Which AI model suits an SME
How an SME chooses the right AI model without technical expertise: three questions, a half-day test and mistakes to avoid.
Reading time: 10 min
One SMEs which evaluates AI almost always starts from the wrong question: which model is the best. The right question is another, and has an answer that can be reached in half a day without technical skills.
This guide proposes the route, with the reference numbers and errors that cost the most.
In summary
- We need the model cheaper that solves well the task.
- For most SMEs, the economic or intermediate zone.
- API spending is often a few tens of euros per month.
- The major voice is the design work, not consumption.
- Evaluate Who knows the process, not a technician.
- The most expensive mistake is starting from the high end out of caution.
The three questions from which to start
What task do I want to automate? It should be written in a sentence. "We want to use AI" is not a project; "classify incoming emails into six categories" is. If you can't write it in a sentence, the project is not yet defined.
How many times a month does that happen? Under a few hundred, automation rarely pays off. Over a thousand, it begins to make sense. Over tens of thousands, it becomes the difference between making and not doing.
What happens if he gets it wrong? A recognizable and correctable error is manageable. An error that enters the data and is not discovered is a serious problem, and requires controls designed from the beginning.
What It Really Costs
Numbers surprise those who expect important figures. Let's take three scenarios typical of an SME, with about 1,500 input tokens and 300 outputs per case.
| Monthly Volume | Economic belt | Intermediate band |
|---|---|---|
| 2,000 cases | ~1 $ | ~12 $ |
| 10,000 cases | ~5 $ | ~60 $ |
| 50,000 cases | ~27 $ | ~300 $ |
List prices verified on August 19, 2026 for Anthropic and OpenAI and on September 3, 2026 for Google on the Standard plan.
Consumer spending for an SME is almost always modest. The true cost of an AI project lies elsewhere: in the design, integration with existing systems, and in the time of people reviewing outcomes.
This has an important practical consequence: it is not advisable to optimize the price list. It is advisable to optimize the percentage of cases that the system closes well, because it is the one that frees time.
The mistake that costs the most
Faced with an important project, the temptation is to choose the most capable model to be at peace. It is the most common mistake and is expensive on three fronts.
It confuses two different things: the importance of the project and the difficulty of the task. A system that sorts the requests of the customers of an important company still does sorting — a simple task, which an economic model performs well.
Correct approach
- Start from the lowest plausible end
- Climb only on a measured failure
- Distinguish importance from difficulties
- Add a second level on doubtful cases
Expensive approach
- Start from the top for safety
- Never check if less is enough
- Choose based on customer importance
- Using a single model for everything
The test, in five hours
- Write the task in a sentence
- Collect ten real cases with the correct outcome known
- Include the three worst cases that you have
- Add a case that lacks information request
- Start from lower band plausible
- Have them assess who do that job every day
The fourth point is the most important and hardly anyone does. A case where there is no information separates a usable model from a dangerous one: the correct answer is to declare the absence, not produce a plausible value.
The sixth point is the one that makes the proof reliable. Whoever does that job recognizes in two seconds an error that to a technician seems a reasonable answer.
What is needed in addition to the model
The model is the most visible and the least determining part. Four things matter the most.
The papers. A system that responds from a messy document base will produce messy responses, with any model. Sorting those documents is worth more than any technical choice.
The instructions. Precise, not long. Each word is paid for with each call, but an ambiguously described field produces errors that look like model tantrums.
Integration. If the system requires you to open another tool or change habits, adoption collapses in a few weeks. It is the most frequent cause of failure of AI projects in the company, much more than the wrong model.
The controls. A weekly sample of outcomes reviewed by one person. Half an hour that protects the entire investment.
Practical examples
Sorting incoming requests. Emails or messages are classified and routed to the right person. Simple task, clear criteria, recognizable errors. It is the project with the best relationship between result and risk, and it is the one from which it is best to begin.
Extracting data from recurring documents. Invoices, bubbles, orders. High volumes, predictable formats, possible verification with arithmetic controls. Economic range sufficient for ordinary documents, with a second level for those wronged.
Assistant on internal procedures. It answers employees' recurring questions by drawing on company documents. Low volumes, but free time for those who answer the same questions today. The condition is that you state when the answer is not there.
First response to customers. It covers recurring intents and passes the rest to one person. It requires more attention than the other three because it produces outward commitments, but it is also the one with the most visible return.
| Project | Risk | Typical SME volumes | Headband |
|---|---|---|---|
| Sorting | Low | Thousands/Month | Economic |
| Extracting documents | Medium | Thousands/Month | Economic + second level |
| Internal Assistant | Low | Hundreds/month | Economic or intermediate |
| First Customer Response | Middle-high | Thousands/Month | Economic or intermediate |
None of the four require the high end. It is a fact that is worth keeping in mind when reading about top models and important costs: those models serve tasks that an SME rarely faces as a first project.
