AI automation in a small company does not start with picking a tool - it starts with sorting your tasks into three piles: the ones a plain rule can handle, the ones AI can draft under supervision, and the ones that should stay with a person. Search for the phrase and you get step-by-step rollout guides and lists of ten tools. What is missing is the answer to the question I hear most often from owners of companies with two to a few dozen people: which task do I hand over first, and where is the line past which artificial intelligence starts costing more than it gives back.
This article is about that line and about choosing the first process. There is no tool list here, because the tool gets picked last, once it is clear what it has to do. There are no savings percentages either, because the only measure that makes sense in a small company is hours: how many a task eats today and how many are left after the change.
What a small company can hand to AI, and what it cannot
Before the word “AI” comes up, it is worth separating two things that agency offers blur into one. Plain automation moves data and executes rules: a new order lands in a spreadsheet, an invoice changes status once paid, a customer gets an email after three days of silence. Every input has exactly one correct output. A connector or a short script does that work, and there is no room for interpretation in it.
Artificial intelligence belongs where the input has no fixed shape. A customer email written three different ways, a photo of a document, notes from a call, a supplier’s product description that needs rewriting in the store’s own voice. A language model can pull structure out of that material or produce text from it, which a plain rule cannot do, because there is nothing for a rule to check against.
Hence my three piles. The first pile holds tasks where the same kind of input always yields the same result. That is ordinary business process automation, the kind I laid out for stores in the piece on WooCommerce automation with Zapier and n8n. AI is unnecessary there, and even harmful, because it injects uncertainty into a task that has no use for it.
The second pile holds tasks where the input is messy but the output can be verified at a glance. A first reply to a routine question, assigning a ticket to a category, a draft description, pulling the fields from a supplier invoice into a table. AI makes sense here, on one condition: a person approves the result before it reaches a customer or a system.
The third pile keeps everything where the outcome depends on a relationship, on liability or on an exception: negotiating terms with a wholesale buyer, deciding to honour a complaint outside the policy, any correspondence somebody has to sign. This is not about what the model can do. It is about who answers for the consequence.
The line where AI hurts rather than helps runs through two places. First: the cost of a single mistake exceeds the hours saved. A model that every so often confidently writes the wrong amount into a quote does not save work, it moves the work to undoing. Second: there is nobody to check. If the only person who could approve results has no time for it, an unsupervised automation runs exactly until the first error reaches a customer.
What automation costs - the billing model, not a price list
“How much does AI automation for business cost” has no single answer, because you pay in three different models, and it pays to know which one you are in before an offer arrives.
A subscription to a ready-made tool. You pay monthly per user or per number of operations. Cheap to enter, first result within days, no code. The catch: the tool knows only typical processes, and your exceptions get handled by hand next to it. For first-pile tasks this is usually the right choice.
Usage at the model provider. You pay for text processed, so the bill grows with the number of cases handled. For one process in a small company it is an item that disappears next to the cost of human work, but you need to know about it, because it jumps when the model has to chew through a whole catalogue or a whole correspondence history.
A one-off build for your company. Here you pay for connecting the model to your systems, for rules and permission boundaries, for tests on real cases, and for a panel that shows what the automation did. What such a quote is made of and when it does not pay off at all, I covered separately in the piece on what an AI agent costs. A narrow scope, one process and one data source, I quote fixed-bid, in writing after a conversation, billed in stages, with the caveat that the amount depends on the number of systems and permissions, not on how “intelligent” the model is supposed to be.
The calculation worth doing before picking a billing model is simple: how many hours a month the task eats today, and what an hour of the person doing it costs. If the result barely covers a subscription, a custom build makes no sense. If it exceeds the build bracket within a few months, a subscription only postpones the decision.
How to pick the first process: time, repeatability, cost of error
The biggest mistake I see in first rollouts is choosing the most impressive process instead of the most profitable one. A chatbot on the website looks modern, while somebody in the company retypes data from emails into a spreadsheet every day and nobody counts it.
I choose the first process with three questions, in this order.
