Chapter 1. Work analysis and automation design

Learning objectives

• Divide a workflow into steps and write down each step's input, output, owner and completion condition.
• Compare automation candidates by repetitiveness, risk and review cost, applying exclusion criteria first.
• Distinguish steps that rules can handle from those requiring AI judgment.
• Place human review points before hard-to-reverse actions, externally visible actions and consequential decisions about individuals.
• Write a design sheet that starts with the simplest delegation level and expands its scope based on evidence.

Introduction

Suppose a company's general affairs manager is told to “automate supplier management with AI.” The manager immediately starts looking for a tool to buy. A few weeks later, the tool is in place, but nobody uses it. Nobody has written down the actual steps of supplier management, which steps repeat, which are risky, or who must check what. A tool cannot accept a broad task called “supplier management.” It can accept a specific task such as “extract business details and the contact person's information from a supplier's email and enter them in the registration form.” Failure in the opposite direction is also common: an enthusiastic employee delegates everything from checking registration documents to notifying suppliers that trading can begin, and in the first week a notification goes to a supplier with incomplete documents. Both failures have the same cause: starting without deciding which steps to delegate and how far.

This chapter explains how to start automation with the work itself. First, break the work into steps and write the inputs and outputs. Then use three criteria—repetitiveness, risk and review cost—to decide which steps to delegate first. Next, identify where a person must check the work, and learn to start with the simplest level of delegation and expand as evidence accumulates. Finally, record these decisions in a one-page design sheet. You will continue using it throughout Chapters 2–5.

Key concepts

Breaking work into steps

Before cooking, a chef prepares the ingredients and lays them out in order. Automation works the same way. Break the work into steps and record five things for each.

Always record the completion condition. “Summarize” has no clear finish, but “all three fields are filled” can be checked. The same condition becomes the review criterion for a step delegated to AI. Since one step's output is the next step's input, the table also reveals how missing information early on can cause problems later.

Triggers, completion conditions and handoffs

Work begins with an event: an email arrives, the month ends, or an application is submitted. This is the trigger. Recording it helps determine whether to run the work on a schedule, respond to an event such as an incoming email, or let a person request it each time.

When drawing the step table, mark the handoffs, where the owner changes. Much of the total workflow duration is spent waiting for the next person to start (waiting time), rather than doing the work (processing time). Even if preparing a quotation takes 20 minutes, a two-day approval delay makes the customer wait two days. Cutting preparation to ten minutes with AI barely changes the customer's experience. Recording processing and waiting time separately distinguishes the time AI can reduce from the time it cannot.

Normal and exception paths

Automation designed only for average cases often stops in practice. Draw both the normal path, taken by most cases, and exception paths, taken by cases with missing information, conflicting materials or missed deadlines.

1. Collect cases that departed from the normal path from past records.
2. Group them by type and record each type's frequency and impact.
3. Decide how to handle each type: request more information, hand it to a person, or put it on hold.
4. Record the signals AI can use to recognize exceptions, such as an empty required field, an amount above a threshold, or words such as “urgent.”

The clearer the criteria for handing exceptions to people, the more sustainable the automation. Forcing exceptions into automatic processing complicates even the normal path.

Three criteria: repetitiveness, risk and review cost

Evaluate each step on three criteria to establish automation priorities.

Do not simply add the three scores: a very high repetitiveness score can hide serious risk. Use this order.

1. Apply exclusion criteria first. Exclude automatic processing of decisions affecting individuals' rights without human review, and work requiring sensitive material to be entered into unapproved accounts, regardless of scores.
2. Compare only the remaining candidates. Top-priority candidates repeat dozens of times a day, are easy to fix when wrong, and take a minute to check.
3. Compare review cost with the original work time. If checking takes as long as doing the work originally, automation saves almost no time.

Also remember that work done once a quarter rarely pays back the time spent designing and testing automation.

Breaking risk down

One “risk” field makes judgment too coarse. The same error may be easily corrected in an internal draft but impossible to recall once sent to a customer. Breaking risk into three components gives a more accurate assessment.

