HomeBlogAI in the business

AI implementation: what to automate first, and how to measure it

Everyone talks about AI, and most businesses that tried it are left with a subscription nobody uses. This is the opening article of a series on AI in small businesses and online stores: not which tool to buy, but which task to take first, how to do it in stages, and how to know after a month whether it really saved time or just moved work around.

A hotel reception service bell with a frosted pale green glass dome and a round polished green marble base, standing alone on a bright studio surface in diagonal window light
The customer rings when it suits them. Most orders arrive while the bell rings in an empty office.
58%
of orders arrived outside working hours, Sunday to Thursday, 09:00 to 18:00. 32,148 orders in 8 online stores, January to October 2026
32%
of orders arrived between 18:00 and midnight, on any day of the week
44%
the share of orders outside working hours in the store where it was lowest. No store has customers who keep office hours

What AI implementation for business means, and why most attempts stall

AI implementation for business is moving one repetitive task, with a result that can be checked, from a person to an AI tool that works inside the systems the business already has, while a person stays responsible for the result. You start with one task, measure how long it took before and after, and only then move to the next. It is not "buying a subscription" and not "bringing artificial intelligence into the company". It is solving one task, then another.

Most implementations stall because they start from the wrong end: the tool. Someone saw a demo, opened an account, sent the team a link, and two weeks later nobody logs in. No defined task, no measurement before, and no single person whose job it is to say "this works" or "this does not". The tool did not fail. The implementation never started.

We grow online stores, and over the past year we have seen up close what happens when repetitive work moves to AI. What we learned has nothing to do with the name of any model. It has to do with choosing the task, the order, and the measurement. This is the first article in the series; the next ones go deep into each area.

How to choose what to automate first: three tests

Every task that is a candidate for AI goes through three questions. The one that passes all three goes first. One that fails any of them waits.

  1. Does it repeat? Not "happens sometimes", but every day or every week, in the same shape. Answering a customer who asks where their order is repeats dozens of times a week. Negotiating with a supplier does not.
  2. Can the result be checked? A person must be able to look at the output and say within seconds whether it is right. A product description: checkable. "Strategy": not.
  3. Is a mistake cheap? If the AI gets it wrong, what happens? If the answer is "a customer gets an imprecise reply and we correct it in the next message", that is cheap. If the answer is "we gave 40% off the whole site", that is expensive, and it stays with a person.

AI for small business: what passes the three tests

This table is the working tool. Copy it, fill in the tasks of your own business, and decide tomorrow morning. The last column is our decision, from experience.

TaskRepeats?Checkable?Cheap mistake?Decision
Answering a customer about an order, a delivery or an exchangeWhatsApp, email, site chatEvery day, at every hourYes, against the orderYesFirst
Product descriptions, Google title and description, image alt textFor every new product, and old ones nobody updatedEvery weekYes, read and approveYesSecond
Copy variants for ads and emailsHeadlines, short and long versions, subject linesEvery weekYes, the ad is measuredYesThird
Weekly summary of the numbersRevenue, media spend, average order, what went up and downEvery weekPartly, you need to know what is rightMediumAfter the first three work
Pricing, promotions and discountsHow much, for whom, whenEvery monthHardExpensiveStays with a person
A complaint, a refund, an angry customerEvery case is differentNoHardExpensiveStays with a person
The decision tool: which task to automate first. The three tests, and what we do with them in an online store. Our experience working with online stores, 2026. Not a sample

Where to start: the customer is active while the office is empty

We did not pick customer replies as the first task on a hunch. We checked when orders actually arrive, in Israel time, in 8 Israeli online stores we work with, from January to October 2026: 32,148 orders. Against the working hours of a small team, Sunday to Thursday, 09:00 to 18:00.

58% of orders arrived outside those hours. 32% came in between six in the evening and midnight, 22% on Friday and Saturday, and 7% between midnight and six in the morning. An order is the moment a customer has already decided. Before and after it come questions: is it in stock, when does it arrive, can I swap the size. They arrive in the same hours, and whoever answers them in the morning is answering a customer who has already forgotten, or already bought elsewhere.

When orders arrive, by hour of the day

The share of all orders that arrived in each part of the day, on all days of the week. In dark: the hours outside 09:00 to 18:00.
0% 10% 20% 30% 6.8% 00-06 7.4% 06-09 20.0% 09-12 18.0% 12-15 15.5% 15-18 15.0% 18-21 17.4% 21-24
0% 10% 20% 30% 6.8% 00-06 7.4% 06-09 20.0% 09-12 18.0% 12-15 15.5% 15-18 15.0% 18-21 17.4% 21-24
Preneuriat data, 8 online stores, 32,148 orders, January to October 2026

This is not one store with a night-owl audience. In the store where the share was lowest, 44% of orders arrived outside working hours. In the median store 56%, and in the highest 64%.

