How to Know What to Automate in Operations

To know what to automate in operations, look for work that is high-volume, repetitive, rules-based, low in judgment, and stable in its inputs, and avoid work that is exception-heavy, relationship-driven, or bound by regulation. That part is easy; the criteria are well known. The hard part, the part that actually decides whether your automation succeeds, is finding where those tasks live inside your operation. They hide inside roles, blended with judgment work, invisible to org charts and output dashboards. You can only see them with activity-level data about what people actually do.

Everyone can recite the rules for what to automate. Almost no one can point to the specific tasks in their own operation that meet them. That gap, between knowing the criteria and seeing your own work clearly, is where automation decisions go right or wrong.

The work hasn’t disappeared. It’s just invisible. This article is about making it visible.

Key Takeaways

• The criteria for automatable work, high-volume, repetitive, rules-based, low-judgment, are easy and widely known.

• The hard part is locating those tasks, because they are blended into roles and hidden from org charts and output metrics.

• Five tests separate good candidates from bad: volume, judgment level, exception rate, regulatory exposure, and dependency.

• Some work should not be automated yet, judgment-intensive, relationship-driven, compliance-bound, and getting this line wrong is how AI initiatives backfire.

• Activity-level ground truth turns the universal criteria into a specific, defensible list of what to automate in your operation.

The Criteria Are the Easy Part

Ask any consultant, vendor, or article what to automate and you will get the same answer, because it is correct: automate work that is repetitive, high in volume, governed by clear rules, low in judgment, stable in its inputs, and structured in its data. That is a good list. It has been a good list for a decade.

The trouble is that the list is a filter, not an answer. It tells you what kind of work qualifies. It does not tell you which of the thousands of tasks running through your operation right now actually fit the description. And that second question is the entire game, because automation does not happen to “repetitive work” in the abstract. It happens to a specific step in a specific workflow performed by a specific role, and you cannot automate what you cannot locate.

So teams apply the criteria to the work they can see, the obvious, named processes, and miss the larger pool of automatable work buried inside roles that look, from the outside, like judgment jobs. McKinsey estimates 30% of work hours are automatable by 2030. Most operations could not tell you where even half of their share of that 30% currently sits.

The Hard Part: Finding Where Those Tasks Actually Live

The reason automatable tasks are hard to find is that they do not exist as tidy, separate jobs. They are interleaved with judgment work inside the same role, often within the same hour.

A claims examiner is, on paper, a judgment role. In practice, a meaningful share of an examiner’s day might be rote data entry, document classification, status updates, and rekeying information between systems, all highly automatable, sitting right alongside the coverage decisions that genuinely require their expertise. The job title says “judgment.” The activity data says “40% automatable.” Only one of those is true at the level automation operates on.

This is why org charts and job descriptions mislead. They describe roles, not tasks. And output dashboards mislead too: knowing a team processed 4,000 claims tells you the result, not the sequence of automatable and non-automatable steps that produced it. To find automation candidates, you need a layer most operations have never captured, the activity layer, what people actually do, moment to moment. That is exactly what ground truth workforce data provides, and how the approach surfaces it.

Five Tests for a Good Automation Candidate

Once you can see the work at the task level, run each candidate through five tests.

1. Volume. Does this task happen often enough that automating it frees meaningful capacity? A perfect candidate that occurs twice a month is not worth the build.

2. Judgment. How much human discretion does the task require? Pure rules-following is ideal; anything requiring interpretation, negotiation, or discretion is risky.

3. Exception rate. How often does the task deviate from the standard path? High exception rates mean the “automatable” task is really a judgment task in disguise, the most common reason automations underperform.

4. Regulatory exposure. Does automating this task touch a regulated decision, a compliance requirement, or an auditable record? If so, it may need human-in-the-loop controls or be off-limits entirely.

5. Dependency. Does the task depend on data, systems, or upstream steps that are not yet ready? A high-value candidate blocked by a dependency belongs later in the sequence.

A task that passes all five is a strong near-term candidate. One that fails on judgment or exceptions is exactly the kind of work that looks automatable and is not.

The Work You Should NOT Automate (Yet)

Knowing what to automate is inseparable from knowing what to protect. The work to leave alone, for now, includes anything judgment-intensive, relationship-driven, exception-heavy, or compliance-bound, and anything too low-volume to justify the risk.

This is not caution for its own sake. It is the lesson of every AI initiative that overshot. 55% of companies regret AI-driven layoffs, and the pattern is consistent: they automated or eliminated the judgment-intensive work along with the rote work, because they could not tell the two apart inside the same role. Klarna automated customer service, then rehired when the complex cases overwhelmed the automation. The complex cases were always there. They were just invisible until the automation removed the people who handled them.

Drawing this line correctly is the single highest-value output of seeing the work clearly. It is also impossible to draw from an org chart.

From Candidates to Sequence

Identifying candidates is not the same as deciding what to do first. Once you have a scored list of automatable tasks, weigh value against risk and dependency to build a sequence, the discipline covered in how to prioritize AI projects in operations. The list tells you what could be automated. The sequence tells you what to automate first, and that is what turns analysis into action. See what the prioritized output looks like.

Seeing the Work Clearly, in 90 Days

The reason most operations automate the obvious and miss the hidden is that the activity data has been impossibly slow to collect, a year of consultants shadowing staff, and even then only a sample.

The Ground Truth AI² Platform^™ captures it automatically. A lightweight client records activity at the click-region level, hundreds to thousands of data points per person per day, and Vision AI classifies each one into a precise picture of the work, not “in Salesforce” but “entering claim data into the intake form.” See the platform. The result, produced in a fixed 90-day engagement and interpreted by 20-plus years of operational expertise, is a specific, defensible list of what to automate in your operation, and what to leave alone.

The criteria were never the problem. Seeing your own work was. Fix that, and “what should we automate” stops being a debate and becomes a decision.

FAQ: What to Automate in Operations

How do you determine what to automate in operations?
Look for tasks that are high-volume, repetitive, rules-based, low-judgment, low-exception, and low-regulatory-risk. The criteria are easy. The hard part is finding which specific tasks in your operation meet them, because they hide inside roles alongside judgment work.

What are the 4 D’s of automation?
Traditionally: Dull, Dirty, Dangerous, and Dear (expensive). In knowledge-work operations the most relevant are Dull and Dear, and a fifth often added is “Data-rich” (well-structured, repeatable inputs). Use the framing as a filter, not a substitute for activity data.

What are the main types of automation?
Robotic process automation (rules-based, screen-driven), intelligent automation (rules plus AI), workflow automation (process orchestration), and AI agents (autonomous, model-driven). Each fits a different work profile; ground-truth data tells you which fits where.

Why can’t I just identify automation candidates from our org chart?
Because the chart describes roles, not tasks. The automatable work is interleaved with judgment work inside the same role, often within the same hour. You can only see the split with activity-level data.

What should I NOT automate?
Judgment-intensive, relationship-driven, compliance-bound, or low-volume work. Getting this line wrong is precisely how AI initiatives backfire, including the Klarna-style reversals. Protecting the right work is the other half of the decision.

Want a specific list of what to automate in your operation, backed by data? Book a 30-minute strategy call and we’ll show you what the activity data reveals.