A useful AI idea starts with a piece of work the team can describe: what comes in, what someone does with it and where the result goes. A chatbot demo can suggest possibilities, but it cannot tell you which task costs your company the most time.
Recurring administrative tasks offer concrete places to look. An accounting firm that stops retyping supplier invoices. A distributor whose order emails turn into draft orders in the ERP. A clinic where the receptionist no longer writes the same five replies forty times a day. They come from looking at the work people actually do.
A first shortlist can come from a few hours of conversations and observation over a week or two. The next step is to compare volume, review effort, available data and the cost of a mistake.
What a good first use case looks like
Generative AI is good at reading messy input and producing text. Rules are good at everything you can write down as “if this, then that”. The best first project combines both: one step that needs reading or writing, inside a process that is otherwise clear.
Good and poor first candidates
| Good candidate | Poor candidate | |
|---|---|---|
| Volume | Daily, dozens or hundreds of cases | A few times a month |
| Input | Emails, PDFs, calls, free text | Already structured fields |
| Output | A draft, a summary, filled-in fields | A final decision about a person or money |
| Checking | A person can verify it in a minute | Needs an expert for an hour |
| Data | In a mailbox, folder or database | In someone’s head or on paper |
| Owner | A team lead who wants it fixed | Nobody in particular |
If a simple rule does the job, use the rule. “Invoices from this supplier go to that cost centre” doesn’t need a model. Our post on workflows, AI workflows and agents explains where the line usually falls.
Step 1: list the repetitive work
Spend a week collecting, not judging. Three ways that work:
- Ask each team one question: “What did you do this week, more than ten times, that involved reading or writing something?” Write down every answer.
- Look at where work arrives. Shared mailboxes, helpdesk queues, contact forms, scanned post. Count what comes in each month and what kind it is.
- Sit next to someone for a morning. People underestimate their own routine work because they stopped noticing it. Two hours of watching shows the copying between systems, the searching for files, the same email written again.
Signs that a task is worth writing down:
- the same email or document gets written again and again with small changes
- someone copies data from one system into another
- a queue backs up every Monday or at month end
- one person is the only one who knows where to find things, and everyone asks them
- a report takes someone’s whole Friday afternoon
A list of 15–30 items is a manageable starting point. Comparing them will narrow it down.
Step 2: estimate the hours
For each item, estimate the volume per month and the minutes per case. Rough numbers are enough here. You’ll measure properly once you’ve picked one.
Here is what that might look like for a 12-person accounting firm:
Hours per month by task
Two things to keep in mind. First, hours spent aren’t hours saved. Someone still checks the AI’s work, so the saving is always smaller than the total, and you only learn how much smaller in a pilot. Second, small items rarely justify a project on their own. Ten hours a month of meeting notes is a job for an off-the-shelf tool, not for custom software.
To compare items with different costs, multiply the hours by what an hour of that person’s time costs the company, including contributions and overhead. An hour of a senior accountant and an hour of a junior assistant aren’t the same number.
Step 3: check risk, data and ownership
Hours show how much work is involved. Risk, data access and ownership determine how much of it you can realistically change.
What does a mistake cost? A wrongly sorted email costs a minute. A wrong amount in a VAT return can cost a penalty. A wrong hiring decision can be unlawful. A high cost of error doesn’t rule a use case out, but a person will check every output, which shrinks the saving. Decisions about people need extra care; see AI in HR and recruitment.
Is the data reachable, and may you use it? The AI needs the input (the emails, the PDFs) and usually some reference data (the client list, the price list, past answers). Where does it live? Can a system read it? Are you allowed to use it this way? Supplier invoices in a shared mailbox: yes. Client documents that arrive on paper and go into a cabinet: not yet.
Who owns it? A use case without an owner doesn’t survive its first problem. The owner is whoever is responsible for the work today and will judge whether the new way is better.
Step 4: score and pick one or two
Score each remaining candidate from 1 to 5 on the questions below. You don’t need precision. You need to compare candidates on the same scale.
Scoring questions, 1 to 5 each
- Volume How many hours a month go into it?
