How to use AI in your business without getting sold snake oil
A practical guide for business owners who aren't from tech and don't want to be fooled.
Most firms that lose money on AI make the same mistake. They start from a tool instead of a problem. They see an impressive demo, feel the pressure of falling behind, and buy. Only afterwards do they go looking for where to use it. The right order is exactly the reverse, and the gap between the two orders is the gap between money invested and money thrown away.
I write this as someone who has been building AI systems for six years, since before ChatGPT existed. I saw this technology from the inside when it wasn't fashionable, and I see it now, when everyone is selling something with "AI" on the label. Most businesses don't need more AI. They need more clarity about which piece of AI solves a real problem for them.
Why most firms lose money on AI
The core mistake has a single cause: you start from "which tool should I get" instead of "what is slowing me down". When you start from the tool, you end up justifying a purchase you have already made. You hunt for somewhere to put it. You force a process to fit a solution, rather than choosing the solution that fits your process.
The local context makes the pressure more dangerous, not less. In Romania, AI adoption among small and mid-sized businesses is still very low. Estimates vary: futurebanking.ro reports around 1% of small firms using AI, while a study cited by AOAR on GhidAI.ro puts it at 2.78%. Whatever the exact figure, the message is clear. Almost no one is doing it yet, so you are not as far behind as you are told you are. That gives you time to choose well rather than buy in a panic.
The most expensive tool isn't the one that costs the most. It's the one you bought before you knew what problem it solved.
The one-bottleneck method
The answer isn't a complicated AI strategy. It's a simple method you can run yourself.
Step 1: Find the task that hurts
Look for a single activity in your business that is, at the same time, repetitive, costly, and dull for the people doing it. Standard replies to customers, sorting documents, writing quotes, summarising reports. Don't look for where it would be "cool to have AI". Look for where you lose hours every week on something that doesn't require real human judgement.
Step 2: Test the cheapest thing that already exists
Before you buy anything custom, test what already exists, in its cheapest form. Often, an ordinary subscription to an AI assistant solves 80% of the problem for a few tens of euros a month. Give it the real task, on real data (anonymised if needed), for a week. See how much of the work actually disappears.
Step 3: Keep it only if it pays back inside a month
This is where the decision is made. If the tool saves you more time and money than it costs, within the first month, keep it. If not, drop it without regret. This simple discipline keeps you out of the "but I already paid, let's give it another chance" trap. A tool that doesn't pay back quickly on a concrete task rarely pays back at all.
If you want to go further and ask whether something is worth building from scratch, I've written separately about how to decide whether you should build the app at all.
What's real and what's hype in 2026
Today's AI is very good at some things and poor at others. Confusing the two categories is exactly the ground on which useless products get sold.
| What AI does well today | Where it stays unreliable |
|---|---|
| Writing and rewriting text (quotes, emails, descriptions) | Decisions with legal or financial consequences, unsupervised |
| Summarising and searching across many documents | Exact figures and calculations where errors aren't allowed |
| Answering repetitive customer questions | Understanding your internal context unless you give it to it |
| Classifying and sorting information | Tasks that require human accountability and judgement |
The practical rule: AI is a very fast, very confident assistant that sometimes gets things confidently wrong. It's excellent for what you can verify quickly. It's dangerous for what you don't have the time or expertise to check. Use it where you remain the one who decides.
Buy off-the-shelf, configure, or build custom
There are three roads to using AI, and most businesses need the first, not the last.
| Approach | Cost | Speed | Control | When it makes sense |
|---|---|---|---|---|
| Off-the-shelf (subscription) | Low (tens €/month) | Immediate | Low | Almost always, as a first step |
| Configured on an existing platform | Medium (hundreds €) | Days to weeks | Medium | When off-the-shelf nearly solves it, but not quite |
| Built custom | High (thousands €+) | Weeks to months | High | Only after you've validated the value and need something unique |
The expensive mistake is jumping straight to "built custom" because it sounds serious. You build custom only after you've proven, with a cheap version, that the problem is worth solving, and you need something that doesn't exist off the shelf. If you get there and genuinely want a system built, that's no longer coaching, it's execution. Then you need a team that delivers, like Sapio, the studio I run, or an already-built product, like ai-aflat.ro. The difference between deciding and building is exactly the difference between a Tech Coach and an agency.
The vendor red-flag checklist
A good vendor asks questions before selling. A weak vendor sells before asking. Here's what should put you on guard:
- They make you an offer before understanding how you actually work.
- They talk about revolution and the future but can't show you concretely what changes tomorrow.
- They promise results without accepting a short trial period.
- They can't explain where your data is stored or whether it's used for training.
- The price is a large, fixed sum with no link to a measurable task.
- They push a long contract before proving value on a single problem.
If you tick two or more of these, stop. A serious vendor loses nothing if you ask for a week's trial on a real problem.
How you know it paid off
The only measure that matters is payback. Before you start, note what the chosen task costs you now: hours per week, multiplied by what that time is worth, plus the cost of errors or delays. After a month of real use, compare. If the saving beats the cost of the tool, you made a good decision. If not, you learned cheaply that it wasn't the right road.
Don't measure success in "how modern it sounds". Measure it in hours freed and money kept in the business. AI used well doesn't look like a technology project. It looks like a week of work that suddenly became easier.
The most expensive tool isn't the one that costs the most. It's the one you bought before you knew what problem it solved.
Frequently asked questions
- Do I need to hire a developer to use AI in my business?
- Usually not, at least not to start. Most bottlenecks in a small business can be solved with off-the-shelf AI that needs configuration, not code. You only need a developer once you want something tailored to your workflow, integrated with your systems, and you have already validated that the investment is worth it.
- How much does it cost to get started with AI?
- You can start with tens of euros a month, not thousands. Many useful tools cost between 20 and 100 euros per user per month. The sensible rule is to never pay for something that does not pay for itself in the first month. Real costs only appear when you build something custom, and that decision deserves separate scrutiny.
- Is ChatGPT enough for my business?
- For a lot of writing, summarising and text-analysis work, yes, it is surprisingly enough and cheap. It is not enough when you need AI to work with your internal data securely, integrate with other systems, or run automatically without someone operating it. That is when you move to a dedicated solution.
- How do I protect my company data if I use AI?
- Ask any vendor three things: where the data is stored, whether it is used to train their models, and whether they offer business options for confidential data. Serious business plans do not train on your data. Never put sensitive data into a free consumer tool.
- How do I spot a tool that is just hype?
- Hype promises transformation without asking how you work. A real tool solves a concrete task you can test quickly and measure. If a vendor talks about revolution and the future but cannot show you in five minutes what changes in your business tomorrow, it is hype.