Before You Automate It With AI: Ask These Questions

AI makes it easier than ever to look at a repetitive task and wonder why you’re still doing it manually.

Maybe it’s summarizing meeting notes, preparing a recurring report, sorting incoming requests or drafting a routine response. When an AI tool can handle at least part of the work, automation can feel like the obvious next step.
Sometimes it is.

But repetition alone doesn’t make a task a good candidate for automation. Work that looks straightforward often relies on context, professional judgment and decisions that are inherently human and easy to overlook until something goes wrong.

Before turning a recurring task into an AI-powered workflow, it’s worth taking a closer look at the work itself.

Start with the workflow, not the AI tool

Seeing what AI can do naturally leads you to think about where you might use it. That’s useful for experimentation, but it’s a less reliable way to improve an actual workflow.

Start instead with the process. Where does the work slow down? Which steps consume time without adding much value? Where are people repeating actions simply because that’s how the process has always worked?

A recurring task may be frustrating because the process is inefficient. Information may arrive inconsistently, or the work may pass through more hands than necessary. Adding AI could speed up one step while leaving the underlying problem untouched.

Clarifying what needs to improve makes it easier to determine whether AI has a useful role in the solution. It also brings another part of the workflow into focus: what the people doing the work already contribute.

Where does human judgment fit into an AI workflow?

Even routine work can depend on judgment. Consider a customer service team reviewing incoming requests before routing them or preparing a response. On paper, the process seems straightforward: read the request, identify what the customer needs and determine the appropriate response.

In practice, the human reviewer may notice something that doesn’t fit the usual pattern. They may remember an earlier interaction that changes how they should handle the request. Context rarely appears in a written process, but it still shapes the outcome.

That critical thinking can disappear quickly when the work is reduced to a set of inputs and outputs.

Identifying where those decisions happen helps clarify the role AI can play. It may be well suited to organizing information or preparing a first pass while a person continues to interpret the situation and determine what comes next.

Once you understand how human judgment fits into the process, you can think more clearly about the work AI is supporting and what could happen when the technology gets something wrong.

What happens when AI gets something wrong?

In 2022, an Air Canada customer turned to the airline’s chatbot after a family death to ask about its bereavement fare. The chatbot told him he could purchase his ticket and apply for the discounted rate afterward. That information was incorrect. After he relied on the response and Air Canada later denied the adjustment, a Canadian tribunal found the airline responsible for the inaccurate information its chatbot provided.

The situation illustrates why the consequences of an AI error matter as much as the error itself. An inaccurate internal summary may create a minor inconvenience. An inaccurate response that influences what a customer buys or how they make a financial decision carries a different level of risk.

The potential consequence should shape the level of human oversight built into the process. As the stakes increase, so does the need for meaningful review and clear opportunities for someone to intervene.

That makes the question more practical than simply asking whether the technology is accurate. Before automating a task, consider what happens when the output is wrong and how quickly someone would recognize it.

That answer may also expose another issue: whether AI has the right information to do the work well — and whether it should have access to that information in the first place.

What information does AI need to do the work well?

AI only has access to the context you give it. People often carry much more context into their work without realizing how much they rely on it.

They remember earlier conversations, understand expectations that were never documented and recognize when an otherwise routine situation deserves more attention.

That creates an important tension. The system needs enough context to produce useful work, but some of that information may be inappropriate to share.

A workflow that depends heavily on institutional knowledge may therefore be harder to automate than it first appears. The same applies when completing the task requires information that shouldn’t be entered into a particular AI system.

Before building AI into the workflow, understand what information the task depends on and what information the technology should be allowed to use. Those boundaries affect both output quality and the responsibilities of the people overseeing it.

Who is accountable for AI-assisted work?

If AI drafts a response or recommends an action, responsibility for that work doesn’t disappear. The Air Canada case makes that distinction especially clear. The tribunal rejected the idea that the chatbot could somehow be separated from the organization providing the information, finding that Air Canada was responsible for information presented through its website.

That same principle matters well before a dispute reaches that point. When something falls outside the normal process, someone needs to recognize it and respond. That human involvement should remain part of the workflow from beginning to end.

The person overseeing the work remains responsible for what moves forward, regardless of how much of the process AI supports. That means reviewing the output in context, applying professional judgment and knowing when something needs a closer look.

It also gives you a clearer way to judge whether the automation is actually helping.

How do you know AI automation is improving the work?

Time saved is useful, but it doesn’t tell the whole story. An automation that saves 20 minutes upfront may create another 20 minutes of corrections later.

Define what improvement looks like before introducing automation. For example, you might compare how long the process takes before and after introducing an AI workflow, alongside how often someone needs to revise an AI-generated output.

Testing also needs to reflect how people actually use the system. A 2025 audit of New York City’s MyCity chatbot found that seemingly minor changes in wording could produce different results, including instances in which the chatbot failed to answer a question but responded when the same words were rearranged. The audit also found inaccurate and inconsistent responses during its independent testing. 

A workflow might therefore appear effective based on broad measures while still creating friction for the people relying on it. If the workflow creates new friction, change it. Automation should improve how the work gets done, and that improvement should be visible in practice.
AI-enabled workflows shouldn’t become “set it and forget it” processes. They can and should be tested, adjusted and reconsidered as the work changes.

Better automation starts with better judgment

That recurring task you started with may still be a strong candidate for automation.

The key is understanding the work well enough to determine how AI can support it and what level of human involvement is needed at each stage. A well-designed workflow should make the work better, not simply make more of it automatic. That starts with asking better questions before you begin to automate.