Guide — AI in business
In brief
Artificial intelligence pays off in an SME when it works on the data the company already has: documents, ERP, tickets. There are three levels of AI, and for each one the deciding question is where the data ends up. AITAKY runs AI on a server inside the company, with local models, and always starts from a small first version, in production within weeks, with code owned by the client.
Marco Martini · Founder, AITAKY
01 / 09 · Why this guide
Over the last two years every software vendor has added the word "AI" to its price list. The result is that an SME owner receives three proposals a week and no answer to the only question that matters: what, concretely, changes for me?
This guide is our answer. It is not a catalogue: it is what we tell clients when they ask where to start, including the things we advise against.
An honest map
The three levels of AI that actually exist, how autonomous each one is and, above all, where your data ends up in each.
Real numbers
Projects we have delivered, with before-and-after figures. No hypothetical examples.
The questions to ask
Six questions to put to anyone proposing AI to you, us included. If they can't answer, they aren't ready.
"We tell you the truth, even when it costs us a sale."
It is the principle we work by. It applies to AI too.
02 / 09 · What it is, and what it isn't
AI doesn't think. It does statistics very well and very fast.
What it does well
It reads thousands of documents, emails, tickets and ERP rows and puts them in order: it classifies, extracts, summarises, links. The tasks that "someone does by hand on Friday afternoon".
Example: 2,065 products reduced to 893 components in two days, at 95% accuracy.
What it doesn't do
It doesn't know your market, your customers, or why a rule has an exception. It proposes, it doesn't decide. A system that "decides on its own" with nobody checking is a risk, not an advantage.
That is why the AI Act requires human oversight and training: not bureaucracy, common sense.
What really changes
The value isn't "having AI". It is that the person in accounting, in the warehouse or in after-sales stops doing the repetitive part of the job and does the part that requires judgement.
It is measured in hours saved per month and in errors that no longer happen. If it isn't measured, it isn't AI: it's marketing.
An AI model is a program that has learned, from huge numbers of examples, to predict the most likely answer to a request: the next word in a text, the category of a document, the value of a sale. It works well with good examples, badly with examples that are few, dirty, or unrelated to your business.
Generic AI knows everything about everything and nothing about you. Useful AI is the kind that works on your documents, your ERP, your tickets. Which raises the question nobody asks you: where does that data have to go?
03 / 09 · The map
Almost every proposal you receive falls into one of these three levels. The right level depends on the problem, not on fashion. The column that changes everything is the last one.
| Level | What it does | When you need it | Where the data goes |
|---|---|---|---|
| 1 · General assistantsChatGPT, Copilot, Gemini | Writing, summarising, translating, drafting. Useful right away, no project needed. | For individual work. They don't know your company: every time you start from scratch. | Vendor's cloudWhatever you paste leaves the company. You need a written rule on what may be pasted. |
| 2 · AI on your documentssemantic search, RAG | Answers using your manuals, contracts, emails, procedures. "What's the warranty on model X sold in 2023?" and it finds the answer in your files. | When people waste time looking for information the company already has. | Cloud or on-premiseBoth are possible. With sensitive data (customers, patents, health) we do it on-premise. |
| 3 · AI inside processesautomations, agents | Reads a ticket and assigns it. Receives an order by email and enters it in the ERP. Checks a catalogue and flags inconsistencies. Works quietly, every day. | When a repetitive process passes through someone's hands several times a day. | On-premiseIt talks to the ERP and to production data: for us it has to run where the data lives, with tracked access. |
Level 1 you already use, and that's fine: all it needs is an internal rule. The value for an SME lies in levels 2 and 3, and that is where data becomes the issue. We run AI on a server inside the company, with local models: for a manufacturer or a professional firm, "the data leaves" is not a technical detail, it's a contractual risk.
Artificial IntelligenceAitera is the AI appliance we install on site: local models, ERP integrations, no data in the cloud. Anima is semantic search across documents. Adcelera connects AI to the systems you already have, with over 40 connectors. Components we reuse, not black boxes.
Aitera04 / 09 · Where it pays off most
They are not the most spectacular applications. They are the ones with a high volume of repetitive work, data the company already has, and a person who is currently the bottleneck.
01 · After-sales and support
Email, phone calls, WhatsApp: the AI understands what the customer is asking, checks contract coverage, proposes the right technician. Whoever runs field service gets back the half hour a day lost to sorting.
That is what FarDesk does, and we use it ourselves first.
FarDesk02 · Documents and knowledge
Manuals, past quotes, contracts, procedures, standards: a natural-language search across the company's documents, with the source cited. It works for the engineering office as well as for a firm with twenty years of case files.
For an accounting firm we turned figures "chased at year end" into a weekly report, right down to billing.
03 · Dirty data and catalogues
Duplicate records, item codes written five different ways, inconsistent BOMs, price lists in Excel. The job that blocks every ERP project, done in days instead of months, with a person validating.
2,065 products of a furniture e-commerce reduced to 893 components in two days, 95% accuracy.
