Today, we hosted an AI-Powered Workplace webinar focused on calling attention to a problem more teams are starting to feel. That problem: AI is making people faster at their tasks, and the work around them has not changed at all. Most teams try to identify AI use cases by starting with a tool. We started with the work.
Teams can now draft the email, summarize the meeting, build the deck, and clear the queue in a fraction of the time. Individually, everyone is more productive. Collectively, the same requests keep arriving, the same questions keep getting asked, and the same handful of people keep getting pulled in.
We opened the session with our own experience. We have a Teams channel that helps our sales, operations, and engineering teams stay on the same page. It works. People get unstuck fast, and it wasn’t a top problem to solve. Then we pointed Copilot at it and asked a different question: what does this reveal about how work actually gets done? We found seven jobs happening inside one chat window.
It turns out that the channel was acting as a resource desk, a reference desk, a proposal support desk, an account assignment desk, an onsite coordination desk, a decision log, and a searchable memory of everything the team had figured out before. None of that was designed. It accumulated.
The issue usually sits in the work design. Requests arrive in whatever shape the sender chooses. Routing depends on who happens to be reading. Decisions live in threads. Closure is implied rather than marked. Every one of those looks like a communication habit, and every one of them is a workflow gap.
Microsoft’s 2026 Work Trend Index put a number on the pattern. Across 20,000 AI users in 10 markets, organizational factors like culture, manager support, and talent practices accounted for 67% of AI’s real impact. Individual mindset and behavior accounted for 32%. Their research team calls the gap the Transformation Paradox: employees are ready to work differently, and the systems around them still reward the old way. That’s why so many teams struggle to identify AI use cases that actually stick.
In our session, the attendees reacted most strongly to one number. In 25 active days, that channel carried 250 messages. Eighteen to twenty people saw every request. Two to five were actually needed. Before anyone could help, we found 25 or more clarifying questions.
Use that as the test: if you counted the questions people ask before they can start, what would the number be?
We did not use Copilot to answer faster. We used it to see what we were actually answering.
– Ashley Pyle, Chief Experience Officer, GadellNet
I want to be clear about where we are in this. We have used this method to understand the work and to come together as a team and plan what comes next. We have not rebuilt the entire workflow yet, work like that takes time, attention and collaboration among teams to get it right. The analysis is done, the opportunity portfolio is built, and the first pilot is scoped. The changes are ahead of us, not behind us. I find that this is the true state of most organizations doing this well, and it is a better place to start than a finished case study.
This is why we do not automatically start with an agent. Building one is the easy part, and it is the wrong first move. Teams need a method that finds the opportunity before anyone picks a tool.
5-Step Method to Identify AI Use Cases
1. Start with the work you already have.
Pick your busiest channel, queue, inbox, or handoff. You need volume, repetition, and visible friction. The point is that the evidence already exists. You are reading what happened, not imagining what might.
2. Analyze the patterns, not the tasks.
Ask what people repeatedly request, what has to be clarified before work can begin, who gets pulled in, who ends up owning the outcome, and how anyone knows it is finished. Our analysis turned 250 messages into 57 recurring requests across four categories. That is the moment the conversation stops being about messages and becomes a map of the work.
3. Build the portfolio before you pick a favorite.
Our analysis produced eight candidates: resource matching, reference matching, proposal support, intake normalization, pricing guidance, evidence for proposals, account assignment, and decision tracking. Every one traced back to something that already happened. When you identify AI use cases from real evidence, you end up with a portfolio, not a favorite.
4. Let governance decide the order.
Governance is not the approval gate at the end. It is how you choose. We scored every candidate on frequency, business value, data availability, risk, human judgment required, validation difficulty, adoption effort, and ownership. Risk has two inputs: how sensitive the data going in is, and how much the output matters. The higher of the two sets the tier.
5. Decide whether it needs technology at all.
Every candidate gets one of four answers: redesign the process, automate it, assist the person, or leave it alone. Most will not be “build an agent.” Account assignment needed routing rules and a named owner first. Pricing guidance scored high risk with a weak validation path. Naming those honestly is what makes the roadmap real, and it is what keeps a roadmap from filling up with work nobody should do.
We are using the same method to choose our own first pilot. Reference matching is where we are starting, because it recurs, the data already exists, and a person can confirm the answer in seconds. It was not the most exciting idea on the board. It was the one we could prove. We have a working version, and we are testing it before we change anything about how the team operates.
This pattern is not unique to us. In the Claude Cowork deployment guide, Anthropic shared that its legal team pointed Claude at 742 intake tickets to understand what the work actually looked like. The analysis helped them see which categories could be templated, which required human judgment from the start, and where the queue was backing up. This is the same operating move we made with our work: use AI first to see the work, then decide whether the right answer is process redesign, assistance, automation, or doing nothing.
I want you to be able to answer this question: before you build anything with AI, do you know what the work is actually doing today?
Pick one busy channel, inbox, queue, or handoff. Analyze what keeps repeating, where people get stuck, who gets pulled in, and how the work actually gets finished.
If you’re trying to identify AI use cases that deliver measurable business value, don’t start with a tool. Start with the work.
That is where the right AI use cases show up.
If you are looking for support in scoping, developing, or governing your own AI use cases, GadellNet can help. Get in touch today!