A step with all three components low may be a candidate for automatic execution. If even one is high, it needs human approval or rigorous review. Voluntary guidance such as the US National Institute of Standards and Technology's AI Risk Management Framework (NIST AI RMF) helps organizations manage AI risks systematically and can inform company-wide criteria.

Work for rules and work requiring AI judgment

Not every step needs generative AI. Calculations using fixed formulas, or branches such as “add an approval level if the amount exceeds KRW 500,000,” are more accurate and easier to explain when handled by spreadsheet formulas, email filters or business-system rules. Generative AI is probabilistic: even with the same input, its wording can vary.

Good designs combine both. AI extracts information from unstructured text into predefined fields, and rules handle subsequent calculations and branching. When rules serve as a gate that checks AI results, errors are caught before reaching the next step.

The delegation ladder

“Automate” is too broad a term. Dividing delegation into a four-level ladder makes the scope concrete.

Anthropic recommends finding the simplest possible solution when building AI systems and adding complexity only when needed. Start at the bottom of the ladder too. Beginning at Level 4 makes it hard to distinguish instruction problems, material problems and execution-connection problems when something fails. Set numeric advancement criteria before starting, so that later you do not move up merely because things “seem to work.”

Where to place human review points

Human checks go at points of high impact, rather than at every step. Anthropic gives examples of actions requiring confirmation: destructive operations such as deletion, hard-to-reverse operations, and operations visible to others such as sending messages or changing shared systems. For work automation, add one more.

1. Destructive actions: deleting files or records, overwriting data.
2. Hard-to-reverse actions: transfers, executing refunds, confirming orders.
3. Externally visible actions: sending customer emails, public posts, sharing with business partners.
4. Consequential decisions about individuals: decisions affecting people's rights and opportunities, such as hiring, credit and insurance.

Distinguish approval from verification here. Approval means “you may perform this action”; verification means “I have checked that the content is correct.” Pressing an approval button does not mean the content has been checked. Chapter 3 covers what an approval screen must show to enable verification as well.

Designing a reviewable workload

Adding an approval button and making real human review possible are different problems. Once you choose review points, calculate and prepare the following.

1. Review time per case and daily volume: at 200 cases a day and two minutes each, review takes 400 minutes, about 6.7 hours daily. Consider whether one reviewer can do that alongside other work.
2. Evidence for the reviewer: rather than requiring a complete rereading of the source, have AI flag key figures, evidence locations and uncertain points alongside the result.
3. Cover for absences: decide who reviews during leave and whether execution stops if nobody is available.
4. Behavior while review is pending: prevent results from being sent automatically while waiting for review.

If the workload exceeds human capacity, keep the review point and narrow the automation to a reviewable number of cases, or use sample review only for low-risk steps to concentrate review time on high-risk steps.

Translating work requirements into feature combinations

Once the step table is complete, choose the feature type needed for each step. Thinking in feature types before product names keeps the design useful even when tools change.

One workflow usually combines several forms. Remember, however, that every additional form also adds connection permissions, sharing scope and review points. Removing unnecessary forms is part of design.

What not to automate

OpenAI's usage policies prohibit automating important decisions without human review in sensitive areas such as employment, finance and credit, insurance, housing, education, law and healthcare. Anthropic's Usage Policy also requires qualified human review for advice or decisions affecting individuals in such high-risk areas, and requires disclosure of AI use when model output is shown directly to individuals. Automation that requires customer or applicant information to be entered into personal accounts the company has not approved violates data boundaries before it is even an automation question.

This does not mean never touching such work. Record why it cannot be delegated under current conditions, and look for ways to change those conditions. For example, leave decisions to people and use AI only to organize materials, or automate selected steps after establishing a company workspace.

The one-page automation design sheet

Record the decisions so far in a one-page design sheet. An incoming task owner should be able to understand the purpose and boundaries.

[Automation design sheet]
Task / Owner / Date
1. Purpose: what to reduce and by how much (include baseline figures)
2. Trigger and inputs
3. Step table: input, processing, output, owner and completion condition for each step
4. Delegation level: ladder level for each step
5. Materials and permitted tools (systems to connect, actions to enable)
6. Human review points: location, checks, reviewer, cover for absences
7. Excluded work and reasons
8. Stop conditions: what should stop automatic execution
9. Success criteria: numbers required to move to the next level

Going deeper

Workflow or agent?