Time. How many hours a month it takes, counting everyone who does it. Do not estimate, log it for a week. Owners tend to underestimate the tasks they do themselves and overestimate the ones that annoy them.
Repeatability. How many times a month the input has the same shape. Ten cases, each different, is not a process, it is ten jobs. A hundred cases in three variants is a candidate. This is also where it becomes clear whether a task belongs to the first pile (same shape, same result: a plain rule) or the second (varied shape, verifiable result: AI with a check).
Cost of error. What happens when the automation gets one case in a hundred wrong and nobody notices for a week. A misfiled ticket costs a few minutes. A wrong amount on a document costs a customer and a correction. The first process should carry a low cost of error even at the price of a smaller saving, because that is where the company learns to supervise an automation before trusting it with something expensive.
The task that wins on all three questions is usually not the one the owner had in mind. In service companies it is most often the initial sorting of enquiries from the contact form and the inbox. In stores it is retyping data from supplier documents, and only later product descriptions. What to take off a store team’s hands first I laid out in the piece on online store automation, so I will not repeat the list here.
One thing that often surfaces during this sorting: a process that looks like an AI candidate turns out to be unstable. It is done differently depending on who is on shift. I do not automate such a process with a rule or with a model. First the people agree on one way of doing it, then it goes to the machine.
What I do differently - the result in hours, not a tool list
When somebody asks how to roll out AI in their company, I answer with a plan that fits in a few sentences, and that is deliberate.
I start with a measurement before the change. A week of logging how much time goes into the chosen task, for every person who does it. Without that number, a month later the conversation is about impressions, and impressions always favour whoever speaks loudest.
The first rollout runs in proposal mode. The automation prepares a result, a person approves with one click or corrects it. Every correction is information about a missing rule. The right to act without approval comes only once corrections have nearly stopped for a few weeks, and only for that one task.
On the technical side I watch three things that offers usually leave out. What exactly leaves for the model provider and what stays inside the company is settled at scoping, not after the fact. Every action the automation takes has to be visible in one place, with a way to undo it. And every connection to your systems needs a defined behaviour for when the other side does not answer, because that is precisely when automations do the most damage.
After a month we go back to the number from the measurement week. Hours dropped clearly: we extend to the next task on the list, in the same order. Hours barely moved: we close it and do not pretend it was a success. Closing after a month is a cheap answer to a question that would otherwise cost a year of subscriptions.
If you have a written list of tasks and do not know which one to take first, write to me through the contact form and paste that list, with the time and the number of repetitions next to each item. I will send back what is a job for a ready-made tool, what is AI with a human check, and what is better left to people, together with a quote for the first step. If it turns out along the way that you need not a model but a plain panel that gathers data in one place, I will say so directly, and how such a system gets priced I described in the piece on what a custom web application costs.
Frequently asked questions
What does AI automation actually mean in a small company? Handing a language model the tasks where the input is messy (an email, a document, a note) and the output can be checked. Anything with one correct output for a given input is plain automation and does not need AI.
Where do I start with AI process automation? With a week of logging how much time repetitive tasks eat, and with sorting them by time, repeatability and cost of error. The first process should be cheap to get wrong, not the most impressive one.
Does AI automation pay off with only a few people? It depends on hours, not headcount. A two-person company where one person spends half the day retyping documents has a better case than a twenty-person company where every task is different. Count the hours and the rate, then set them against a subscription or against a build bracket.
Can I start with a ready-made plugin instead of a custom build? Yes, and it is often the right first step as long as the process is typical. Which plugins really save time and which sell a promise, I covered in the piece on AI plugins for WordPress. A custom build starts paying off when the automation has to know your rules or reach a system that nothing off the shelf connects to.
How do I check that the automation is not making things up? Proposal mode with human approval at the start, a test on archived cases before go-live, and a panel where every action is visible and reversible. A model without access to your data will always invent, so supervision means giving it something to answer with and making the answer checkable.
Building an application, a panel or an integration rather than a website? I write B2B systems to order - backends, APIs, client panels, billing modules. Let’s talk about your project.