One of the more interesting parts of bringing AI into a company is that the question changes over time. At first, the question is simple: where can AI help? That is a good starting point because it gets people experimenting, it gets teams comfortable, and it helps people see that AI is not just a toy or a chatbot sitting off to the side. It can summarize, draft, review, classify, extract, compare, and answer questions inside the work people are already doing.
But after a while, that question is not enough. AI can help with a lot of things, but that does not mean it belongs in all of them. The better question becomes: where does AI actually create leverage? That distinction matters because there are workflows where AI makes someone’s day a little easier, workflows where AI removes hours of repetitive work, workflows where AI creates a better first pass so a human can focus on the important judgment call, and workflows where AI just adds cost, complexity, and another thing for the team to babysit.
Those should not all be treated the same. Over the last year, as we started applying AI beyond engineering, this became one of the biggest lessons. The work was not just finding places to use AI. The work was learning how to tell the difference between a useful AI workflow and an expensive demo.
Start by Watching the Work
The easiest mistake to make is starting with the tool. Someone sees a new AI product, a new model, or a new workflow and immediately starts looking for places to use it. That can create interesting experiments, but it does not always create durable value because the tool is not the center of the work. The department is.
The better starting point is to watch how people actually work. Sit with the team and observe what happens during the day. Look at where requests come from, where information gets copied, where decisions get made, where people wait on someone else, and where the same checks happen over and over again. A lot of the best AI opportunities are not hidden inside some giant transformation project. They are sitting inside the small repetitive moments that everyone has accepted as part of the job.
This is especially important when bringing AI to departments outside engineering. Engineers are used to changing tools, automating work, and rethinking systems. Other departments may not think about their work that way, even if their workflows are full of automation opportunities. If you come in too quickly with a solution, it can feel like disruption. If you start by understanding the work, it feels like partnership.
When we started looking at other departments, the most useful thing was not asking, “What AI tool do you want?” It was asking better questions about the work itself. What do you review every day? What do you always double-check? What takes longer than it should? What do people ask you for repeatedly? What information do you need before you can move something forward? What makes a request easy versus painful?
That is where the signal usually is. AI tends to be useful where there is repeated context, repeated review, repeated classification, repeated summarization, or repeated decision-making against known patterns. It is less useful when the work is rare, highly relationship-driven, poorly understood, or so nuanced that a person has to redo most of the output anyway.
Improve the Current Workflow First
The first version of an AI workflow should usually make the current process easier, not completely reinvent it. This helps the team build trust because they can compare the AI output against something they already understand. They know what good looks like. They know what bad looks like. They can tell when the system is useful and when it is just confidently wrong.
That might mean using AI to summarize a document before someone reviews it. It might mean extracting key details from a file before someone enters them into another system. It might mean drafting a response that a person can edit before sending. It might mean checking a record for missing information before it gets handed off. None of those are revolutionary by themselves, but they are valuable because they introduce AI inside the normal flow of work.
This stage matters because AI adoption is not just a technical problem. It is also a comfort problem. People need to see AI work in the context of their actual day, not in a polished demo. They need to build a feel for when they can trust it, when they need to correct it, and what kind of context makes the output better.
It also gives you a safer way to learn. When AI is assisting the workflow, the human is still close to the output. Mistakes are easier to catch, edge cases become visible, and the department can help shape the workflow without feeling like the entire process is being taken away from them. That feedback is what makes the next version better.
This is where a lot of teams stop, and it is still useful. Making the existing workflow faster is a real win. But if AI only helps people do the same work a little quicker, you may still be leaving the biggest value on the table.
Move From Assistance to First Pass
The bigger unlock comes when the question changes from “How can AI help this person do the task?” to “How much of this workflow can AI do before a human needs to step in?” That shift is where AI starts to change capacity instead of just speed.
In an assistive workflow, the human still carries the whole process. They gather the information, make the decision, perform the review, draft the response, and move the work forward. AI might help along the way, but the person is still responsible for every step.
