How to Find the Best AI Use Cases in Your Business, Task by Task

In two survey rounds, in late 2023 and late 2024, the economists Anders Humlum and Emilie Vestergaard surveyed about 25,000 workers in Denmark each time. The workers came from 11 occupations, including accountants, financial advisors and marketing professionals. Their study found that most chatbot users reported saving time, but their average reported saving in 2024 was only 2.8% of work hours. Controlled trials in occupations covered by the surveys often found productivity gains above 15%. The authors give two reasons: the trials focused on the occupations where chatbots help most, and many workers lacked the employer support that trial participants had. They also note that trials show the effects of chatbots "can turn negative if applied to the wrong tasks."

Our reading is that the gain from AI sits in particular tasks, done often, with the right information at hand. So when your team is asked to find AI use cases, a list of ideas is a weak place to start, and so is giving everyone a tool and hoping the savings add up. Neither gives you a way to compare one task with another. The alternative is a task inventory: a list of the recurring tasks your team actually performs, each with what starts it, what it produces, how often it happens and how long it takes. It also records who does each task and what information it depends on. The five sections below build one and sort it into a short list, drawing on published research at each step. Choosing which candidate to do first is a separate decision, and depends on your own business circumstances.

1. Look first at the work your clients pay for

For professional services businesses that sell their expertise, start the inventory in client delivery, meaning the research, analysis, drafting, reviews and reporting your clients pay for. BCG's The Widening AI Value Gap is a September 2025 report based on a survey of 1,250 senior executives. It found that 70% of the potential value from AI sits in core business functions such as sales and marketing, manufacturing, supply chain and pricing. The same report says that some companies falling behind "experiment too widely," automating processes here and there instead of focusing on a few important functions.

That evidence needs a bit more nuance before you apply it. The report's methodology shows that 90% of respondents work at companies with revenue of at least $1 billion. Its value figures are self-reported, and professional services is not one of its nine industries. So reading "core functions" as client delivery is our interpretation, not BCG's finding. We think it holds because in a business that sells its expertise, the core is the work the client sees and pays for.

In practice, pick one or two service lines and list the roles that deliver them. Billing, scheduling and onboarding still belong in the inventory, because they take time too. But if every row comes from the back office, the list has missed the work that matters most to your clients.

2. Ask your team where they already use AI

The people doing the work are a useful source of candidates, because many of them already use AI on parts of it. A Federal Reserve note from April 2026 reports that about 62% of workers in professional services said they used generative AI (tools that write text or create images) for work in November 2025. In the same note, about 33% of professional, scientific and technical services businesses reported using AI in the Census Bureau's business survey at the end of 2025. The note counts lawyers, accountants, consultants and advertising agencies, among others, in professional services.

Those two figures do not measure the same thing. One survey asks workers about generative AI and the other asks businesses about AI more broadly, and the author points to differences in sampling and weighting. So the gap does not show how much use is unapproved or how well businesses have organized it. But the worker answers are reason enough to ask your team what they already do. The Census Bureau's May 2026 review adds that less than 20% of businesses with four or fewer employees reported using AI, against 37% of those with at least 250 employees. Those surveys asked about use in any business function during the previous two weeks.

Microsoft's guidance on AI adoption makes the same point in general terms: employees "understand the business processes they perform every day and can identify opportunities where AI might improve their work." Ask each person which task they use AI for, which tool they use, and what they check or fix afterward. Make clear that the purpose is to understand the work, and record each answer as a task rather than a tool, because a tool name does not tell you where the time goes.

3. Turn each AI use case idea into a counted task

To compare an idea from a colleague, a vendor or a conference with the others, write it as a task. Each row records what starts the task, such as a client request, a month-end or a new document, and the finished result. It also records how many times a month the task happens, the minutes each instance takes, who does it and how the result is checked. Our AI strategy work describes this step as mapping "each routine task, with what it costs today."

The Danish data show why the count matters. In Humlum and Vestergaard's surveys, between 64% and 90% of chatbot users reported saving time, depending on the occupation, yet their reported saving averaged 2.8% of their total work hours. When you are looking for time savings, a task needs to recur often enough, and take long enough, for the saving on it to add up. An idea list leaves out those two facts, which is why ideas are hard to compare.

The difference shows up as soon as you write a row down. A row such as "first draft of the monthly client report, 40 a month, prepared by two analysts, checked by the account lead" gives you a task, a count and the people involved. It still needs what starts it and the minutes per report before you can compare its time with other tasks, but you can see what is missing. A row that says "use AI for reporting" gives you far less, because it has no start, no end and no count. The numbers in the example only show the format; they are not a benchmark.

Leave room for tasks that AI itself creates, such as checking drafts or keeping instructions up to date. The same study found that 8.4% of surveyed workers reported new job tasks created by chatbots. Those tasks cost time too, so they belong in the inventory next to the work they support.

