Most companies have more ideas for AI than time to try them. The risk is not choosing badly so much as choosing too big: a project that tries to fix everything, drags on for months, and never reaches daily use.

The first project should be small, useful and low-risk. Here is a simple method to find it, in five steps you can do with a pen and paper.

Step 1: List the repetitive tasks

Start with the work, not the technology. Ask the people who do the work, not just their managers. A short question works well: what do you do every week that feels like copying, sorting or answering the same thing again?

Typical answers:

  • sorting the shared inbox
  • answering the same customer questions
  • typing invoice details into the accounting system
  • searching for an old quote or contract template
  • writing up meeting notes and chasing action points
  • putting together standard quotes

Write down every task, without judging yet. Ten to twenty items is a good list.

Step 2: Score by time and cost of errors

For each task, estimate two things.

Hours per week. Roughly how much time does the task take across everyone who does it? Rough is fine. You are comparing tasks, not producing a report.

Cost of an error. If the task is done wrong, what happens?

  • Low: easily noticed and fixed, such as an email sorted into the wrong folder.
  • Medium: costs some time or annoys a customer, such as a reply with a wrong delivery date.
  • High: costs real money, legal trouble or a key customer, such as a wrong payment or a wrong contract term.

The best first candidates are high on hours and low on cost of error. They save real time, and a mistake will not hurt much while you learn.

Step 3: Check the data is available

An AI agent can only work with information that exists. For each promising task, check:

  • Is the input digital? Emails, PDFs and documents work. Paper forms and phone calls need an extra step first.
  • Are the rules or answers written down? Price lists, terms, instructions, past examples. If the answer lives only in one person's head, the agent cannot use it.
  • Can the agent get access? Is the data in a system that can be connected to, and are you comfortable giving access? See is your company data safe with AI? for what to check.

A task that scores well but fails this check is a good second project, after the information has been collected.

Step 4: Start with low risk

Among the tasks still on the list, prefer ones where:

  • A person can approve the output before anything leaves the company or changes in a system.
  • Little sensitive data is involved. Internal instructions are safer than HR records.
  • It is easy to stop. If it does not work, the team can go back to the old way the next day.

This does not mean the first project must be trivial. It means the downside should be small while you find out what the tool can do in your business.

Step 5: Define success before building

This step is the one most often skipped. Before anything is built, write down in one or two sentences what "working" means. For example:

  • "Most drafted replies can be sent with little or no editing."
  • "Every invoice reaches the right approver on the day it arrives."
  • "New staff find the right contract template without asking a colleague."

Also write down how you will check it, for example by having the team mark each draft as used, edited or rejected for the first two weeks. Without this, you will end up judging the project on feelings.

A worked example

Here is a made-up shortlist for a company of about 30 people, to show how the scoring works.

TaskHours per week (estimate)Cost of errorData available?Verdict
Sorting the shared inbox5LowYes, all emailStrong first candidate
Answering delivery questions4MediumYes, terms are writtenGood, with a person approving replies
Typing in supplier invoices6MediumYes, PDFsGood, approvals stay as they are
Drafting large project quotes3HighPartly, pricing is in people's headsLater, after pricing rules are written down
Checking employment contracts1HighYesNot a first project

In this example, sorting the shared inbox wins. It takes real time, mistakes are cheap, and the data is already there. Invoice handling would be a sensible second step.

Common mistakes

  • Choosing the most impressive idea rather than the most useful. A customer-facing chatbot sounds exciting, but internal routine work often saves more time at lower risk.
  • Trying to automate the whole process at once. Take one step, such as sorting, and leave the rest as it is. You can extend later.
  • Leaving out the people who do the work. If the person who does the task today is not involved, the tool is built on the wrong description and nobody uses it.
  • Not agreeing who owns the tool. Someone needs to notice when results slip or the price list changes.

Then build it, and learn

Once you have chosen, keep the project tight: one task, a fixed scope and a short timetable. You will learn more from one working tool than from a long list of plans. Inside a 2–3 week AI sprint describes what that looks like in practice.

If you would rather have someone go through the list with you, that is what an AI Audit is for: a week-long review that ends with a prioritised list of what is worth automating, and an estimate of effort and benefit for each item.