Anthropic distinguishes two kinds of AI systems. Workflows move AI and tools along predefined paths; agents let AI determine its own process and tool use. Complex structures such as agents generally gain performance at the expense of latency and cost, so Anthropic recommends them only when simpler methods fall short. It also explains that work with clearly separated steps suits a workflow that chains prompts with checkpoints between them.

For no-code work automation, the criterion is this: if you can know the number and order of steps in advance, design it as a workflow. For a fixed sequence such as “classify email → look up information → draft reply,” running each step separately and checking intermediate results makes errors easier to locate. Research or exploration where the number of steps is unknown may suit an agent better. Agent errors can accumulate across steps, however, so design stop conditions and human intervention checkpoints too.

Traps in scoring tables

A table scoring repetitiveness, risk and review cost from 1 to 5 and adding them up can be persuasive in a team meeting. But it has three traps.

1. Offsetting effects: a repetitiveness score of 5 offsets a risk score of 5, pushing a dangerous step up the ranking. This is why exclusions come first.
2. Differences between raters: two people score the same step differently. Define each score, such as “3 means 10–30 cases daily,” to reduce variation.
3. Scoring whole tasks instead of steps: scoring “customer support” as a whole averages classification (low risk) with compensation payments (high risk). Always score individual steps.

A scoring table does not make the decision for you. Use it to structure the discussion, and write the final judgment and its reasons in the design sheet.

Rubber-stamp review

Errors sometimes pass unchanged despite a human review point. As reviewers experience mostly correct AI output, they become less attentive and eventually merely approve it. This book calls that formal approval without substantive checking “rubber-stamp review.” Causes and responses follow.

Another method is to periodically mix in outputs with deliberately inserted errors to see whether reviewers detect them. Evaluate review by its ability to catch errors, rather than merely by the presence of a person.

At work

Sunny Pine Facilities: organizing repair requests

The control room at Sunny Pine Facilities, which manages twelve office buildings, receives about 150 repair requests daily. Tenant contacts submit requests freely by email, messenger and telephone notes. Coordinator Do Eunseo was tasked with automating the organization and assignment of requests to field technicians.

Do first split the work into five steps: ① receive the request, ② organize location, symptoms and urgency, ③ assign a technician, ④ send an acknowledgment of receipt to the tenant, and ⑤ record completion. Measuring processing and waiting times showed that ② took longest to process at four minutes per case, while ③ had an average two-hour wait because the control room manager made assignments in one morning batch.

Before applying the three criteria, Do checked exclusions. Safety-related requests involving a gas smell, electrical leakage or people trapped in elevators already had a separate procedure for immediate telephone dispatch. Do excluded these from automation, adding only an exception path: when AI found such expressions, it would stop organizing the request and alert the control room by phone.

Of the remaining steps, ② was repetitive, errors could be corrected by technicians on site, and review took 30 seconds. Do chose it first. Assignment in ③ involved judgment about technicians' skills and travel distances, so people would continue doing it, using the AI-organized table. Acknowledgment in ④ was visible to tenants, so it began at Level 2, draft for review.

One colleague suggested that automatically texting technicians would eliminate the two-hour wait. Do did not accept this immediately. The wait resulted from assignment decisions being concentrated in the morning; automated SMS notifications would not help if assignments were still delayed. Instead, Do also proposed changing the procedure so the manager assigned work twice daily, morning and afternoon.

For requests missing a location, AI would draft a request for more information, which Do would check and send. Success criteria were “at least 80% of summary tables used without edits, and no more than two incorrect assignments per week.” The design sheet ended with a stop condition: “If even one case of missed safety-related wording occurs, stop automatic organization and investigate the cause.”

Word Loft: first review of manuscript submissions

The editorial department of humanities publisher Word Loft receives nearly 200 manuscripts monthly. Three editors open submission emails, skim manuscripts, decide whether to consider them for publication, and send rejection replies. Editor-in-chief Seo Haneul said, “It would be helpful if AI could read manuscripts, choose which to reject, and send the replies.”

