Which job to automate first

“Automate the business” is not a job

“We want AI in the business” and “all of this should run by itself” are understandable things to want. The trouble is that nothing can start from them: no first step, no way to test them, and no way to tell next week whether anything changed.

A job is different. “Answer everyone who asks whether we’re open on Friday” is a job, and so is “message everyone who left something in their cart and didn’t come back”. Each has a first step (your opening hours, your store), a test (one question on a Friday morning, one abandoned cart), and a clear sign it works (you stopped typing the answer, the carts get a message without you).

The research points the same way. An MIT report from August 2025 found that 95% of the enterprise AI pilots it looked at showed no measurable financial return. The figure has been argued with since, and fairly. The useful half is what the pilots that worked had in common: they were built for one specific process, judged by a business outcome rather than a demo, and connected to the systems the work already happened in. More on that report in You don’t buy AI. You deploy it.

It shows in our integration meeting too. Owners who arrive with one specific job leave with it working. Owners who arrive wanting to “automate everything” leave with a good conversation. The rest of this page is for picking the job.

The three-question test

Repeated work comes in two kinds, and the same test works for both.

Answering: the same questions, over and over. Your hours, your prices, “do you have it”, “where is my order”.

Chasing: the things you’re supposed to start yourself and often don’t. The abandoned cart, the customer who stopped ordering, the lead who went quiet after you sent a price, the reminder before an appointment, the nudge to reorder.

Take any job of either kind and ask three questions about it. About a minute per job is enough.

• How often does it happen? Every day is a strong candidate, every week can work, and once a month is not worth the setup yet. For chasing, count how often it should have happened, not how often you got round to it.
• Is the right answer, or the right action, the same every time? “We open at nine” is the same for everyone, and so is “you left something in your cart, want me to hold it?”. “Can you do me a better price?” depends on who is asking and how your month is going. Same every time: automate it. Your judgement: keep it.
• Can a system see it? For answering, the answer has to live somewhere a system can read: a price list, a store, a calendar, a spreadsheet, the FAQ you already wrote. For chasing, the moment to act does: an abandoned checkout in your store, an appointment tomorrow in your calendar, a lead with no reply for three days. If it only lives in your head, step one is writing it down.

A quick example. “Do you have it in size 40?” comes up every day, the answer is whatever the stock says, and the stock lives in your online store: three yeses. If carts get left in your Shopify store most days, the nudge is the same for every cart and the store records every abandoned checkout: three yeses again. “Can you knock something off the price?” also comes up every day, but the answer depends on the customer, the item and your month. It fails the second question, so it stays with you, however often it comes up.

Jobs that usually pass

These are the ones worth checking first, in both kinds. Every row is something AssistantLabs runs today.

Answering

All four run at eleven at night too. If the questions pile up after you close, answering them then is a first job in its own right.

Chasing

The first two need your store on Shopify.

One honest limit. WhatsApp lets you message freely only within 24 hours of the customer’s last message. Outside that window, the first message has to be a template WhatsApp approved in advance, and WhatsApp charges for templates. Chasing usually starts outside the window, because the customer hasn’t written to you lately. What WhatsApp charges.

Each one, step by step. The Academy walks through your FAQ, quote requests, follow-ups and WhatsApp notifications. Which of your systems can connect, by type of business, is in What can you automate in your business?

What should stay with you

Some jobs come up often and still fail the test, because the right answer depends on the customer in front of you. Don’t automate these, not first and not later. That isn’t a limitation to work around. It is the right answer.

• Complaints. An unhappy customer should hear from a person, and quickly.
• Refunds and returns. Every case is a little different, and yes or no is your call.
• Price negotiations. “Can you do it cheaper?” is a decision about your margin, not a fact to look up. An offer code you set in advance can go out on its own. Haggling over it is yours.
• Chasing someone who is upset. A customer in the middle of a complaint shouldn’t get an automatic cart reminder or a “we miss you” message. Sort it out with them yourself first.
• Anything where one wrong message costs you the customer. A delivery date you can’t promise, a question about someone’s health or a contract, a regular who is angry. If a mistake there would cost more than the time you save, it stays with you.

“Stays with you” doesn’t mean nothing happens. The agent recognises the situation, tags the conversation, and a WhatsApp alert reaches your phone with the customer’s name, number and last message. What it says in the meantime is decided with you when the limits get set, and the moment you reply yourself, it stops answering in that conversation.

Find yours in ten minutes

You don’t have to guess. Both lists are already in your WhatsApp. Put a timer on and do this:

• Scroll back through last week’s WhatsApp. Customer chats only. One week is enough.
• List one: every question you answered more than three times. The same question in different words counts once: “Open on Friday?” and “Until when are you open Friday?” are one line.
• List two: everything you meant to start and didn’t. The cart nobody chased, the quote nobody followed up, the no-show nobody had reminded, the regular who hasn’t ordered in a while.
• Cross out what needed your judgement, on both lists: discounts, complaints, anything where you had to think about who was asking.
• Circle the one that cost you the most, in time or in a lost sale. That’s the first job.

If it’s on the first list, you’re in the Answering table above. If it’s on the second, Chasing. And if nothing survived the crossing out, that is worth knowing too: the work that keeps coming back to you is judgement work, and the first job is having the agent tag it and alert you, so you stop hunting for it between chats.

Write it as a sentence. “Answer the ‘is it in stock’ messages” is a job. So is “remind everyone the day before their appointment”. “Customer service” and “marketing” are not. Bring the sentence and both lists: your real cases are what the job gets tested on, and whatever else survived is your next job.

Bring the job to the meeting

The job you circled is the most useful thing you can bring to our integration meeting, because in that hour it is the job that gets built first. The first meeting is free.

What gets built depends on the job. It may be the agent answering the questions on your first list, a cart-recovery sequence on your store, appointment reminders, or the leads list with its follow-ups. On a shared screen, with an AI expert, we connect your WhatsApp, connect the one system the job needs (the store, the calendar, the CRM, the price list), set what runs on its own and what comes to you, and test it on your real cases.

You decide when it goes live. The agent can start muted, answering nobody while you read what it would have said, and a sequence can be tried with a test send before it reaches a single customer.

After the hour it can go two ways, and both are fine.

• You keep it. You choose a plan, any plan, and your first three months come with a weekly session with an AI expert. Each one adds the next job from your lists: another question the agent answers, a reminder, a win-back sequence.
• You don’t. We disconnect what we connected, and everything goes back to exactly how it was.

Either way you’ll have seen one real job run on your real cases, which tells you more than any demo. If you have ten more minutes before the meeting, run Is your WhatsApp ready? too.