Chapter 5
Automate work and let AI manage decisions
Keep the Goal Loop running without starting every action yourself
On the commute, the salesperson realizes that he forgot to send yesterday's request to AI. The work never started.
Chapter 4 prepared an environment where a short request can start useful work. One problem remains: on days when the salesperson says nothing, AI does nothing. Work stops during a holiday or after a forgotten request. This chapter addresses that dependence.
Without this step, the gains remain limited to efficiency. A document may take ten minutes instead of an hour, but requesting it, correcting the result and asking again can still take 20 or 30 minutes. Work still depends on you remembering to start it.
The aim is to have yesterday's email read and the relevant calendar entries prepared before you wake up, without needing to direct attention to the process every morning.
In the five-step Goal Loop, the change is the trigger that starts the work. Steps 1 through 5 remain the same. A time or event starts the loop instead of a person saying “Please do it.”
Automation and autonomous work
Automation: At the end of each month, invoices are prepared and sent to fixed recipients without your involvement. The procedure repeats, so it can be specified in advance.
Autonomous work: On Monday morning, a list shows stalled projects, why they are stalled and the next action for each. Different projects appear each week because the situation changes what needs to be done.
The distinction comes from the nature of the work, rather than how familiar you are with it.
Use automation when you can specify a fixed procedure. Use autonomous work when each situation requires a new decision.
Automate routine work first
Begin with repeated procedures. As Chapter 3 explained, AI can do work performed on a computer. When you notice that you are repeating the same action, make it a candidate for automation.
Monthly invoices, weekly sales reports, and checking and filing received contracts all have procedures you can describe. Fixed actions are also easier to diagnose and correct when something fails. Start with that work.
Before automating, check the Goal. Which difference between a Goal and its current state does the routine resolve? If you cannot answer, consider stopping the work first. Reports nobody reads and activities that exist only out of habit can otherwise continue automatically every week. Remove unnecessary work before automating the remainder.
Use autonomous work when the next action depends on the situation
Some work cannot be reduced to a fixed procedure. Following up on stalled sales opportunities, responding to applicants and reminding people about missing submissions all change with the current state. The recipient, message and decision to send may differ each time. AI must read the Goal and current state and choose the next action.
More difficult work often belongs in this category. An article or proposal cannot be fully specified by an unchanging sequence. What you should write depends on recent publications, current responses and which topics have enough supporting material. Give AI the information needed to decide: the publishing Goal, previous work and its results, and writing rules. Then delegate choosing the topic as well as drafting it.
| Work | Approach | Reason |
|---|---|---|
| Monthly invoice generation | Automation | The targets and procedure are fixed |
| Weekly sales aggregation and reporting | Automation | Sources and recipients are fixed |
| Checking receipt of contracts and filing them | Automation | Received documents go to a defined location |
| Following up on stalled sales opportunities | Autonomous work | The customer's reply changes the message and whether contact is needed |
| Responding to applicants | Autonomous work | Questions and guidance differ by applicant |
| Reminding people about missing submissions | Autonomous work | The set of people who have not submitted changes each time |
| Planning and drafting articles | Autonomous work | Recent publications and responses change what should be written |
Work toward handling all recurring work through these approaches. Busy weeks can reduce the consistency of manual work. A configured AI process can produce consistent output across repetitions, making repeated work particularly valuable to delegate.
A single job can contain both approaches. Creating and sending invoices is automation; deciding how to contact a late payer requires situational judgment. Keep fixed steps automated and use autonomous decisions only where the action changes.
For the salesperson, AI starts at 9 a.m., reviews the Goals for his projects, checks their current states, advances work it can handle, and reports reasons when it cannot proceed. Follow-up wording depends on each project, so the process involves autonomous work.
Choosing to wait is also a valid action. One morning, AI notes that the customer still has two business days before the agreed reply deadline and decides against additional contact. After the deadline, it sends one follow-up. Running every morning does not mean sending a message every morning. The current state determines the action.
At Addness, a process checks completion of an internal assessment three times a week. A broadcast reminder would be automation. This process checks who has not completed the assessment and contacts only those people. The recipients change each run, so it requires autonomous decisions. When everyone has completed the assessment, no difference remains and the loop stops.
The salesperson asks AI to check the projects at 9 a.m. each day and proceed, while stopping to ask when something is unclear. Work can then continue without another request from him.
After setting AI to begin at 9 a.m., the salesperson arrives to find work already progressing. A colleague looks at the screen in surprise.
Start AI with a time or an event
AI needs a trigger to start without a person making a request. There are two types.
- Scheduled execution: review projects at 9 a.m. daily, review last week's sales meetings on Monday, or prepare invoices at month-end.
- Event triggers: start when a customer email arrives, a quotation is approved, a form is submitted or a new file is added to a folder.