- Needs AI Messy input or text output, or would a rule do?
- Easy to check Can a person verify the result in a minute?
- Low cost of error What happens when the AI gets one wrong?
- Data available Can a system reach the input today?
- Owner Is there someone who wants it and can decide?
- Small integration How few systems does it touch?
For the accounting firm, the worksheet might look like this:
Comparing the shortlist
| Invoices | Emails | Reminders | Monthly notes | Meetings | |
|---|---|---|---|---|---|
| Volume | 5 | 4 | 2 | 3 | 1 |
| Needs AI | 4 | 5 | 2 | 5 | 5 |
| Easy to check | 4 | 3 | 5 | 3 | 4 |
| Low error cost | 3 | 3 | 5 | 3 | 4 |
| Data | 5 | 4 | 3 | 4 | 4 |
| Owner | 5 | 4 | 4 | 3 | 2 |
| Integration | 3 | 4 | 4 | 4 | 5 |
| Total | 29 | 27 | 25 | 25 | 25 |
Look at the “needs AI” row before the total. Chasing missing documents ties with two AI candidates, but a scheduled reminder with a good template covers most of it without any model. Finding that out is a good result too.
Then place the strongest candidates on a value and effort grid:
Value against effort
Value →
- 1 Supplier invoice entry
- 2 Client email answers
- 3 Monthly client notes
- 4 Document reminders (a rule)
- 5 Meeting notes
- 6 Tax advice chatbot for clients
Effort →
- 1 Supplier invoice entry
- 2 Client email answers
- 3 Monthly client notes
- 4 Document reminders (a rule)
- 5 Meeting notes
- 6 Tax advice chatbot for clients
Pick one, at most two. The first should be a quick win you can pilot within weeks. If you have the capacity, the second can be a bigger item you start planning. Don’t start five at once.
Examples by department
Finance and accounting
- Reading supplier invoices and receipts into the accounting system, with a person checking amounts and VAT. See AI document processing.
- Payment reminder drafts that take the client’s history into account.
- A short monthly commentary on the numbers for each client or cost centre.
Sales
- Turning call notes or meeting transcripts into CRM updates and a follow-up draft.
- Background research on a company before the first call.
- First drafts of quotes from past quotes and the price list. More in AI for sales teams.
Customer support
- Sorting incoming emails by topic and urgency and routing them to the right person.
- Suggested replies based on your help articles, sent by an agent after a check.
- Answering simple questions such as order status without a person. See AI in customer support.
Operations
- A distributor’s order emails and PDF attachments turned into draft orders in the ERP.
- Reading carrier emails and flagging the delivery problems that need action.
- Checking supplier documents (certificates, delivery notes) for missing information.
HR
- Job ad drafts from a short brief.
- Reading CVs into structured fields in the applicant tracking system.
- Answering employee questions from the staff handbook.
- Be careful with ranking or rejecting candidates automatically. It’s a high-risk area under the EU AI Act; keep a person making the decision.
Red flags
Drop or postpone a use case when:
- Nobody owns it. “It would be nice if…” isn’t ownership.
- The data isn’t there. Getting the data into a system is the first project.
- It’s mostly judgement. If every case needs an expert’s call, AI can prepare the file but won’t save much time.
- The volume is tiny. Twenty cases a month rarely justify building anything. A chat assistant will do.
- A rule would do. If you can write the logic as “if X, then Y”, build the rule.
- The process is broken. Automating a messy process gives you a faster mess. Fix the steps first.
- The real goal is a press release. “We use AI” isn’t a business case.
What to do next
- This week, ask each team the one question from step 1 and collect the answers.
- Estimate the hours for each item and drop the red flags.
- Score the rest, pick one, and measure its baseline properly. How to measure the ROI of an AI project explains how.
- Decide whether an off-the-shelf tool fits or you need something built (build or buy), then run a short pilot following the 90-day plan in our guide.
If you’d like a second opinion on your list, bring it to a 30-minute call. It’s usually enough to say which item to build first and roughly what it would cost.
