Decide What to Automate and What to Keep Manual

Choose the process that meets the need with an amount of complexity you can justify and maintain.

A learning-center coordinator spends Friday afternoon preparing next week's activities. Some tasks are repetitive: count registrations, copy confirmed dates, and assemble a draft schedule. Others require judgment: decide whether a struggling student needs a different class, or resolve two instructors' conflicting commitments.

It is tempting to hand the whole afternoon to an agent. Before doing that, separate the decisions. Repetition, language, ambiguity, and consequence do not always appear in the same places.

The right automation may be a spreadsheet formula, a rule, a short script, a model-assisted draft, or a bounded agent. Sometimes the best improvement is a clearer checklist for a person.

Observe the current process

Watch a few real instances or reconstruct them from authorized records. Identify the trigger, inputs, decisions, outputs, exceptions, and owner. Record where people wait, where information is missing, and where errors require rework.

For the center's weekly schedule, the trigger is a cutoff time. Inputs include approved activities, room availability, and confirmed instructors. The output is a draft schedule. Exceptions include missing confirmation, a room conflict, or an activity that exceeds capacity.

Now ask which facts are authoritative. A casual email proposing a date is different from a confirmed event record. Automating the movement of uncertain information can make uncertainty travel faster without resolving it.

The process map should therefore identify both work and evidence. “Copy the date” is incomplete until you specify where the confirmed date comes from.

Compare the available approaches

A rule performs a predefined action when a condition holds. A script carries out a defined sequence of operations. A model-assisted step uses generation or interpretation within a larger process. A workflow fixes the main path while allowing branches. An agent gives the model some control over which permitted steps to take next.

These categories overlap in real products. The useful question is who controls the next action and how the result is checked. Anthropic's guide to effective agents distinguishes predefined workflows from systems in which a model dynamically directs tool use, and recommends adding complexity when the task warrants it.

For the center, a comparison might look like this:

The table is a starting hypothesis, not a permanent classification. If the archive has only six predictable questions, a predefined retrieval workflow may be sufficient.

Assess variability and consequence separately

A task can be highly variable but low consequence, such as generating fictional writing prompts. It can be predictable but consequential, such as updating an authoritative bank-account field. Variability tells you something about the need for flexible interpretation. Consequence tells you something about verification and control.

Ask what happens when the output is wrong. Is the result easy to inspect before use? Can the action be reversed? Would a mistake affect one disposable draft or many people's records? How quickly would anyone notice?

A draft can often tolerate more experimentation than an external commitment. That does not mean drafts need no review: a fabricated date may become a commitment if someone copies it into a message without checking.

Place the review where it can still prevent the consequence. Reviewing a report after automatic distribution cannot prevent the original disclosure.

Calculate total effort, including the work automation creates

Suppose a fictional center processes 200 routine entries each month. Manual handling takes six minutes per entry, or 1,200 minutes. A proposed workflow needs 12 hours of initial setup, then one minute of review per entry, plus 60 minutes of monthly maintenance and 40 minutes of exception handling.

The recurring effort would be 300 minutes per month: 200 for review, 60 for maintenance, and 40 for exceptions. Compared with 1,200 manual minutes, the estimated saving is 900 minutes, or 15 hours per month.

At that steady workload, a 12-hour setup would be recovered in 0.8 months of time savings. That calculation excludes software fees, model charges, training, and any quality difference. It is a fictional planning example, not a measured result.

Now reduce the volume to 20 entries. Manual work is 120 minutes. The same assumed maintenance and exception allowance plus 20 minutes of review also totals 120 minutes. The setup effort no longer pays back under those assumptions.

The lesson is to measure the actual workload and recurring costs. A dramatic improvement per run can be economically unimportant when the task rarely occurs.

Find the smallest useful pilot

Choose a representative, bounded task and define success before building. For the schedule workflow, a pilot might prepare drafts for ten synthetic activities, identify every missing confirmation, and create no external event.

Compare it with the current process on the same inputs. Measure factual errors, missed exceptions, review time, and total time to an accepted result. Include failed runs in the accounting.

A workflow that produces a draft in three seconds but requires ten minutes of correction may be slower than the manual method. A slightly slower draft that reliably flags missing evidence may be more useful.

Set a stop rule for the pilot. If the workflow cannot distinguish proposed dates from confirmed dates after a defined revision cycle, reconsider the task design rather than adding another agent by reflex.

Keep a person where responsibility belongs

Some decisions depend on relationships, values, or information that is not appropriately captured in a system. A coordinator may need to speak with an instructor before changing a commitment. A student may need encouragement and a chance to explain a difficulty.

AI can help prepare questions, summarize approved records, or draft options. The person still owns the judgment and the relationship. The goal is to remove needless clerical work while preserving the interaction that makes the service useful.

This principle applies across fields. A researcher can automate file preparation while retaining decisions about study design. An artist can generate variations while choosing the meaning and final composition. A small business can automate inventory arithmetic while reviewing unusual supplier changes.

Document why the chosen design is adequate

Write a short decision record: the problem, relevant requirements, options considered, chosen approach, evidence from the pilot, and conditions that would trigger reconsideration.

For example: “Use a fixed workflow to assemble weekly schedule drafts from confirmed records. Use a model only to improve descriptions. Route missing confirmations to the coordinator. Reconsider an agent if evidence gathering becomes the main source of delay.”

That record helps a future maintainer avoid rebuilding the system around a new tool without understanding the original constraints.

It also gives you permission to simplify. If a model step merely changes true to Confirmed, replace it with a rule. Removing an unnecessary component can reduce cost, latency, and the number of ways the process can fail.

A reusable prompt

Analyze this process using its trigger, inputs, decisions, outputs, exceptions, and owner. Compare manual work, rules, scripts, model-assisted steps, fixed workflows, and agents. Estimate setup, review, failure, and maintenance effort. Recommend the least complex approach that meets the acceptance criteria, and define a small pilot that could show the recommendation is wrong.

For students: automate the administration of learning thoughtfully

Students can use a formula to count practice sessions, a script to organize permitted files, or a model to generate questions from approved material. Those uses differ from outsourcing the thinking an assignment is designed to develop.

Ask which skill the task is meant to teach. If the objective is to learn a mathematical method, automating the answer may remove the practice. If the objective is to analyze an experiment, automating a repetitive filename conversion may preserve time for the analysis.

Follow the instructor's AI rules, and include unaided checks where learning depends on independent performance. A useful student pilot compares not only speed but also what the student can explain afterward.

Practice: choose five approaches deliberately

List five recurring tasks from school, work, home, or a volunteer project. For each, identify variability, consequence, volume, and an observable success criterion. Choose an approach and explain why a more complex one is unnecessary or justified.

Completion check: Each choice follows from the task's requirements and evidence, with review and maintenance included in the estimate.

Stretch: Remove one AI step from a proposed workflow and compare the result on the same sample inputs. Keep the simpler design if it satisfies the requirements equally well.