While breaking down the work, editor Noh Jihoo realized that the request combined three different tasks: organizing a manuscript's subject, length and structure; judging whether to consider publication; and replying to the author. The first was repetitive and errors could be corrected immediately by an editor. The second involved the publisher's editorial judgment, and the third was externally visible.

Noh compared three options: (A) AI makes decisions and sends replies; (B) AI only organizes and recommends, while editors decide and reply; and (C) AI organizes and recommends, then drafts replies after editors decide. Noh rejected (A) because a mistaken rejection could not be reversed and a relationship with a rejected author could be difficult to repair. Option (B) was safe but left the time spent writing replies unchanged.

Noh chose (C). For each manuscript, AI produces a one-page summary of subject, estimated length, table-of-contents structure, and fit or mismatch with the publisher's editorial direction, marking “recommend review” or “recommend hold” with reasons. After an editor decides, AI drafts the reply; the editor reads and sends it. Since manuscripts are the authors' copyrighted work, processing would take place only in the company's contracted business workspace.

Noh also calculated review volume: 200 manuscripts monthly at three minutes per summary meant ten hours a month, manageable across three editors. Before starting, Noh recorded the baseline: about 70 hours of combined editorial time handling submissions the previous month, and a 19-day average from submission to reply. Two months later, processing took about 40 hours; as editors gained time, the unread backlog shrank and average response time fell to eight days. Noh's report began, “AI does not judge manuscripts. It organizes what editors need to judge them.”

Common mistakes

1. “Automating important work first brings the greatest benefits.” Important work also carries greater error impact. Start with frequent, low-risk work that is easy to review.
2. “Adding the three scores gives the priority order.” Adding lets repetitiveness hide risk. Apply exclusions first, then compare the remaining candidates.
3. “Having AI handle calculations too solves everything in one step.” Formulas or business systems are more accurate for fixed calculations. Use AI to organize unstructured input.
4. “More checks mean more safety.” Approval at every step creates approval fatigue and superficial checking. Focus on points of high impact.
5. “An approval button guarantees review.” Approval differs from verifying content. Reviewers need time and evidence to check it.
6. “High pilot accuracy means approvals can be removed.” Even with high accuracy, one irreversible error can cause serious harm. Set delegation levels using both accuracy and difficulty of reversal.
7. “You must choose the tool before designing.” A step table and success criteria must come first to provide a basis for comparing tools.

Summary at a glance

Check the official docs

• Anthropic — Building effective agents: Read the distinction between workflows and agents and the principle of starting with the simplest solution.
• Anthropic — Prompting best practices: Check examples of actions requiring confirmation in “Balancing autonomy and safety.”
• Anthropic — Define success criteria and build evaluations: Read how to make success criteria specific and measurable.
• OpenAI — Usage policies: Check the prohibition on automating important decisions in sensitive areas without human review.
• Anthropic — Usage Policy: Read the human review and AI disclosure requirements for high-risk use cases.
• NIST — AI Risk Management Framework: Check the voluntary organizational AI risk framework and its generative AI profile.

Self-check

1-1. This chapter recommends breaking risk into three components rather than assigning a single score. Name them and give an example of each being high.

1-2. Explain why processing time and waiting time should be recorded separately in the step table. State which AI can reduce.

1-3. For internal mail reception, decide which step should use rules and which should use AI: “sort by the department name on the envelope” and “determine whether the enclosed letter is a complaint or a simple inquiry.” Explain why.

1-4. How do approval and verification differ? In one or two sentences, explain why approval without verification is dangerous.

1-5. A team plans human review of 90 outputs daily. Each review takes four minutes, but the reviewer has only three hours a day. Calculate the problem with this plan and suggest one adjustment to review or automation scope.

1-6. Translate this requirement into the feature forms in this chapter (knowledge space, custom assistant, connector, agent, scheduled task): “Every Monday morning, gather last week's meeting minutes from the team's shared drive, summarize only the decisions, and show them to the team lead.”

1-7. Choose two required items in the one-page automation design sheet that are particularly important during an ownership handover. Explain why.

Hands-on tasks