Where to start adopting AI in a small company
Almost all the money lost on AI is lost before anything is switched on: choosing the tool before choosing the problem.
The conversation about artificial intelligence in a small company almost always starts backwards. Someone saw a tool, tried it out on a Sunday, and on Monday asks how to bring it into the company. A year later there are three active subscriptions, two that nobody uses, and not one process that has changed.
The waste does not happen during implementation. It happens before: choosing the tool before choosing the problem.
The first decision is not what to buy
Start by identifying one specific process that already hurts, one that happens many times a month, whose result can be measured and whose errors are tolerable. The tool is chosen afterwards, once you know what it has to solve. Sometimes the answer turns out to be simpler and cheaper than expected.
That is one of the best possible outcomes of a diagnosis. A good part of what people set out to solve with AI is better solved by putting the data in order, connecting two systems or removing a step that no longer makes sense: it arrives sooner, costs less, and leaves the ground ready for when automation really is the next move.
The four questions that put the conversation in order
1 · How many times a month does the process you want to automate happen?
Automating has a fixed start-up cost —analysis, building, testing, correcting during the first weeks— and a small variable cost. That is why frequency rules: a process that happens four hundred times a month pays for its automation even if it saves two minutes each time. At four times a month, on the other hand, the start-up cost takes a long time to come back, and it is better kept for later.
It is worth counting before deciding. It is a ten-minute calculation and it usually brings surprises in both directions: processes that seemed constant happen rarely, and others nobody had on the radar turn out to be the ones that repeat the most.
2 · What does it cost today to do the process badly?
Not what it costs to do it: what it costs to do it badly. An appointment lost because nobody answered a message at nine at night, a quote that went out with the old price, an order shipped to the wrong customer.
That number —the cost of the error, not the cost of the work— is the one that usually justifies the project, and it is the one that is almost never in the spreadsheet used to ask for the budget.
3 · Where does the information live today that the automation would need as a reference?
This is the question that stops the most projects, and it is best that it stops them early.
An automated system needs somewhere to look up the truth. If the real opening hours live in the manager's head, if the stock is in three files that do not match, if the current price depends on who you ask, there is nothing to look up. AI does not put disordered data in order: it repeats it with more confidence.
When the answer to this question is not clear, the first project is not an AI project: it is putting the data in a single place. It sounds less modern and it is what makes everything else possible, so it is not a wasted step but the first one. It is also work we do: putting the information in order and connecting the systems you already have to each other is part of the trade, and it is usually the step that unblocks everything waiting behind it.
4 · What happens if the system gets it wrong?
No system is right one hundred percent of the time, neither the automatic one nor the one run by a tired person at six on a Friday. The difference lies in what is built around it so that a failure is caught and corrected before it reaches anyone, and that is designed differently depending on what is at stake. So it is worth sorting the candidates into three groups:
- What a person reviews before it goes out — sorting email, writing a first draft, summarizing a meeting, proposing replies. The system proposes and you decide, so the worst that can happen is discarding a draft. Start here: it is where you see how it behaves with everything under control, and where the team gains confidence in it.
- What talks to the customer — booking, confirming, reminding, answering frequently asked questions. Here the work consists of keeping the system from improvising: it looks up the schedule and the price in your system instead of inventing them, and it hands the conversation to a person as soon as something does not fit. Built well, what could have been a wrongly booked appointment stays a conversation that someone picks up.
- What is never left in the system's hands — charging, invoicing, dispatching, anything with legal or health consequences. Here the system does the heavy work and leaves it ready, and a person approves before it goes out. It is not a limitation of the technology: it is a design decision, and it leaves almost all the work done — what is left is to read and approve.
The rule we use in house is simple: no change reaches a production system without a responsible person approving it. To that are added the three things that make a failure visible in time: a record of what the system did, an alert when something falls outside what was expected, and a way to go back. This is not decorative caution: it is what lets you sleep while the system is on duty in the small hours.
What a first project that goes well looks like
A reasonable first project has four traits, and none of them is “using the newest technology”.
| A single process |
|---|
| Two processes at once double the variables and, when something fails, nobody knows which of the two caused it. |
| An owner with a name |
| Someone in the company who answers for the result. Projects without an owner are abandoned without anyone noticing. |
| A number agreed before starting |
| Decide up front which figure will say whether this worked —minutes saved per week, appointments that stop being lost, errors that no longer happen— and write it down. If the measure is chosen at the end, the one that came out well gets chosen, and then it measures nothing. |
| A date to decide whether it continues |
| With an explicit criterion for switching it off. A pilot with no decision date becomes a permanent expense. |
The four mistakes that cost the most
- Buying the tool first. The tool is chosen once the problem has been described. The other way round, the problem gets bent to fit the tool.
- Starting with the most visible thing. What is seen from outside is usually the most delicate. The first project is better kept internal, where a mistake is corrected without a customer finding out.
- Not deciding what information may leave the company. It is settled in an afternoon and it avoids serious trouble: it is in what data should not go into an AI model.
- Confusing a pilot with production. A pilot that works in a demonstration is not ready to serve customers. Between the two there are rules, failure handling, monitoring and a way out to a person.
What to do this week
Without hiring anything, and in an hour:
- Write down the five processes your team does by hand the most times a month.
- Next to each one, put how many times and what it costs when it goes wrong.
- Mark separately the ones whose information does not yet live in any system. They are not ruled out: what changes is where you start, because that data has to be put somewhere before automating anything. It is a job with a beginning and an end, and it is usually shorter than it looks.
- Of the ones that do have their information in a system, choose the one with the cheapest error. That is your first project. And if all of them were marked in the previous step, you already know what yours is: putting the data in order.
If you are ready to start automating and bringing AI into your company, the technical division will go through your list with you in a thirty-minute meeting and advise you on which of your processes to start with, why, and what would be needed first. No cost, no obligation to hire us.
Frequently asked questions
Where does a small company start with AI?
With a specific process that happens many times a month, whose errors are cheap to correct and whose information already lives in some system. The tool is chosen after describing the process, never before.
Do you need your data in order before using AI?
For anything that looks up company information, yes. An automated system needs a source of truth to consult; if the data is spread across files that do not match, the system will repeat the inconsistency with more confidence than a person would have.
How long does a first AI adoption project take?
It depends on the process and on the state of the information. An assistant covering a simple, well-defined process can be ready in a matter of days; a process with many exceptions or with scattered data takes weeks. Any estimate is sharpened by looking at those two things, so ask for it as a range and not as a date. What is worth fixing up front is when the decision will be made on whether the pilot continues or is switched off.
Do you need to hire someone technical to adopt AI?
You do not need to hire anyone on staff. The technical part can be contracted by project, and for a small company that is usually the sensible route: you pay for the work you need instead of for a permanent position, and you get a team with more craft than a single hire could reach. What does have to stay inside the company is the project owner: someone who answers for the result and who knows the process. That part is not subcontracted, and it is the part that decides whether the project goes well.
Sources and notes
- The decision framework in this article is Navhera's working method (discover, diagnose, design, implement) described on the home page, applied to the phase before any engagement.
- The order this article proposes comes from the practice of putting automation in place and from the mistakes that repeat when starting with the tool.
Written by the Navhera team and reviewed before publishing. If you spot an error, write to us and we will correct it with a note.