specific and complete that briefing, the more reliably it performs.

OpenAI’s ChatGPT calls them Custom GPTs. Google’s Gemini calls them Gems. Microsoft Copilot has its own version. The name varies; the idea is identical across all platforms.

Step 1: Personalise Your AI

Before building assistants for your business, it’s worth setting up ChatGPT’s personalisation features for yourself. This is the foundation everything else builds on.

In the personalisation settings, you can add a custom instruction that tells the AI who you are, how you think, what you’re working toward, and how you want it to communicate. Ian’s own instruction, shown in the lesson, tells ChatGPT to think like a co-founder rather than an assistant, adapt to his fast-moving working style, and filter its knowledge toward multi-billion dollar tech and AI — because that’s his world.

The point isn’t to be fancy. It’s to stop explaining yourself every time you open a new conversation.

• Include your role and goals. This helps AI filter its knowledge base toward what’s actually useful to you, rather than giving generic answers.
• Describe how you like to work. Do you want short, direct answers? Structured sections? Plain language? Say it once here.
• Be honest about your constraints. If you move fast, switch topics, or want to be challenged on your thinking — tell it. AI adapts to what you give it.

Step 2: Build Your First Custom GPT

Once you understand what a system instruction is, building a Custom GPT is straightforward. You’re essentially writing a detailed brief for a new team member who never forgets, never gets tired, and works at the speed you set.

In the lesson, Ian builds one live: an AI Editorial Analyst that searches for recent AI industry news and produces branded editorial content for a website, social media, and other channels. The whole build takes minutes — ChatGPT’s conversational interface walks you through it.

A good system instruction is not a paragraph. It’s a document — often 100 to 200 lines. That length is fine. The more specific you are upfront, the less you have to correct later.

Step 3: The Real Shift — Assistants vs. Agents

Here’s the distinction most people don’t get told.

An AI assistant waits for you. You open it, give it a task, it does the work, you review the output. That’s useful. But you’re still in the loop. You still have to remember to use it.

An AI agent acts on its own. You configure it once. You define its triggers — a schedule, an incoming email, a change in a spreadsheet — and it runs automatically, without you. It reads data, makes decisions, takes actions, and reports back. You find out what it did, not what it needs you to do.

That’s the real shift: from using AI to deploying it.

A Real Agent in Action: Microsoft Copilot Studio

In the lesson, Ian demonstrates a working AI agent built in Microsoft Copilot Studio — a platform he describes as “extremely sophisticated and pretty well developed for the enterprise environment” that most teams aren’t paying enough attention to.

The agent is a task execution monitor. Here’s what it does, entirely on its own:

• Reads a project task spreadsheet to identify what each team member is responsible for and when things are due.
• Identifies overdue tasks and tasks approaching their deadline based on today’s date.
• Drafts a personalised follow-up email for each person — including AI-generated recommendations on how to complete their specific task successfully.
• Sends the emails to the relevant team members automatically.
• Runs daily on a schedule — no human trigger, no button to press, no one needs to remember to run it.

This isn’t a concept or a mockup. It ran live in the demo. The email it produced was well-written, contextually relevant, and included genuinely useful task guidance. All generated automatically.

What makes Microsoft Copilot Studio particularly powerful for this kind of work is its tool connectivity — the ability to read from and write to your actual business software (spreadsheets, email, task managers, CRMs) — combined with built-in scheduling triggers, authentication policies, and an execution log that lets you trace exactly what the agent did on each run.

What This Looks Like Inside a Real Business

ADAIA built a Custom GPT eight months ago that handles one specific job: turning client meeting transcripts into full proposals.

When someone at ADAIA finishes a discovery call, an AI records and transcribes it. That transcript gets pasted into the assistant. The assistant — which knows ADAIA’s services, pricing, proposal structure, past examples, and how to frame ROI — produces a ready-to-send proposal document. No additional prompting. No back-and-forth.

The team feeds it notes. The AI produces the output. That’s the whole process.

This is the kind of delegation that changes how a team operates. Not saving 10 minutes — removing an entire step from a workflow.

• Identify the task first. The best candidates for Custom GPTs are processes where you receive one type of input and always need to produce the same type of output.
• Write the system instruction like an SOP. Cover every exception, format requirement, and business rule. The more detail, the less you have to supervise.
• Upload your documents. Proposals, templates, guidelines, past examples — all of this becomes the assistant’s knowledge base.
• Enable the right capabilities. Web search if it needs current information. Code interpreter if it processes data. Actions if it needs to write to external tools.

Frequently Asked Questions