How to Choose Which Tasks to Automate: 5 Criteria for RPA and AI, and Where to Keep Human Checks

Five Criteria and a Scoring Table for Choosing Tasks to Automate

Compare candidate tasks on five points — volume, steps, input, impact when something goes wrong, and whether the task will keep going — and choose the ones with lots of annual work time, steps you can put into words, and input that already arrives as data. The order is task first, tool second — not the other way around.

This article assumes you already have a few candidates for automation. Listing candidates in the first place is covered in "How to Do a Work Inventory." In terms of the six steps in "How to Improve Business Processes," this is the decision you make in Step 4 (design countermeasures) when automation is one of the options.

The decision flow: ask "can we get rid of it?" before "can we automate it?"

The trick is not to start with "can this be automated?" First ask whether the task can be eliminated, then run whatever remains through three filters in order: annual work time, steps, and input. Two of the five criteria aren't in Figure 1: "lifespan" already drops out at the first question as a task that's about to be retired, and "impact" doesn't decide whether to start — it decides where to put the human check.

A task that lands on "Standardize the steps first" or "Digitize the input first" in Figure 1 hasn't been ruled out. It just needs one more round of preparation before it's ready for automation.

The five criteria, and how to score them ◎ ○ △

The five criteria are in the table below, scored ◎ (strong fit), ○ (fair fit), or △ (weak fit). The thresholds for ◎ and △ are example values set for this article. They vary by industry and company size, so when you use the table yourself, it's enough to score candidates relative to one another.

Calculate annual work time as "time per case × cases per day × business days per year." For example, at 8 minutes per case, 40 cases a day, and 20 business days a month: 8 minutes × 40 cases × 20 days × 12 months = 76,800 minutes, or about 1,280 hours a year. The higher the volume, the more a difference of a few minutes per case adds up over a year.

Three kinds of tasks that aren't a good fit for automation

• Tasks that happen only a few times a year: building and maintaining the automation tends to take more effort than doing the work by hand
• Tasks done differently every time and driven by judgment: like price negotiations or evaluating job candidates, where the right answer changes case by case
• Tasks about to be retired or changed: if a system replacement or rule change is coming, whatever you build will soon be useless

"Not a good fit" doesn't mean "never a good fit." A task that each person does differently comes back as a candidate once the steps are standardized, and a task that arrives on paper comes back once the input is digitized.

Example: a parts distributor compares three candidates

Say a 30-person parts distributor has a staff member who handles both office admin and digital initiatives. The president asks, "Can't we automate something with AI?" — so this person compares three candidates.

All three assume the task will continue unchanged for now. Order processing scores △ on impact, but instead of dropping it, the verdict is "automate it, but keep a human check." That's because the size of the impact is less about whether to automate and more about where to place the human check.

Minami

Process improvement lead

At our company, too, people keep asking if AI can just write the reports for us...

Spark

DrillSpark consultant

Before automating a 72-hour-a-year job, ask who actually reads that report. On volume alone, order processing — which takes more than five hours every day — comes first.

Take the task that ranked highest in your scoring and start by drawing it as a process flow diagram. In the second half of this article, you'll mark each step in that diagram to decide what AI handles and where a person checks.

Three Reviews Before You Automate: Eliminate, Combine, Rearrange

If you don't first ask "can we eliminate it?", "can we combine it?", and "can we change the order or who does it?", automation just makes unnecessary work go faster. That's why "can it be eliminated?" sits at the very top of the decision flow in Figure 1.

Automate unnecessary work and you just get unnecessary work, faster

ECRS is a set of four improvement principles applied in order: Eliminate (remove it) → Combine (merge it) → Rearrange (change the order or owner) → Simplify (make it easier). Automation and AI belong to the last one, S (Simplify).

The four principles themselves are explained in "What Is ECRS?" Here, we use the first three as questions to ask before automating.

Three questions: eliminate, combine, rearrange

With the process flow diagram in front of you, check each step against the three questions below. Any step that fits one of them either drops off the automation list or changes shape.

Example: what the review did for the report and fax orders

Take the monthly sales report, which the scoring table flagged as low volume. Before even considering automation, the team asked who read it — and it turned out only two people did. In that case, the 72 hours a year of work mostly disappears without any automation at all.

Order processing is similar. Many Japanese B2B customers still order by fax, but if those customers switch to email, you don't even need to build a system for AI to read faxes. Only if some customers can't switch do you use the marks in a later section to decide how to handle that step.