Scheduled execution
Review projects at 9 a.m., review the week on Monday, or prepare invoices at month-end.
Event trigger
An email arrives, approval is granted, or a form is submitted.
The trigger changes. AI still runs the same Goal Loop.
Most recurring work fits either “at a particular time” or “when something happens.”
The trigger changes; the Goal Loop stays the same. AI started at 9 a.m. still reads the Goal and current state, identifies the main constraint, takes an action and records the result.
Run the Goal Loop automatically, including decisions and result updates.
A loop that reassesses the work can adapt in ways a fixed action cannot.
Chapter 3's limitation still applies: current models can misjudge the main constraint in a large Goal. Schedule only Goals small enough for AI to handle in one iteration. An oversized Goal can otherwise produce the same mistaken assessment every morning.
Recording results is equally essential. Without it, AI starts each day unaware of its previous actions and creates the same proposals or drafts again. Updated context allows the work to progress.
Keep people involved while the process improves
Do not begin by deciding to remove people from the work. Where people participate should change with the reliability of the loop. The salesperson develops it in stages.
The salesperson initially reviews and corrects every follow-up draft. As he records corrections in the working environment, the process improves until only an exceptional case needs his attention.
Human review helps because a reliable loop cannot be designed perfectly on the first attempt. Inspecting actual output and correcting it is the fastest way to improve.
There are three stages, with a clear signal for moving forward.
| Stage | Human involvement | Signal to proceed |
|---|---|---|
| 1. Review everything | Check and correct every output | Three consecutive outputs need no correction |
| 2. Review exceptions | Check cases involving money or a new recipient; AI handles other sending within the agreed scope | AI correctly distinguishes the exceptions |
| 3. Complete within the agreed scope | Receive reports while the defined work finishes without individual review | Move the reliable process to cloud execution |
The improvement comes from recording corrections. If the wording is wrong for your company, save the correction in the environment from Chapter 4. Do the same for report structure: remove unnecessary fields or expand coverage when needed. Reusing corrections reduces repeated human intervention. Without that record, you keep making the same edits.
Even at the third stage, review a week's output together at least once a week. Each daily action can look reasonable while the overall pattern moves away from the objective. A week of project reviews may reveal that easy opportunities get attention while the largest deals remain stalled. A single run cannot reveal every pattern.
Ask AI to set up the process
You can start by saying which work you want to automate.
As you run the Goal Loop, notice ToDos that recur in the same form. Ask AI to set them up:
Automate the sales report we prepare every Monday.
Automate the initial handling of inquiry emails.
You do not need to memorize every settings screen. Explain the intended behavior and ask AI to configure it. AI can ask for missing details, such as the execution time, delegated scope and response to uncertainty.
Follow one sequence: run the process on your own computer first. Early outputs will need adjustment. Local execution lets you inspect, correct and try again immediately. Move to a form that runs while your computer is closed, such as cloud execution, once the process no longer needs that close review.
Starting unattended too early can allow incorrect work to continue around the clock.
Use the supplementary lessons for implementation details
This chapter explains what to automate and in what order. The supplementary lessons cover practical setup.
- Supplementary lesson 1: A complete guide to Claude Code — Video demonstrations of scheduled execution and moving from local to cloud execution.
- Supplementary lesson 3: Examples of automation and autonomous work — Implemented examples, including the information and requests used. Further examples are added on note.
Open those lessons when setup becomes unclear.
Define when the loop should stop
An autonomous loop needs completion and stopping conditions. Define three outcomes:
- Continue: a difference remains and work can proceed.
- Complete: the Goal is achieved; mark it complete and report the result.
- Stop: information is missing, work exceeds the agreed scope, or an unexpected situation occurs. Stop and return the decision to a person.
The third condition matters: “When uncertain, stop and ask before proceeding.” It makes the delegated scope explicit.
Incorrect work continuing around the clock is a consequence of skipping the structure from Chapter 2. Define the Goal, establish the current state from evidence and specify when to stop. That sequence makes autonomous work useful.
Review
- Start AI through scheduled execution or an event trigger.
- Describe the work you want automated; ask AI to configure it.
- The trigger changes. The five steps of the Goal Loop remain the same.
- Start with fixed procedures, then add work where the current state determines the next action.
- Before automation, verify which Goal the recurring work advances. If it advances none, consider ending it.
- Adjust human involvement as reliability improves. Save corrections in the environment.
- Define stopping conditions first. Missing information, scope limits and unexpected situations return to a person.
Continue to the next chapter
Routine work can now run automatically, and AI can run a Goal Loop autonomously. Chapters 2 through 5 covered work within one person's scope.
The next chapter extends the approach to a team, addressing why people in the same company can have AI agents with different knowledge.