The short answer
Artificial intelligence (AI) enters a company in two ways. Top-down: top management sets objectives, rules and budget, then the projects start. Bottom-up: people and departments entrust individual tasks to AI assistants and agents, and what works spreads. The first gives direction but risks staying on paper; the second gives quick but scattered results. What works is the combination: direction and rules from the top, use cases from below.
- Top-down: objectives, rules and budget decided by top management. It is needed for processes that cut across several departments.
- Bottom-up: one task at a time, with quick results and little spending. Without coordination it remains individual use.
- In Italy, just 19% of people who use AI at work use only company tools.
- Some decisions cannot be delegated: objectives, usage rules, responsibilities, spending, preparing people.
- The way that works: direction from the top, projects chosen with the people who do the work, one process at a time.
Note. The timings given are indicative and depend on the systems to be connected; this guide does not give prices. The parts on the rules, European and Italian, are for information only and are no substitute for legal advice.
Two ways of bringing AI into a company
Artificial intelligence (AI) finds its way into a company anyway, whether or not anyone has decided it should. What makes the difference is the point at which it comes in.
From the top, or top-down: the people who run the company decide that AI is a priority and establish where it should deliver results, under what rules and with what spending. The projects follow from those decisions.
From below, or bottom-up: you start from individual tasks. One person uses an assistant to summarise a contract; an office entrusts an AI agent with reading orders in PDF. What works is repeated and extended, one task after another.
Neither route is the right one in absolute terms. They have different advantages and risks, and the companies that get results almost always use them together. This guide compares them for those who have to decide how to organise adoption: business owners, company directors and senior managers.
Top-down adoption: the strategy decided at the top
How it works. Top management, meaning the business owner, the CEO or general management, does four things:
- it chooses two or three company objectives that AI must contribute to: response times to customers, errors, margin, continuity of know-how;
- it identifies the processes to work on, and in what order;
- it assigns a responsibility, a budget and a measurement criterion;
- it sets the rules: permitted tools, data that may be used, checks.
What it brings.
- Direction. The projects serve the company’s objectives, not one department’s curiosity.
- Whole processes. Only top management can change a process that cuts across several offices, such as the path that runs from quote to order and from order to invoice. No department, on its own, has the authority to do so.
- Shared resources. The data connections, the contracts with suppliers and the rules are built once and serve everyone.
- Rules from the start. Confidentiality, security and legal obligations are not something to chase after the event.
Where it gets stuck.
- The strategy stays on paper. Months of analysis, a well-made document, no project.
- The projects are born far from the real work. Those who decide do not see where time is lost every day, and end up choosing what sounds best in a meeting.
- Starting too big. A project that touches everything takes a year to show anything, and in the meantime the technology has changed.
- People have the tool imposed on them. What arrives from the top without having been asked for is used little and badly.
Bottom-up adoption: one task at a time
How it works. It takes two forms, very different from each other.
The first is spontaneous. People discover an AI assistant and start using it to write, translate, summarise, do calculations. Nobody decided it and often nobody knows about it.
The second is guided. You choose a precise, repetitive and measurable task and entrust it to an assistant or an AI agent within a fixed scope: reading incoming orders, filling in a customer or supplier record, preparing a draft quote. Then you move on to the next task.
What it brings.
- Speed. The first result arrives in weeks, not quarters.
- Practicality. You work on problems that someone really has.
- Little risk. If the chosen task does not work, little has been spent.
- Acceptance. Whoever proposed the improvement uses it and stands up for it.
- Discovery. Uses emerge that no analysis done from the top would have imagined.
Where it gets stuck.
- Confidential data out of control. Price lists, contracts and customer data copied into personal tools, with private accounts.
- Results that remain individual. Each person saves a few minutes, but the process stays as it was and nothing shows in the company’s accounts.
- Duplication. Three offices solve the same problem in three ways, with three different tools.
- No measurement. Without a number before and one after, nobody can say what worked.
- Dependence on a few people. When the enthusiast changes jobs, the use stops.
