- Pick a task that happens often, is mostly reading or writing, and has an answer a person can check.
- Start with one team and real examples, not a company-wide rollout.
- Measure the task before you build, so you can prove the result.
- Keep a person in the loop until the numbers say otherwise.
Start with a task, not a tool
When a business says "we want to use AI", the conversation often turns to which model or which product to buy. That is the wrong first question. AI is useful when it takes over part of a specific job that people already do every day. So the first question is simpler: which task, done by which people, takes up time that you would rather spend elsewhere?
Write down three or four candidate tasks. For each one, note who does it, how often, and how long it takes. You will usually find one or two that stand out.
Four signs a task is a good fit
- It happens often. A task done fifty times a week is worth automating. One done twice a month is not.
- It is mostly reading or writing. Answering questions, drafting replies, pulling details out of documents and summarising long text are where today's AI models are strongest.
- A person can check the result. If someone can tell in a minute whether the output is right, you can use AI safely while you build trust.
- A mistake is cheap to catch. Start where an error costs a few minutes of rework, not a lost customer or a wrong payment.
Good first projects
In our work, these four types of project deliver value quickly for small and mid-sized businesses:
- Answering repeat questions from your own documents. Customers, parents or staff ask the same things again and again. An assistant that answers only from your website and documents, and shows where each answer came from, saves hours every week.
- Drafting first versions. Lesson plans, proposals, product descriptions, replies to common emails. The person still edits and approves, but starts from a draft instead of a blank page.
- Pulling fields out of documents. Invoices, claim forms, CVs and purchase orders all hold the same few fields. AI can read them and fill a spreadsheet or system for a person to check.
- Summarising long material. Reports, contracts and meeting notes turned into a one-page brief for the person who has to decide.
What to leave for later
- Decisions about money or people made without a human review.
- Projects that need data you do not yet have in one place. Fix the data first.
- "An AI for everything." Broad assistants are hard to test and hard to trust. Narrow ones get used.
Measure before you build
Before any work starts, record three numbers for the task: how long it takes today, how many times it happens each week, and how often it goes wrong. Without these, you will not be able to show whether the project worked, and the conversation becomes about opinions.
A simple test: if the task takes 10 minutes, happens 100 times a week and AI saves half the time, that is more than 8 hours a week back for one team.
Keep a person in the loop
For the first months, treat AI output as a draft. A person reviews, edits and approves. Track how often they change the output. When changes become rare for a type of task, you can decide whether to reduce the review for that type.
A 30-day plan
- Week 1: pick one task and collect 20 to 30 real examples, including tricky ones.
- Week 2: build a small working version on those examples.
- Week 3: let two or three people use it on real work and note every problem.
- Week 4: compare time and errors with your starting numbers, then decide to stop, fix or expand.
A month is enough to know whether a project is worth more investment. Most of the cost of a failed AI project comes from skipping this step and building something large first.
Not sure which task to start with?
We run a short assessment to find where AI will save your team time and money, before you spend on building anything.
Book a free call