Chapter 3

Delegate work to AI

Make your work clear enough for AI to execute

The salesperson asks AI to make a good comparison document. A generic comparison table comes back, leaving him unsure why AI failed to understand the job.

Now we begin working with AI.

The first two chapters focused on work itself: running the Goal Loop to bring the current state closer to the desired state. From this chapter onward, AI runs parts of that loop. The process stays the same; you decide which parts to delegate.

By the end, you will be able to decide which part of your work to give AI and what information to provide. We will continue with the corporate training proposal from Chapter 2.

There is little new structure to learn. You can run the loop because you know the Goal and can see the situation. Delegation means giving AI those same conditions. No difficult technical explanation is needed.

People often underestimate how much work AI can do. They stop at writing, summarizing and research. Yet AI can do the work that can be done on a computer: prepare documents, aggregate spreadsheet data, draft emails, compare sources and organize information for an approval decision. Before deciding a job is beyond AI, ask it to try. If something is unavailable, AI can explain what it cannot do.

An initial inability does not always mean the job is impossible. Asking “Use another available method” or “Complete the parts you can” often allows work to continue.

Ask questions whenever something is unclear. AI can explain a term, a tool or the purpose of a file you just created. Keep asking as you work.

Delegate the steps of the Goal Loop

The loop from Chapter 2 has five steps:

1. Compare the desired and current states to identify the main constraint.
2. Choose a ToDo that can resolve it.
3. Complete the ToDo.
4. Update the current state with the result.
5. Identify the main constraint in the new situation.

Give AI these steps one at a time. You do not need to delegate everything immediately. What changes is who does each step, rather than what the steps involve.

Step 3, completing a ToDo, can be delegated from the start. Step 4 can be included by asking AI to record the result in the Goal's current state.

Steps 1 and 2, identifying the main constraint and choosing a ToDo, require more judgment. The right division depends on the size of the Goal.

• Small Goals can be delegated. If AI can see the desired and current states, a short request to resolve the difference can cover all five steps.
• Large Goals can exceed current AI capabilities. More issues create more opportunities to choose the wrong ToDo. It is often faster for a person to decide steps 1 and 2, then give AI steps 3 onward.

For the corporate training proposal, the options look like this:

The same scopes apply to other work:

Start with one action.

Whatever the scope, AI needs the same kinds of information.

Why delegation has failed before

“I gave AI the job, but the result was unusable.” People who have that experience often study better prompts: prompt engineering, AI agents, harnesses or context engineering. Those subjects can help.

The underlying problems usually fall into two categories.

1. AI does not have the Goal and context. It does not know what you want to achieve or how far the work has progressed. Without that information, it returns generic advice.
2. The Goal is too large. It exceeds what AI can handle in one attempt. Without a useful breakdown, AI chooses its own starting point and may pursue something different from your intention.

Without a purpose, AI cannot know what work is needed. Given a Goal and the current state, today's AI can often work out how to resolve the difference, provided the Goal is not too large.

The size it can handle increases as models become more capable. A scope that failed last year may work this year. Try it again occasionally instead of treating an earlier failure as permanent.

Providing both does not guarantee success. A difference that exceeds the model's capabilities can still lead AI to take the wrong action.

With a clear Goal, the current state and a manageable scope, a short request can be enough:

Bring the current state into line with the Goal's completion criteria.

You need not assign an elaborate persona or specify every step. Prompt wording often matters less than the information available.

The salesperson in Chapter 2 had already tried a vague request:

“Make a good comparison document for the training proposal.”

The result looked like a generic comparison template, with little connection to his actual proposal. He nearly concluded that AI was not ready for real work.

Two things were missing:

• The current state: AI had seen none of the training content, prices, sales discussions or lessons from similar past proposals.
• The desired state: The request did not explain what the customer needed to decide or what the training was intended to achieve.

With neither evidence nor a clear objective, AI produced a general comparison table. It gave the most plausible response based on the information it did have.

Three conditions for effective AI work

What should you give AI? Three things.

Missing these conditions explains why AI's response can diverge from what the work needs.

Apply them to the corporate training proposal. Two iterations have been completed, and the salesperson is beginning a third.

The vague “Make a good comparison document” request asked AI to carry out the third iteration without the necessary information.

Condition 1: Provide the Goal

AI needs a Goal connected to a higher-level objective and a Definition of Done that makes completion clear.

For the third iteration, the salesperson provides:

Goal: The decision-maker has the information needed to compare options A and B and make a go/no-go decision by Friday.
Higher-level connection: Win the proposal to contribute to this quarter's corporate training order target.

The Goal specifies who needs to act, by when, and what they must be able to do. AI can check its own work against the result: “Can the decision-maker make a decision using this information?”

Before delegating, check two points:

• Is completion described in a way that can be verified?
• Can you explain the connection to the higher-level Goal in one sentence?

If writing a Goal feels difficult, use a simple format:

Who: The decision-maker
By when: Friday
Required state: Able to compare options A and B and decide whether to proceed

The format works beyond sales. In hiring: “By next week's interview, the candidate can picture what working with us would involve.” In accounting: “By month-end, every customer has received the month's invoice without any being returned for correction.” Who, by when, and in what state gives AI a basis for checking its work.

Condition 2: Provide access to context

AI now has the Goal. So far, however, it can only read the sentence you typed. The project documents and previous discussions are still unavailable.

Context is more than document files. It includes the current state, industry knowledge, working methods, higher-level Goals and anything else that informs a decision.

The principle is simple: you can make the decision because of what you have seen. Let AI see the same information.