RPA vs. AI: Every Step Falls Into One of Three Types

RPA is a tool for "repeating fixed on-screen operations exactly," and generative AI is a tool for "reading, sorting, and drafting text that doesn't follow a fixed format." What neither can take on — judgment and accountability — stays with people. Look inside the task you chose, and its steps split into these three types.

What RPA is good at: fixed on-screen procedures

RPA (Robotic Process Automation) records routine operations on a computer in software and repeats them automatically. It suits work where the steps are completely fixed, such as keying data into the order management system, looking up stock, or sending emails with a set text.

What AI is good at: reading, sorting, and drafting unstructured text

Generative AI is good at pulling the fields you need out of emails and PDFs that everyone writes differently, sorting inquiries by type, and drafting replies. It can take on the "input that's a little different every time" that RPA struggled with.

In order processing, for example, pulling three fields — part number, quantity, and requested delivery date — out of an order email is an AI step; entering those three fields into the order management system is an RPA step; and discussing the delivery date with the customer when an item is out of stock is a human step. It's common to combine both tools within a single task.

The two fail in different ways

RPA tends to stop when the layout of a screen or field changes — but because it stops, you notice something is wrong. Generative AI can produce output that is grammatically fine but factually wrong; it doesn't stop, it just gets things wrong in a plausible way.

This difference between "errors that stop" and "errors that slip by" decides where to put human checks. When you mark steps in the next section, always pair every step marked "AI" with a decision about how a person will check it right after.

Four Steps to Put Five Marks on Your Process Flow

Once you've decided to automate a task, the next move is to mark each step of its process flow diagram as "Eliminate," "RPA," "AI," "Human check," or "Human decision," so that what you hand over is decided step by step. Try to automate a whole task in one go, and the exception-heavy steps tend to drag everything back to manual work.

The five marks and when to use each

The line people most often struggle with is between "AI" and "Human check." When a person reviews what the AI read, draw that as its own "Human check" step, separate from the "AI" step.

Simply write the mark in front of the step name, like "AI: Read the order." You don't need special symbols or formatting.

Step 1: Diagram the current flow

Start by diagramming today's order processing from beginning to end. The key is to draw what you actually do now — retyping faxes, the manager checking every order — not the ideal flow. Symbols and drawing conventions are covered in "How to Map a Business Process Flow."

Step 2: Mark each step

Go through the steps in the diagram one by one and give each one of the five marks. If you're unsure, ask in this order: "Is this step's input a fixed screen, or unstructured text?" then "Does the result go outside the company?"

In Figure 2, retyping the fax is marked "Eliminate," because in the earlier review we decided to ask those customers to switch to email. Discussing the delivery date when an item is out of stock involves a promise to the customer, so it stays "Human decision."

Step 3: Narrow human checks with conditions

Once the steps are marked, the manager's check of every single order starts to stand out. If a person still reviews every order, automating the earlier steps won't shorten the wait for the manager. So narrow the check with a condition.

Suppose you decide that the manager checks only orders of ¥500,000 or more, and for everything else the staff member compares the AI's extracted fields against the original email. If 4 of the 40 daily orders are ¥500,000 or more, the manager's checks drop from 40 a day to 4.

Step 4: Redraw the flow after marking

Finally, redraw the flow to match the marks. Remove the "Eliminate" steps, add a conditional branch to the "Human check" steps, and for the "AI" and "RPA" steps, write in the step name who (or what) does the work.

Put Figures 2 and 3 side by side and the changes are clear at a glance: the fax step is gone, the manager's check is conditional, out-of-stock discussions stay with people, and the order confirmation goes out only after a check. The diagram also lets you estimate that, of the 8 minutes per order, only the few minutes of checking will still need a person. But that's an estimate. Measure the actual time after you try it.

In DrillSpark, you describe the process in plain language, the AI drafts a flowchart, and you refine it through conversation. If you write marks like "AI:" and "Human check:" at the start of step names, anyone on the team looking at the diagram can see right away how much has been handed over.

Three Questions for Deciding Where to Keep Human Checks

Don't put human checks on every step. Keep them only right before steps where "the result goes outside the company," "a mistake can't be undone," or "you need to be able to explain the reasoning." Put a person everywhere, and there's no point in having automated.

The three questions

Draw the line by "what happens if it goes wrong," not by "what kind of work it is." Even for the same AI draft, the wording of an internal memo and a delivery-date answer to a customer call for very different levels of human review.

