"What should we automate?" is the wrong first question. The right one is "what should we automate first?", because the first project decides whether there is a second.
A good first automation earns trust. The team sees hours come back, customers notice faster answers, and the next idea gets a yes. A bad first automation, even a clever one, teaches everyone that "AI does not work here", and that belief is expensive to undo.
So the choice deserves more thought than it usually gets. Here is the filter I use, and the jobs that usually pass it.
Four filters for a first job
Run every candidate through four questions. The job you want passes all four.
1. Does it happen every day?
Frequency is where the hours hide. A task that takes ten minutes and happens forty times a day is worth far more than a task that takes two hours once a month, even though the monthly one feels bigger because it is painful.
Frequent work also gives you feedback fast. If the automation handles something daily, you will know within a week whether it works. If it runs once a month, you will need half a year to learn the same lesson.
2. Does your team dislike doing it?
This sounds soft, but it matters. When you automate work people dislike, they want the tool to succeed. They report problems, suggest fixes and defend it to sceptical colleagues.
When you automate work people are proud of, or work that makes them feel needed, they quietly hope it fails. Adoption is emotional before it is technical. Start where you have allies.
3. Is a mistake cheap and easy to catch?
Every automation will get something wrong eventually. For a first project, choose work where an error is visible and fixable in minutes: a draft someone reviews before it is sent, a reminder that went out a day early, a message that was sorted into the wrong folder.
Avoid, for now, anything where one wrong answer can cost a client, money or legal trouble. Not forever, just not first.
4. Can the rules be written down?
Ask the person who does the job today to explain how they decide. If they can say "when a customer asks X, I check Y and reply with Z, unless W", a system can follow that. If the honest answer is "I just know", the job depends on judgement you have not captured yet, and it is not a good first candidate.
This filter also tells you where the preparation work is. Writing the rules down is often half the project.
Jobs that usually pass
In most small and mid-sized businesses, a handful of jobs pass all four filters again and again.
First response to incoming messages. The questions that arrive every day on WhatsApp, Instagram, email and the website: opening hours, availability, how to order, where you are, what something includes. An assistant can answer the routine ones instantly, at any hour, and pass everything else to a person with the conversation attached. This is often the single biggest source of saved hours, and customers feel it immediately.
Lead follow-up. Many sales are lost not to a "no" but to silence. Someone asked for a quote, got it, and nobody checked in. A system that drafts a polite follow-up after a few days without a reply, for a person to approve or send automatically, protects revenue you already earned the right to.
Booking and reminders. Confirming appointments, sending reminders, offering new times when someone cancels. Frequent, disliked, low risk, fully rule-based.
Sorting and routing. Tagging incoming requests by type and urgency and sending each to the right person. Nobody enjoys triage, and a wrong tag is easy to correct.
Turning messy input into structure. Pulling the details out of an order sent as a voice note, an email with an attachment, or a form filled in badly, and putting them where they belong. Your team checks; the system does the typing.
Internal reporting. The weekly summary that someone spends Friday afternoon assembling from three places. A draft that is ready on Friday morning gives that person their afternoon back.
Jobs that should wait
Some work fails the filters and is better left for later, once you have a system the team trusts:
- Complaints and angry customers. These need empathy and authority. Let a system gather the details and alert a person quickly, not handle the conversation.
- Pricing and negotiation. Anything that commits the business to money should stay with someone who can be accountable for it.
- Sensitive advice. Medical, legal, financial or safety questions. The cost of a confident wrong answer is too high for a first project.
- Creative direction. AI is a strong assistant for exploring ideas. Deciding which idea represents your brand is a judgement call, and it should be yours.
Notice the pattern: the system can still help with all of these by preparing, summarising and drafting. What waits is letting it act on its own.
How to run the first one
Once you have picked the job, keep the first version deliberately small:
- Measure the before. How long it takes today, how many a week, how long customers wait.
- Write the rules and examples with the person who does the job now. Make them the owner.
- Start with a person in the loop. Drafts that someone approves, or answers that go out only for the most routine questions.
- Review every week. What did it get wrong, what was missing, what should it stop doing.
- Widen slowly. When the team stops checking every output because they trust it, that is the signal to give it more.
This pace can feel slow. It is faster than restarting a project that lost the team's trust in its first month.
The real goal of a first automation
The point of the first project is not to save the most time possible. It is to prove, inside your own business, that this works and that it is safe. Once your team has seen one job taken off their desk properly, the second and third ideas come from them, not from you.
If you want help choosing yours, or checking whether the job you have in mind passes the filters, you can book a call at cal.michelbaakliny.com.