In a first-pass workflow, AI does more of the initial work. It reviews the file, prepares the checklist, identifies missing information, drafts the recommendation, summarizes the issue, or flags the exceptions. Then the human comes in to approve, correct, escalate, or apply judgment where it matters.
That is a much better role for AI in many departments. The human is not removed from the process, but their time is focused on the parts that actually need them. They are no longer spending as much time on the repetitive middle of the workflow. They are spending more time on review, judgment, and exceptions.
This is also where the quality of the workflow can improve. A human doing repetitive checks all day can get tired, interrupted, or inconsistent. An AI workflow can apply the same review pattern every time and create a structured output for the human to inspect. That does not mean the AI is always right, but it does mean the process can become more consistent and easier to audit.
The goal is not full automation for the sake of automation. The goal is the right amount of automation. There are workflows where AI can do 70 percent of the work and a human should always make the final call. There are workflows where AI can do 95 percent of the work and only escalate exceptions. There are workflows where AI should only assist because the risk of being wrong is too high. Knowing the difference is the work.
Human in the Loop Has to Mean Something
“Human in the loop” is one of those phrases that sounds responsible, but it does not mean much unless you define it. A human in the loop should not mean someone vaguely looks at the AI output at the end and hopes everything is fine. It should mean the workflow is designed around specific human checkpoints.
Sometimes the human belongs at the beginning of the workflow, providing context or making sure the request is framed correctly. Sometimes the human belongs in the middle, reviewing exceptions or deciding how to handle an ambiguous case. Sometimes the human belongs at the end, approving the final output before anything is sent, filed, or completed. Sometimes the human only needs to be involved when the AI has low confidence or hits a known risk area.
The key is to be intentional. If the human checkpoint is too early, the AI may not remove much work. If it is too late, the risk may be too high. If it is too broad, the person may end up redoing everything. If it is too narrow, the workflow may miss the nuance that made the human important in the first place.
This is where the department’s expertise matters. The people doing the work already know where mistakes are expensive. They know which edge cases are common. They know which parts of the workflow are repetitive and which parts require experience. They know when something looks technically correct but still feels off.
AI workflows get better when those people help design the checkpoints. The goal is not to make the department adapt to the AI. The goal is to make the AI fit the reality of the department’s work.
Look for Repeated Judgment, Not Just Repeated Tasks
A lot of automation starts by looking for repetitive tasks, and that is still a good place to look. But with AI, some of the best opportunities are not just repetitive tasks. They are repeated judgment patterns.
A repeated task is something like copying information from one system to another. That can be automated, but it may not need AI. A repeated judgment pattern is different. It is when someone reviews a document and decides whether information is missing. It is when someone compares a request against a set of requirements. It is when someone reads a long thread and figures out what needs to happen next. It is when someone looks at a record and decides whether it is ready to move forward.
Those are the places where AI can be especially useful because it can handle messy inputs. It can read, summarize, compare, classify, and produce a structured first pass. That is different from traditional automation, which often needs the inputs to already be clean and predictable.
This is also why observing the work matters so much. People do not always describe these judgment patterns as workflows. They may just say, “I review this,” or “I check this,” or “I make sure everything looks right.” But inside those phrases are often dozens of small decisions that happen the same way every day.
Once you identify those patterns, you can start to design around them. What context does the AI need? What rules should it apply? What should it return? What confidence level is acceptable? What should get escalated? What should a human always approve? What should be logged so we can audit the decision later?
That is where AI becomes more than a helper. It becomes part of the workflow.
Measure the Workflow, Not the Demo
One of the traps with AI is that demos are easy to love. A good demo can make almost any workflow feel like it is worth building. The real test is what happens after the 50th run, when the novelty is gone and the workflow is just part of the day.
That is why the workflow needs to be measured in production. Did people actually use it? Did it save time? Did it reduce interruptions? Did it catch things that would have been missed? Did it create better consistency? Did it reduce the number of back-and-forth messages? Did it help the department move more work through the system without adding more people?