Ask yourself, “What would I open to prepare a comparison of the two options?” The salesperson's answer is:

• Two sales meeting records, showing the decision-maker's concerns
• The instructor availability schedule, showing which option is feasible
• The draft delivery plan, showing the content and staffing

Give AI access to those sources. Avoid these common mistakes:

• ✕ Retelling everything as a summary. Summaries can omit details needed for judgment. An original meeting transcript may be more useful than rewritten minutes; AI cannot recover information removed from your summary.
• ✕ Providing only a fragment and expecting AI to infer the rest. AI cannot help with work it cannot see.
• ✕ Pasting the full contents every time. Give AI the location and let it retrieve the source.
• ✕ Adding unrelated material. More irrelevant documents make it harder to identify what matters to the current action.

○ Provide direct access to the originals. AI can establish the current state itself, applying Chapter 2's rule to use evidence.

“Keep the originals” and “Do not provide everything” are compatible. Relevance to the next action determines what to include.

There are three ways to provide access:

The first two can be used today. Chapter 4 covers tool connections.

Some context exists only in your head.

“The decision-maker cares more about the start date than the price.” “The last proposal failed because the delivery team opposed it.” Those facts may be absent from files and email. AI cannot know them unless you record them. One sentence is enough. Type it as soon as it matters. When repeating it becomes inconvenient, keep it in a note in the workspace, as Chapter 4 explains.

For every job, ask: “What would I open if I were doing the work?”

• Sales: recent meeting notes, the quotation you sent, the customer's organization chart and similar past projects
• Hiring: the job description, previous interview records, required skills and reasons earlier candidates declined offers
• Accounting: last month's close discussions, each supplier's payment terms and the same month's figures from last year

Those sources are the context to provide. Chapter 4 organizes their locations so you do not need to supply them every time. For now, be able to identify what AI needs to see.

Condition 3: Keep the Goal Loop intact

What happens after AI acts? The rule from Chapter 2 still applies: finishing a ToDo is followed by updating the work. AI's result must return to the Goal's current state for the loop to continue.

Choose the destination when you delegate. For example: “When finished, add the result and what you learned to the Acme project notes.” The destination can be a note in the project folder, a document, or the Goal management tool used in the course. It needs to be the place the next person or AI opens first to continue the work.

Without a destination, useful output can remain in an old chat without changing anyone's next decision. You get the problem from Chapter 1 at AI speed: deliverables without results.

The result of the third iteration

With all three conditions in place, the salesperson asks again: “Proceed with the next action.”

The salesperson prepares the Goal, sales meeting records and availability schedule first. AI reads the sources as it starts creating the comparison.

With the Goal and meeting records available, AI returns material the salesperson can use directly in the customer meeting. The output addresses the decision-maker's concerns.

The comparison is organized around the decision-maker's actual questions and is ready to use.

Better prompt wording was unnecessary. The Goal and context made the difference.

The first attempt had nearly convinced the salesperson that AI could not do practical work. The problem was what AI had been given.

If the output is poor, check the three conditions: Was the Goal clear? Was enough relevant context available? Was there a destination for the result? Fix the missing condition and try again. A disappointing result can still improve your understanding of the current state.

As the environment improves, requests can become shorter and less formal. “Please prepare a good comparison” can produce useful material when AI already has the Goal and context. The original request failed because the information was missing. Spending hours refining the wording was unnecessary.

Delegate the largest Goal AI can complete reliably

With the Goal and current state available, a request to resolve the difference often suffices. Elaborate wording and personas add little.

The decision that remains is how large a Goal to delegate.

• ✕ Increase the business's revenue. Too large; the response is likely to be generic.
• ✕ Change the ending of one sentence. Too small; you remain responsible for issuing instructions continuously.
• ○ Use the meeting records to identify why five stalled sales opportunities are waiting. A single attempt can return an assessable result.

When the scope works, try a larger one. When it fails, reduce it. The practical limit reflects the model's capability, so the Goals you can delegate grow as models improve.

Expand delegation while checking the decisions

Start with one action, then a child Goal, and then a larger scope after successful attempts.

With current models, decisions about what to do—defining the Goal, choosing the main constraint and narrowing ToDos—are often faster when you think them through with AI.

AI does not automatically know your experience, the history of the work or what you want to achieve next. It generates a plausible response from the information available.

AI may know a procedure in detail without knowing whether that procedure serves your purpose. It can also agree with your stated preference. Agreement alone does not establish that a decision is right, even when the response sounds convincing.

Unrestricted delegation can create unnecessary ToDos and disconnect work from the Goal. Chapter 2's warning about starting from ToDos applies at AI speed. Keep responsibility for the direction, execute with AI, and expand the scope within that division.

The division is expected to change as models improve. Given a Goal and context, more capable models should be able to make more of the decisions about what to do. The quality of the Goal and current state you provide will still matter. The preparation remains useful.

Review

• Delegation means letting AI run parts of the Goal Loop you already use. The structure stays the same.
• Choose the scope: one action, a child Goal, then a larger Goal.
• Retain responsibility for deciding what work should be done and why.
• Provide three conditions: a verifiable Goal, access to context, and a destination for the result.
• Let AI read the same relevant originals you use.
• Before replacing the model, check what information it lacks.
• AI can do work performed on a computer. Try a request before assuming it cannot help.
• Focus on the size of the Goal, and delegate the largest scope AI can handle in one attempt.

Continue to the next chapter

You can now decide which steps to delegate and what AI needs to carry them out.

Collecting the Goal and documents for every request is still time-consuming. The salesperson has been doing that each time. The next chapter makes the information available from the start, so repeated preparation becomes unnecessary.

• Chapter 4: Create an environment where AI can work