04 · Reports and decisions
Sales, margins, stock, delays, hours per job: the AI aggregates ERP, e-commerce and scattered spreadsheets, makes them readable and flags what is out of the ordinary. The owner stops asking "can you pull that figure for me?".
For a retail chain with 120 stores one query went from 18 minutes to 12 seconds.
05 / 09 · Real cases
Fashion retail · 120 stores
Data lake with natural-language search, entirely on-premise. Questions that used to require an IT extraction are now asked by the people who sell.
380+ hours a month recovered · 6-week project · 850 employees
Read the caseManufacturing · furniture e-commerce
A catalogue of modular products brought back to its real components, with the AI proposing and a person confirming.
2 days of work · 95% accuracy · clean base for the new ERP
Read the caseAccounting firm · Padova
From logged activities to billing, with the firm's figures available every week instead of at year end.
"We used to chase the numbers at year end." — Studio Bortoletto
See the case studiesA warehouse app built on inconsistent master data, without waiting to "clean everything first": the AI normalises on input, operators fix the exceptions, the data improves through use. First version in production within a few weeks.
None of them started from "we want AI". All of them started from a measurable problem, got a small working first version, and extended it only after seeing it work. It is the only way we know not to waste money.
06 / 09 · How to introduce it
You don't "install" it. You start from a problem and grow in small steps.
The surest way to fail with AI is to buy a platform and then look for something to use it on. The surest way to succeed is to pick a process that hurts today, measure it, and put a first version into production in weeks, not months.
We look together at your processes, the data you have and where it lives. You leave with three candidates ranked by return and an estimate. If the right answer is "don't do AI for now", we tell you.
One process, one metric. In production with the people who will use it, not in a demo. The data stays where it is; if local models are needed, we install them on site.
Hours saved, errors avoided, response times. A simple report, in person, with the people using it. If the number isn't there, we fix it or close it.
The next process, with what we have learned. Each step is a contract of its own: no multi-year commitment, and the source code is yours from day one.
It depends on the process, and we say so before we start, in writing. We work on a fixed-price project, by the day, or on a monthly fee that includes development, hosting and maintenance. The first version is deliberately small: if it doesn't pay off, it stops there and you have spent little.
The AI Act makes it mandatory for anyone using AI systems: we do it together with the go-live, not as a separate course.
Diagnosis: 45 minutes, free
07 / 09 · Data, security, rules
Since 2025, anyone using AI systems must ensure staff literacy and, for higher-risk uses, document how the system works and who supervises it.
In practice: recorded training and a fact sheet for every system in use. Aitera generates it automatically for the systems it manages.
Mandatory AI training: what to doNames, emails, contracts, health or employee data entered into a cloud assistant leave the company and end up under a contract nobody has read.
In practice: a written rule on what may be pasted, and on-premise models when the data belongs to customers or employees.
If you supply companies subject to NIS2 (automotive, food, energy, healthcare, retail) you will receive security questionnaires. An AI that talks to the ERP is inside the perimeter.
In practice: tracked access, backups, and a supplier who can answer the questionnaire on your behalf.
The NIS2 guide for SMEsAI runs where the data lives: a server inside the company with local language models, integrated with the ERP and with the access controls you already have. We use the cloud when it is needed, not out of habit.
Source code, repositories, keys and documentation are yours. No penalty for changing supplier. If one day you want to take everything elsewhere, you can: it is the best guarantee that we will keep earning the work.
08 / 09 · Before you sign
No technical skills required. Precise answers are. Whoever answers vaguely, or changes the subject, is selling you a word, not a result.
Ask to speak with a client of your size who has been using the solution for at least six months. Not a demo: a phone call. It is worth more than any presentation, this guide included.
If the answer is "efficiency" or "innovation", it isn't an answer.
Whose cloud, in which country, under which contract. "In the cloud" is not enough.
It will get it wrong. The question is whether there is a person in the loop and a way to correct it.
If the first version takes six months, you are buying a project, not a result.
Repositories, keys, documentation, penalties. Ask for it in writing.
If the supplier doesn't know what the literacy obligation is, they don't know the law that governs their product.
09 / 09 · Where to start
A 45-minute diagnosis. Free, and with no product to sell you.
We look at your processes and your data and tell you where AI pays off, where it doesn't, and what the first version would cost. If the right answer is to keep what you have, we say so plainly.
FAQ
Yes. Language models today run on a server inside the company, integrated with the ERP and with the existing access controls. It is how AITAKY installs AI (Aitera) when the data belongs to customers, employees or production processes.
It depends on the process. The first version is deliberately small, with the price in writing before we start: fixed-price, by the day or on a monthly fee. If the number isn't there after 30 days, it stops there.
One with a high volume of repetitive work, data already available and a person who is currently the bottleneck: after-sales and tickets, document search, cleaning master data and catalogues, recurring reports.
Literacy for the staff who use AI systems and, for higher-risk uses, documentation of how the system works and who supervises it. Training is already mandatory.
The first version goes into production in weeks, not months, and is measured after 30 days on hours saved, errors avoided and response times.
Further reading
We use technical cookies required for the site to work. Only with your consent do we also use analytics and marketing cookies. Read the Cookie Policy and the Privacy Policy.