- A ceiling. Until AI is connected to the company’s systems it remains a matter of copying and pasting between windows. To make it work inside the ERP (the company’s management software) you need access rights, investment and decisions that nobody can take from below.
The two approaches compared
| Aspect | From the top (top-down) | From below (bottom-up) |
|---|---|---|
| Who decides | Top management | People and departments |
| Starting point | The company’s objectives | Individual tasks |
| First result | In months | In days or weeks |
| Initial spending | Higher, planned | Low, spread out |
| What changes | Whole processes, including across departments | Single tasks |
| Rules and confidentiality | Defined beforehand | To be chased afterwards |
| Measurement of results | Planned from the start | Rare |
| Main risk | A strategy without projects | Many projects without a strategy |
What the numbers say
The figures for Italy describe an adoption that today happens mostly from below and struggles to become a matter for the company as a whole.
- In Italy, according to ISTAT (the national statistics institute), 16.4% of businesses with at least ten employees used at least one AI technology in 2025, double the share of the year before. Among manufacturing businesses the share is 14.7%, among large ones 53.1%. The European Union average, according to Eurostat, is 20.0%.
- In basic metals and fabricated metal products the gap is wider: 8.8% of Italian businesses use AI, against a European average of 14.3%.
- One business in three among those that say they use AI (33.4%; it was 15.5% in 2024) does not link it to any of the business areas proposed in the questionnaire, such as sales, administration or production. They are mostly small businesses. ISTAT speaks of an adoption that is more widespread but still not very structured, made up of initial and experimental uses.
- According to the Artificial Intelligence Observatory of the Politecnico di Milano (Milan’s technical university), 47% of workers in Italy use AI tools at work, but only 19% of them use company tools alone. The others also use their own tools.
- The time gained stays with individuals. According to the HR Innovation Practice Observatory of the same Politecnico, with AI Italian workers save 30 minutes a day on average, but only 9% of organisations manage that time in a structured way.
- Among large companies in Italy, 71% have started at least one AI project and 84% have bought generative AI licences. Yet only one in five uses AI widely across several functions, only 9% have structured governance and, among those that use ready-made tools, only 54% try to measure the benefit.
- Four large companies in ten have written guidelines on the use of AI; 24% ban tools not provided by the company.
- For Italian small and medium-sized enterprises there is no comparable figure. In an OECD survey of more than 5,000 SMEs in seven countries (Italy is not among them), only 28.6% of those that use generative AI have given guidelines to their staff; according to the OECD, this may mean that the use stems from individual initiative rather than from a strategy decided at the top.
Many people use it, few companies have organised it. This is the gap that the Politecnico’s Artificial Intelligence Observatory points to between individual adoption, by now high, and the transformation of organisations, still limited.
A German study published in January 2026 on 2024 data, carried out by five public research institutes on almost 10,000 employees, measures the same gap. Only about a third of those who use AI at work received their main tool from their employer; the others obtained it themselves. But among those who use a tool introduced by the company almost half use it regularly, against a quarter of the others, and they report more training and more benefits. Bottom-up use arrives sooner and reaches more people; the use organised by the company goes deeper.
People in the company already use it on their own: problem or starting point?
Both. Spontaneous use, what is known as shadow AI, is a bottom-up adoption that nobody governs.
It is a problem for data: anyone who pastes a contract or a price list into an assistant used with a personal account takes them outside the company, under terms of use that nobody has read. And it is a problem for the rules, because the legal obligations fall on the company even when the tool was chosen by a single employee.
It is a starting point because it tells you where AI is really needed. If people in three offices use an assistant to reread quotes, that is the first task to entrust to a company tool, with the data kept safe and a number to measure.
Banning rarely works: the use continues, but out of sight. What works is replacing personal accounts with company tools, giving a few rules and listening to what people are already doing.
The other routes
Between the two extremes there are intermediate forms, which often coexist.
