Over the six chapters so far, you've built the skill of having AI speed up each individual task — email, minutes, documents, data, research. The theme of this final chapter is what lies beyond: the next stage of work, from "one-off instructions" to "delegating a whole sequence of work." Ask, wait, check, ask the next thing. What if AI could arrange and carry out that repetition on its own? In this chapter, we'll introduce the "delegating" work style powered by AI agents, in a form you can picture without any specialist knowledge.
The goal: get a feel for delegating a "chain of work" to AI
From "instructing" to "delegating" — the next stage of work
How you've used it so far has been, in effect, a "one round-trip conversation." You ask, AI answers, you ask the next thing. It gets faster, but you always hold the steering wheel of the progress. The next stage is a work style where you convey only the goal and leave the intermediate steps to AI. Say "compile this list every morning" and AI, on its own, chains together a sequence of steps — gather, tidy, summarize, deliver — and runs them. The human checks the result and focuses on judgment: the image is of standing one level "above" the busywork.
What is an AI agent
An AI agent is an AI that, when you convey a "goal," thinks out the steps to reach it on its own and carries out multiple steps in order. Where ordinary chat is a one-time "question → answer," an agent runs the loop of "goal → plan → execute → check → redo" largely by itself. In human terms, it's the difference between a part-timer waiting for instructions and a staffer you can trust to arrange the work.
Ask one thing, get one answer. What to do next is instructed by you every time. It takes as many conversations as there are steps.
Hand over one goal and it breaks the steps down itself and runs them in sequence. The human moves to checking the key spots and the final judgment.
Another important trait: an agent can "use tools." Beyond just writing text, it combines external actions — opening files, aggregating tables, drafting emails, searching the web — to get closer to the goal. That's exactly why it can be entrusted with "a sequence of work." If you want the terms explained a bit more carefully, also see what is an AI agent (a primer).
💡 The line between "smart chat" and "agent." Roughly speaking, if it ends in one reply, it's chat; if it chains multiple steps and proceeds on its own, it's an agent. Lately it's increasingly common to switch between both within the same service.
The difference from RPA — the hand and the head
Hearing "automating work," some will think of RPA. RPA (Robotic Process Automation) is a mechanism that repeats a set procedure exactly, as many times as you want. An agent, on the other hand, moves while judging "what to do next" based on the situation. To put it bluntly, RPA is the "hand," and an agent is the "head." Their roles differ.
Repeats the same procedure fast and without errors. Good at set inputs and set screen operations.
Weak at: unexpected forms and ambiguous judgment. It stops when the procedure changes.
Builds and adjusts the steps to fit the situation. Strong with ambiguous instructions and exceptions.
Weak at: perfect accuracy and strict reproducibility. It sometimes gets things wrong.
The key point is that this is not "which is better" but "they're strong combined." Have the agent think through the parts that need judgment and let RPA handle the rote repetition — divide the head and the hand and you get far more stable automation than either alone. If you want to dig deeper into the difference, read the difference between AI agents and RPA.
Picturing automation at work
Abstract explanations don't quite click, so let's look at concrete examples of how you'd "delegate" in familiar work. None of these are about some special department — they're uses happening in ordinary desk work.
Classify incoming email by content and even prepare draft replies for common inquiries. A human checks before sending.
Take the weekly data, gather it, aggregate it, and sum up the key points, generating a first draft in the usual format. The human just adds interpretation.
Regularly gather and summarize new information on a set theme, and share it first thing in the morning. An extension of Chapter 6 on research.
Draft answers to common questions on the spot, and route ones needing judgment to a person. It frees up human hands.
What they share is that the more a task is "templated, high-volume, and prep-heavy," the bigger the payoff of delegating. Conversely, work like the following is a domain where the human should still be the lead.
Prep-type work where the procedure is fairly fixed, repetition is high, and mistakes are recoverable.
Decisions where mistakes can't be undone — final judgment, human negotiation, finalizing money or contracts.
📚 If you want to see more cases, the collection of AI-agent business-automation examples is a good reference. Start by finding one "templated, high-volume task" in your own workplace.