Right after an AI step, spell out exactly how to check

RPA errors stop, so you notice them, but AI gets things wrong in a plausible way. If the step just says "check the AI's result," the checker doesn't know what to look at and tends to skim it and move on.

For order processing, write both what to look at and what to compare it against into the step — for example, "compare the part number, quantity, and requested delivery date against the original email." Setting a rough time budget for the check, such as 1–2 minutes per order, also helps you notice when checking has become too heavy.

Aimless visual checks of every case don't last

Automation research has long pointed out that people can't sustain monitoring a screen where almost nothing changes, starting with Bainbridge's 1983 paper "Ironies of Automation." A setup where people glance over every case without deciding what to look for may start out careful, but gradually turns into checking that only feels like checking.

That said, a short comparison with fixed fields to look at (part number, quantity, requested delivery date) and a fixed reference to compare against (the original email) — like the staff check in Figure 3 — can be sustained even for every case. What doesn't last is "looking over everything" without deciding what to look at. Heavy checks like a manager's approval should be narrowed by conditions such as order amount, or done by spot-check sampling.

It's safest to start with a broad check and narrow it as you watch how many cases the check actually corrects. Check patterns such as approval, escalation, and periodic audits are covered in detail in "Designing Workflows for AI Agents and Humans."

Minami

Process improvement lead

Even for the parts AI does, I'm scared not to look at every single one...

Spark

DrillSpark consultant

If you just look over everything without deciding what to check, there's no point in automating. A short comparison with fixed fields to check can keep going even for every order. For the manager's check, write "only review when this condition is met" on the diagram, and everyone knows who needs to look and who doesn't. Start broad, and narrow it once the number of corrections goes down.

Three Things to Prepare Before You Request Approval

Japanese companies often decide on new initiatives through a formal written approval request called a ringi. For that request, prepare three things: "the annual work time the task takes today," "a diagram showing which steps are automated and which stay with people," and "the scope and duration of a small trial." With these three in hand, your manager can judge in one pass "what we're handing over, how far, and what happens if it fails."

Record current work time for a few weeks and convert it to an annual figure

Before automating, record the current time per case and the number of cases. Two to four weeks is a good guide. If the task has busy periods — such as orders piling up at month-end — note whether your records include that period. Without numbers from before the change, nobody can say afterward whether things actually got faster.

Attach one marked-up diagram

Attach a marked-up diagram like Figure 3 to show, on a single page, how far you're automating and where whose check remains. Write a person's name or role on every "Human check" and "Human decision" step to make it clear who is responsible. Compared with a text-only request, it's much easier for readers to spot what they should ask about.

Start small: one task, one month

Don't automate the whole task from day one; try it on part of the steps. For order processing, run "just reading order emails and entering them into the order management system" for one month, and count how many cases the human check corrected. If corrections go down, that's your evidence for extending automation to the next step.

How to design the metrics for measuring impact is covered in "How to Do an As-Is/To-Be Analysis." At the approval stage, it's enough to have the before numbers and the names of the numbers you'll look at after the trial.

When you diagram order processing, starting from a template that covers everything from order to shipment speeds things up. It's a template for manufacturers, so if you're a distributor, remove the production and manufacturing lanes, keep only sales and logistics, rewrite it with your own steps, and then add the marks. You can try AI drafting on the free plan, too.

Five Common Automation Mistakes and How to Avoid Them

Automation failures happen less because of tools or technology and more because someone skipped the steps of choosing the right task and deciding how to hand it over. Of the five mistakes below, two happen when choosing the task and three when deciding how to hand it over.

Mistakes in choosing

Mistakes in handing over

The last mistake means the automation itself has come to depend on one person. Ways to reduce tasks that only one person understands are covered in "How to Reduce Key-Person Dependency."

Minami

Process improvement lead

The RPA we brought in before stopped when the screen changed, and the person who built it had already transferred to another team...

Spark

DrillSpark consultant

That's less a technology failure and more that no diagram or written steps were left behind. If the diagram shows which steps the robot handles, you can spot where it stopped right away — and hand it over to someone else, too.

Summary: Diagram One Task First, Then Mark Each Step

Decide on automation in this order: choose the task → eliminate and combine first → mark each step → keep human checks, based on conditions.

"Which task to automate" is decided at the task level; "how much to hand over" is decided at the step level. The first is settled with the scoring table, the second with the marked-up process flow diagram. When you request approval, submit that diagram together with the current annual work time and the scope of a small trial.

Start by diagramming the one task that ranked highest in your scoring table, and put one of the five marks on each step.

Related Templates

Related Guides

FAQ