The correction rate matters too. If humans are rewriting most of the output, the workflow may not be good enough yet. If they are only making small edits, the workflow may be creating real leverage. If they are ignoring the output completely, the AI may be solving the wrong problem.
Cost matters as well. AI is easy enough to try in a lot of places, but it is not free enough to run everywhere forever. Every workflow has a cost, whether that is model usage, engineering time, maintenance, monitoring, or the operational cost of people reviewing outputs. If the workflow does not save enough time, improve enough quality, or unlock enough capacity, it may not be worth keeping.
That does not make the experiment a failure. It just means the workflow did not earn its place. That is a healthy outcome if you are honest about it.
Be Willing to Turn Things Off
One of the most important parts of an AI strategy is being willing to stop using AI where it does not make sense. That sounds obvious, but it is harder in practice because teams can get attached to the idea of a workflow even when the value is not there.
Some workflows are too small to justify the cost. Some are too rare to justify the maintenance. Some are too sensitive to trust beyond an assistive role. Some require so much human review that the AI step does not save anything. Some create a better demo than a durable process.
That is why AI workflows need a way to be audited. You should be able to look at how often they run, how often they succeed, how often humans change the output, how often they get ignored, and what they cost. Without that, it is easy to confuse activity with value.
Turning off a workflow should not feel like failure. It should feel like part of the process. The goal is not to prove that AI belongs everywhere. The goal is to find the places where it creates meaningful leverage and keep improving those.
In some cases, the best decision is to leave the workflow alone. In other cases, the right answer is traditional automation, not AI. In other cases, the workflow needs to be cleaned up before AI can help. The maturity is in knowing which situation you are looking at.
The AI Workflow Test
A useful way to think about this is to give every potential workflow a simple test before investing too much in it. Is the work frequent enough to matter? Does it involve repeated review, summarization, classification, extraction, or decision-making? Can the inputs be made available to the AI with enough context? Is there a clear definition of what good output looks like? Can a human review or approve the output at the right moment? Can we measure whether the workflow is saving time, improving quality, or increasing capacity?
If the answer to most of those questions is yes, it may be a good candidate for an AI workflow. If the answer is no, it may still be a good candidate for a small assistive tool, a traditional automation, or no change at all.
That framing helps keep the conversation grounded. Instead of asking whether a department should “use AI more,” you can ask which workflows pass the test. That is a much more useful conversation because it connects AI to the actual operating model of the team.
It also helps avoid the two extremes. One extreme is assuming AI is magic and should be added to everything. The other extreme is waiting for perfect certainty before trying anything. Neither is helpful. The better approach is to experiment close to the work, measure what happens, and keep the workflows that prove their value.
The Real Goal Is Capacity
The reason to bring AI into departments is not to make the company feel more modern. It is not to say every team is using AI. It is not to add another tool to the stack.
The reason is capacity.
Can the department handle more work without burning people out? Can the team spend less time on repetitive review and more time on judgment? Can requests become more self-service without lowering quality? Can the system catch issues earlier? Can humans focus on the parts of the process where their experience actually matters?
That is the real promise of AI in the company. It is not replacing the department. It is changing where the department spends its energy.
The best workflows we built were the ones where AI took on the repetitive, structured, reviewable parts of the work and brought humans in at the right moments. That made the process easier to scale, easier to audit, and easier to improve over time.
The worst candidates were the ones where AI added another layer without removing enough work. Those are the workflows that feel exciting at first but eventually become noise.
Where AI Actually Belongs
AI belongs where it creates leverage. It belongs where it can produce a useful first pass, reduce repetitive work, improve consistency, or help a department move faster without losing quality. It belongs where the human checkpoint is clear and the workflow can be measured.
AI does not belong everywhere. It does not belong in a workflow just because the workflow exists. It does not belong where the output cannot be trusted, where the cost is higher than the value, or where people have to redo the work anyway.
That was one of the bigger lessons from the last year. The question is not whether AI can help. Most of the time, it probably can. The better question is whether it helps enough to justify becoming part of the way the company works.
That is the bar.
Not every workflow needs AI. But the right workflows can change the capacity of an entire department.