Start small with no-code
You might think "agent = programming required," but these days there are more and more tools that business users can build without writing code (no-code). Specific product names and steps change fast, so grasp this as a "map of categories" here. Rather than fixating on a specific tool, knowing the difference between the kinds pays off for longer.
A type where, on a chat service, you can build a "dedicated assistant" with a role and reference material set. The kind of feature known as a custom GPT. The most approachable.
An AI assistant built into the apps you use daily — email, documents, spreadsheets. It stays within your existing work tools, so the barrier to adoption is low.
A type where you build "when this arrives → process it this way → pass it here" by connecting parts on screen. Suited to automation spanning multiple apps. Takes a bit of learning.
For reference, ① is the kind of feature like GPTs, ② is business-suite systems like Copilot Studio, and ③ is flow-building systems like Dify or n8n — those names come up as representatives of each (the lineup changes frequently, so it's best to remember by "kind" rather than by name). Whichever you pick, the knack for success is the same.
✅ Don't delegate everything at once. First pick "one single routine task that takes 10 minutes every morning" and automate just that. Run it small, get a feel for it, then gradually widen the scope — that's the sure-fire path that doesn't fail.
Cautions when delegating
The more your "delegating" power grows, the more the design of how you delegate matters. Just as when you entrust work to a person, decide the three things — permissions, approval, and monitoring — up front. Skip this and the risk of accidents outweighs the convenience.
Give the agent only the access range it needs. If "read-only" is enough, don't grant permission to edit or send.
For irreversible actions like sending, paying, deleting, or publishing externally, set it to require human confirmation before it executes.
At first, check the results every time. Look at the activity records (logs) to confirm there's no odd behavior or looping.
The principle from Chapter 1 doesn't change. Don't carelessly hand over customer data or confidential material. Check company rules and data handling first.
⚠️ The watchword: "before delegating all of it, delegate part and verify." Don't hand over an important production task wholesale right away. First try it on a low-impact scope, watch the results, and build up trust before widening. The mindset for using it safely is on the same line as the basics covered in Chapter 1.
The next step — toward the building side
Up to here we've looked at it from the view of "using ready-made agents smartly as a business user." If you've come to feel "I want to build an agent that fits my own work more seriously," the learning of the "building side" awaits next. But there's no need to rush. First, automating one task with no-code today is the entrance to it.
"When the weekly sales data arrives, calculate the difference from the previous week and make an internal draft summing up the key changes in three lines. Keep the numbers exactly as in the source data."
This sense of "handing over the goal and the conditions to keep, together" is the first step of delegating to an agent. Once it works, widen the scope little by little.
🛠 If you feel like building one yourself. As a beginner's guide, there's how to build an AI agent (a beginner's guide). It gets into some development-leaning content, but it's a shortcut to understanding the mechanics of "delegating" from the inside.
- An agent is an AI that, given a goal, builds the steps itself and runs multiple steps in sequence. The stage beyond one-question-one-answer.
- RPA is a precise "hand," an agent is a thinking "head." Not opposites — strong combined.
- What suits it is templated, high-volume prep work. Final judgment, negotiation, and finalizing money are still human-led.
- Start small with no-code, one task at a time. When you delegate, always: minimal permissions, human approval for key actions, checking the behavior, and handling confidential data.
Well done making it through all seven chapters. Email, minutes, documents, data, research, and even "delegating" agent use — you now have the full picture of the "work skills" for finishing your daily tasks in half the time. All that's left is to try one of today's tasks together with AI. From here it's the stage of polishing as you go. For those who want to go further, we've prepared two sister courses.
For those who want a gentle, thorough refresher on how AI works and its vocabulary.
Go to the basics course →For those who want to build their own tools, apps, and agents.
Go to the indie development course →There are other courses too — see the course list →
Thank you for reading to the end. AI is a partner that frees your time from busywork and lets you focus on the work only humans can do. Start today by trying one thing you learned in the previous chapter, "Research," or in Chapter 1, on real work. Your own "AI at work" starts here.