How not to sell the wrong project
A few useful questions to ask anyone who proposes an AI solution, regardless of who they are.
- Which model band we will use, and why that?
- You have tried a lower bandWith what result?
- What happens when the model does not know answer?
- How do I get to verify an outcome without doing the job again?
- How long of Human Review Does it stay after automation?
- The system lives Inside the Instruments that we already use?
The third and fourth are the most revealing. A proposal that does not have a precise answer about what happens when the model does not know, and how an outcome occurs, describes a system that cannot be trusted —no matter how good the demonstration is.
Mistakes to Avoid
- Starting from the most ambitious project instead of the simpler one.
- Choosing the high end for caution, paying up to fifty times.
- Have outcomes assessed for those who don't do that job.
- Testing only on easy cases and discover problems in production.
- Not planning for checks to sample after starting.
- Building Out of Tools which people already use.
The first mistake is the one that sinks more pilot projects. An ambitious first project is likely to disappoint, and an initial disappointment closes the door on everything else for years. A simple project that works well, on the other hand, creates the confidence and experience to deal with a more difficult one.
The last mistake is what you discover later, when the system works technically and no one uses it. It is worth deciding where he will live functionality before you even choose the model.
How to Review Choice in Time
An AI system is not a project that closes: it is an aging configuration. A revision every six months takes a few hours and almost always produces savings.
| Verification | Why |
|---|---|
| The model in use is still in the catalog | Price lists often change |
| A lower band would suffice | The bands reposition |
| Instructions contain outdated fixes | Costless context |
| Error percentages are stable | Input data changes |
The second line carries more value than all of them: it happens regularly that a lower band reaches in a few months the quality that used to require the upper one. The criterion for deciding when to change —and when to stop —is in when it is best to change the AI model.
How to apply it in the company
For an SME the concrete path lies in five hours and does not require technical skills.
Start by choosing the right project, which is not the most ambitious: a repetitive process, with clear criteria, sufficient volumes to justify automation and recognizable errors. An ambitious first project is likely to disappoint, and an initial disappointment closes the door on everything else for years.
Then write the task in a sentence, collect ten real cases with the known outcome —three of which are difficult and one that lacks the required information.and start from the lowest plausible end. Have those who do that work every day assess their outcomes.
Four elements matter more than the model. The documents: a system that responds from a disorderly basis will produce disorderly responses. The instructions: precise, not long. The integration: if it requires you to open another tool, the adoption collapses in a few weeks, and is the most frequent cause of failure. The controls: a weekly sample reviewed by a person.
On the economic level, the calculation that is needed is the cost per completed task, not the list price, according to the method of cost per task or price per token. And the criterion for figuring out whether an economic belt —which in most cases SMEs suffice —is in when an economic model is enough.
Finally, schedule a revision every six months: is the model still in the catalog? Would a lower band be enough? Does the prompt contain any now useless instructions? It is three hours that often produce a recurring saving, with the criterion of when it is best to change the AI model.
If you want to go from testing to a truly integrated AI system in your company's processes, Giallo Studio can help you turn the problem into a real workflow.
To complete the reasoning. The method for building the test bench is in Mixed model strategy: light and powerful. The calculation of cost per completed task, which is the number on which you really decide, is in how to read AI benchmarks. To understand when it is advisable to question the configuration, the criterion is in because vendor benchmarks are not comparable.
Conclusion
For an SME the right model is the cheapest one that solves the task well, and in most cases it is lower than expected. Consumer spending is modest; the real cost lies in design and integration.
The route takes five hours: define the task in a sentence, collect ten real cases, try starting from the bottom and have the results evaluated to those who know the work.
If you want to go from testing to a truly integrated AI system in your company's processes, Giallo Studio designs agents and automations tailored to SMEs.
Giallo Studio guides: methodology and sources
How we apply it in the company
Related resources
FAQ
Which AI model suits an SME?
The cheapest model that performs the specific task well. In most business cases — classification, extraction and answering questions from documents — a low-cost or mid-tier model is sufficient.
How much does it cost to start an AI project in an SME?
For typical SME volumes, API spending is on the order of a few tens or hundreds of euros per month. The largest cost is almost always design and integration work, rather than usage.
Do you need an in-house technician to choose?
Not to choose the model: you need someone who knows the process. Results should be evaluated by the people who do that work every day, rather than by someone who knows AI.
Where should we start?
From a repetitive process, with clear criteria and sufficient volumes to justify automation, where an error is recognizable and correctable.
What is the most common mistake?
Choosing the most capable model as a precaution. It can cost up to fifty times as much, is slower and does not produce better results on simple tasks.
How long does it take to decide?
About five hours: define the task, collect ten real cases, run a trial, have the results evaluated and analyse the errors.