4. Record the information each task depends on

Add a column for what each task has to read, such as client files, past deliverables, a pricing sheet or a policy. Then note whether that information is current, kept in one place and cleared for use with AI. Anthropic's September 2025 Economic Index report studied 1 million prompt-and-response records from August 2025, taken from customers who use its API (the connection that lets their own software use Claude). It used the length of a response as a rough measure of how complex a task is, and found that longer responses tended to come with longer inputs. From that, it concluded that "deploying AI for complex tasks might be constrained more by access to information than on underlying model capabilities."

The report has limits worth stating. It covers one vendor's platform, and the finding rests on a pattern in the length of inputs and outputs, not on a measured cause. The same report found that 77% of those business records showed automation patterns, mostly handing a whole task to the model. So the practical question for each task is whether the model can reach the information that step needs.

In your inventory, this column shows which tasks have their information ready for a test and which need work on their records first. It also shows early which tasks touch client information. Your rules on permitted use apply to those tasks, so it is better to know before anyone builds anything. When many tasks read the same documents, our guide to building an internal AI knowledge base covers how to decide which version counts and who may see it.

5. Sort the inventory into a short list of AI use cases

Once each row has a count, a time and an information column, a first sort comes down to two questions: how large is the task, and how ready is its information? Size comes from how often the task happens and how long it takes, while whether a client sees the result helps you judge how much it matters. Readiness comes from the column you added in the previous step. First set aside the tasks you cannot yet describe clearly or whose outcome depends mainly on judgment, and then use size and readiness to sort the rest. That gives you four groups, each with its own next action:

• Large and ready. These make up the short list: the tasks where you decide what AI could do and whether a test is worthwhile.
• Large, but the information is scattered. Prepare the records first, name the person who will gather them, and test once they are in place.
• Small or occasional. Leave these to individuals using the general assistant you have approved, where it suits the task and your rules on client information.
• Unclear or mostly judgment. Park them, note why, and look again when the task or the tools change.

Aim for three to five candidates rather than thirty. Focus on a few important functions, and keep the list small enough for a leader to read, question and fund. For each candidate, you still have to decide how much of the task AI should do. Our guide to an AI readiness assessment covers that choice, from preparing a draft for a person to running defined steps on its own. Choosing which candidate goes first is the next decision, and it has its own method, set out in how to choose the first workflow to automate.

Questions to test your list against

Before the list goes to anyone who will fund it, check it against these questions:

• Does every row describe a task with a start and a finished result, or are some still ideas or tool names?
• Does every row have a count and a time, even a rough one taken from a single week?
• Did the people who do the work write or check the rows, or only their managers?
• Is at least one row from client delivery, not only from administration?
• For each candidate, do you know where its information lives and who is allowed to see it?
• Is the short list five items or fewer?

A "no" shows which part of the inventory to finish before the list can support a decision.

What to do this week

Ask five people who deliver client work, or everyone who does if your team is smaller, to list the tasks they repeated this week, with a rough count and a rough time for each. Ask them the three questions from the second section as well: which task they use AI for, which tool, and what they check afterward. Put every answer into one shared sheet with the columns from the third and fourth sections. That sheet is the first draft of your task inventory, and you can sort it before anyone spends money on a study or a tool.

Questions buyers ask

Should we use a vendor's list of AI use cases to find ours?

Use it to prompt questions, but not as your list. A vendor library describes tasks in general terms, such as summarizing documents or drafting proposals. It cannot tell you how often your team does that work or what information it needs. Take each entry that sounds relevant and ask who on your team does that task and how often. Add it to the inventory only if someone can name the task and give it a count.

Should we stop staff from using AI tools we have not approved while we work this out?

It depends on the tool and the information it touches, but keep any restrictions you already have in place while you decide which uses to permit. Asking staff which tasks they use AI for does not mean you approve the tools they name, so you can collect the tasks without changing your rules. Write a short rule this week on which tools may handle client documents and which may not. Then put the reported tasks into the inventory, and note next to each one any approval it still needs.

Do we need clean data before we look for AI use cases?

No, because you can list a task before its information is ready for AI. The inventory then shows which records matter, since each short-listed task names the information it depends on. Microsoft's adoption guidance makes a similar point when it says AI does not require "an entirely new data strategy before adoption begins." For your first tests, organize the information behind your short-listed candidates, starting with the one closest to ready.

Who should own the list of AI use cases?

One named person should keep it, usually the operations lead, with each row checked by the people who do the task. The owner keeps the columns consistent and decides when a parked row comes back, while the people doing the work keep the counts honest. Review the list each quarter, because the tools and the work both change. A task that was out of reach last quarter may be ready now, and when a candidate moves into testing, it needs an owner of its own.

How long should building the inventory take?

It depends mostly on how many roles you include, but we suggest starting with about a week of notes per role, without waiting for precise timings. Rough counts and times are enough for the first sort, because you are looking for large differences between tasks. Add any recurring work that did not happen that week, such as quarter-end reporting. Start with the one or two service lines from the first section, add roles once the first sort has produced a short list, and measure times more carefully only for the tasks you decide to test.

Your best AI use cases are in the work your team already does, counted task by task, rather than in a longer list of ideas. If you want help building and ranking that list, our AI strategy work starts with interviews with the people who do the work and includes one ranked list, with costs.