- From the middle. The initiative starts with the heads of function: the sales director, the purchasing manager, the head of production. They know the processes and have the authority to change them within their own department. It works well for processes that sit within a single function, less well for those that cut across more than one.
- The competence centre. A small central group, even just two or three people with a supplier alongside, collects the departments’ requests, chooses the tools, builds the data connections and makes them available to everyone. It is the form usually adopted by larger companies and by groups with several sites.
- The pilot department. You start with a single department or a single plant, get it fully up and running and use it as an example for the others.
- From outside. AI arrives inside the programs the company already uses: email, the office suite, the ERP, the CRM. It requires no project, but no choice either: the priorities are decided by the supplier, you usually pay per user, and what you get is the same as what your competitors get.
The way that works: direction from the top, projects from below
The two routes are not alternatives, because they do different things.
From the top comes what only top management can give: the objectives, the rules, the resources, the authority to change a process that involves several departments.
From below comes what only the people who do the work know: where time is lost, which data can be trusted, which exceptions come up every day, what would really be used.
In between you need a meeting point: someone who collects the departments’ proposals, compares them with the company’s objectives and chooses which one to start with. Then you proceed one process at a time, with a fixed-scope pilot project and a number to improve. Every successful project widens the scope and makes the next step more precise.
It is also the approach of the guidelines of Italy’s Ministry of Labour on AI in the world of work, adopted in December 2025 and non-binding: an initial assessment, then planning and governance with objectives, budget and responsibilities, then pilot projects in limited areas and finally the wider roll-out. Monitoring and training accompany the whole path.
In this scheme the strategy is a set of decisions that the projects put to the test, more than a document to be written beforehand.
When an idea born in a department becomes a company project
A use that started from below deserves to become a project when it has a good answer to four questions.
- Is it repeated often enough? A task that comes up ten times a day is worth more than one that comes up once a month.
- Can it be measured? Minutes per case, errors, days of waiting: you need a starting number.
- What data does it work with? If the data is confidential, the tool and the terms of use are decided by the company, not by whoever had the idea.
- Who is accountable for it and who maintains it? As long as it depends on one person alone it is a habit, not a process.
From that moment the tool has someone responsible for it, is described in two lines that everyone can read and is available to the other departments.
The decisions that fall to top management
Whichever route you choose, some decisions cannot be delegated.
- Why. The two or three results the company wants from AI. Without this answer one project is as good as another.
- Where to start. The first process, and what can wait.
- Under what rules. Which tools are permitted, which data must not go into them, how results are checked, who the internal point of contact is: this is what is written in a company AI policy. The choice between cloud tools and tools installed in the company is part of it: we discuss it in the guide AI in the cloud or on-premise: how to choose.
- Who is accountable. One person with a clear mandate. It does not have to be a technical person: it has to be someone who can talk to all the departments.
- How much, and until when. The spending on the first step and the number that decides whether to take the second.
- How people are prepared. This is also a legal obligation. Article 4 of the AI Act requires those who use artificial intelligence systems to take measures to support the AI literacy of their staff. In Italy, Article 11 of Law No. 132/2025 (Italy’s AI law) obliges the employer to inform workers when it uses fully automated decision-making or monitoring systems that affect recruitment, management of the employment relationship, allocation of tasks, surveillance or evaluation of workers. The details are in the guides AI literacy: what the AI Act requires of companies and AI Act: what a manufacturing company must do.
Everything else is left to the departments, and it is a lot: proposing the tasks to entrust to AI, trying them out, saying what works and what does not.
These decisions fit into a few pages. There is no need for a plan of a hundred.
Who leads, in an SME with no IT department
You do not need a dedicated department. In a small or medium-sized business three roles are enough.
- The people who run the company take the six decisions listed above and review them every three months. It is a few hours, but they cannot be delegated.
- A point of contact for adoption, with a few hours a week: this person collects the proposals, handles relations with suppliers and checks that the rules are followed. Often it is whoever looks after administration (accounts) or organisation, because they see all the departments.