Orders outside working hours, by store

The share of each store's orders that arrived outside Sunday to Thursday, 09:00 to 18:00. The lowest, the median and the highest store.
Preneuriat data, 8 online stores, 32,148 orders, January to October 2026

That is why customer replies are the first place. The task passes all three tests, and it happens precisely in the hours when nobody is there to do it. While the team is at home, the AI answers the questions that have an answer inside the system: where the order is, what the delivery time is, what the exchange policy says. In the morning a person goes over only what the AI could not answer.

Choosing the first task is the most important decision in the whole implementation. A task that repeats, that can be checked, and where a mistake is cheap. If it also happens when nobody is in the office, it goes first.

The order of work: one month, one task

The implementation that works looks boring. There is no "AI project". There is one task, one month, and one person who is responsible.

  1. Week 1: measure before. Change nothing. How many times the task happened this week, how many minutes each time took, and who did it. Quality too: how long a customer waited for a reply, how many product descriptions were written, how many mistakes there were. Without this week there is nothing to compare against, and every discussion afterwards is about feelings.
  2. Week 2: implement with a person behind every result. The AI prepares, a person approves before anything goes out. Yes, it takes time. This is the week you learn where it gets things wrong and what it is missing: a shipping policy that is written nowhere, product names only the team knows.
  3. Weeks 3 and 4: release what has proven itself. Replies of a kind a person approved twenty times in a row without a correction go out on their own. The rest stays behind approval. Count how many corrections a person made each week. That number has to fall.
  4. End of the month: measure after, and decide. The same measures as week 1. If the saving is real, the task stays with the AI and you move to the next task in the table. If not, you stop, without arguing with yourself. A "no" after one month is also a good result: it cost little, and it taught something.
  5. Only then expand. A second task, the same order. Not three in parallel. A team that gets three changes at once goes back to the old way in all of them.

How to know it really saves time

"Saves time" is the most common claim in the world of AI, and the least measured. Without measurement, a tool that produces a draft in three seconds and then needs ten minutes of corrections looks like a success. It moved work, it did not save it. Three numbers tell the truth.

These numbers live in a simple spreadsheet. You do not need a system. You need someone to fill it in during week 1 and week 4, and to be willing to read what came out.

What stays with a person

Some things are not handed over, even when the tool can do them. Not because it would do them badly. Because a mistake there is expensive, or because nobody would check.

The simple rule: AI prepares, a person approves, and what a person has approved a hundred times without a correction gets automatic approval. Not the other way round.

Common mistakes in AI implementation

The rest of the series

This is the opening article. In the coming weeks we go into each area in depth, in the same order we recommend implementing:

If you want to go through this process with someone who has already done it, without replacing your systems, read what AI implementation with Preneuriat includes.

How we measured, and what we did not

The numbers in this article come from one sample: 8 Israeli Shopify stores we work with, 32,148 paid, non-cancelled orders, from 1.1.2026 to 31.10.2026. What was measured is the time of each order in Israel time against working hours of Sunday to Thursday, 09:00 to 17:59. No amounts, customers or products were measured, and no customer enquiries were measured: the order serves here as a marker of the hours in which the customer is active. Everything else in the article, the three tests, the order of work and the way of measuring, is based on our experience working with online stores in 2026, not on a sample, and the numeric example in the measurement section is an illustration only.

Questions we get

What is AI implementation for business?

AI implementation for business is moving one repetitive task, with a result that can be checked, from a person to an AI tool that works inside the business's existing systems, while a person stays responsible for the result. You start with one task, measure time before and after, and only then move to the next.

Where do you start with AI in a small business?

With a task that passes three tests: it repeats every day or every week, it has a result a person can check within seconds, and a mistake in it is cheap. In online stores that is almost always customer replies: 58% of the orders we measured arrived outside working hours, which means the customer is active exactly when there is nobody to answer.

How long does it take to see whether AI really saves time?

One month on one task: a week of measuring before, a week of implementation with a person approving every result, two weeks of gradual release, and then a comparison. You measure minutes before and after including the check, how many corrections a person made, and one quality measure of the result.

Does AI for business suit a three-person company?

Yes, and sometimes more than a large organisation, because there the repetitive task sits with the owner. The conditions are to pick one task, not to replace systems, and to keep a person approving at the start.

What should you not let AI do?

Money: prices, discounts, refunds. An angry customer. And anything that goes out in the business's name to the whole list. The AI prepares, a person approves, and only what a person has approved again and again without a correction moves to automatic approval.

Fit check

Want to see your real ROAS?

Every Preneuriat client gets a live dashboard: store revenue, all media spend, real ROAS, updated all day. The first call goes both ways.

Check the fit