- A contact person in each department: the person colleagues turn to, and who reports proposals and problems.
The technical side, that is data connections, security and development, can be outsourced. The decisions stay in-house.
Which route for which company
Four signals help you understand where to start: size, the confidentiality of the data, the state of the ERP and how used people are to putting forward proposals.
- Small business. The people who run the company know the processes first-hand and decide quickly: top and bottom almost coincide. The limit is the owner’s time. It pays to choose a single process, follow it closely and give a couple of rules on the use of assistants straight away.
- Medium-sized manufacturing company. The offices already use assistants on their own, production has data in the machines and in the ERP, and the processes that matter pass from one department to another. Here direction from the top is indispensable, and the first projects are best looked for where two departments pass information to each other by hand.
- Group with several sites or plants. Without a central point every site reinvents the same things. You need common rules and a competence centre, even a small one. To get started, one plant used as a pilot works well.
- Company with highly confidential data. Drawings, recipes, purchase prices, customer data covered by confidentiality agreements: before any trial you need the rules, so you start from the top.
- Company with a closed or dated ERP. From below you soon get stuck at copy and paste. To go further you need a decision on the systems, which falls to top management.
- Company where many people already use AI on their own. This is the most common situation, and it has a good side: the curiosity is there. You start with a survey of what is being used and turn the best uses into projects.
In the offices of a manufacturing company
When people talk about AI in manufacturing they think of the factory: maintenance, quality control, production scheduling. The first tasks to entrust to an assistant or an agent are more often found in the offices:
- quotes and order confirmations to prepare and check;
- delivery notes, material certificates and technical data sheets to read and file;
- requests for quotation to suppliers to compare;
- orders, price lists and invoices to match against each other;
- reminders and answers to customers’ recurring questions.
For those who buy or process metals and commodities (raw materials) there is one more issue. The cost of metal enters purchasing, price lists, quotes, contracts and the accounts: it is information that runs through the whole company. The useful projects are almost always cross-departmental, such as bringing prices into the ERP or keeping indexation clauses under control, and they require decisions from the top on data and responsibilities.
In all these cases AI sorts, compares and checks information. It does not forecast prices and it does not decide purchases.
How long it takes
The two routes have costs and timings of a different kind: from outside, you pay per user and start within days; a project is quoted on its scope and takes weeks.
| What | Indicative timings |
|---|---|
| Ready-made tools for people | Days |
| Initial analysis of the processes | One or two weeks |
| Pilot project on one process | Four to six weeks |
| Agent integrated into the company’s systems | From a few weeks to a few months (our estimate) |
The timings depend above all on the systems to be connected. We discuss the cost, and what makes it go up or down, in the guide What an AI project costs: what it depends on: to everything you must add the time of the company’s own people, which is the cost most often forgotten.
Two warnings. Licences for everyone cost little per head but, without training and without an objective, they produce little: thirty users are still an expense that recurs every month. And a strategy decided at the top costs nothing as long as it remains a document, but it yields nothing either.
A path in five moves
- Look at what is already happening. A no-blame survey: who uses which tools, to do what, with which data. An anonymous questionnaire or a round of conversations is enough. It is the most useful map you can have, and it costs very little. For a first snapshot of the company’s digital maturity there are also public tools: in Italy, the SELFI4.0 self-assessment and the ZOOM4.0 guided assessment of the Punti Impresa Digitale (PID, the digital support desks of the Chambers of Commerce).
- Give a few rules, straight away. One page: which tools are permitted, which data must never be entered, who to ask in case of doubt. Better a simple rule today than a perfect rulebook in six months. There is an example in the guide Company AI policy: usage rules and an example.
- Provide safe tools. Company accounts in place of personal ones, with contract terms that rule out the use of the data to train the models, and basic training for everyone.
- Collect the proposals and choose one or two. Each department points out the tasks where most time is lost. Senior management chooses on three criteria: frequency, the possibility of measuring, the data available.
- Measure and review every three months. With the measures of the people who run the company: response times to customers, errors, hours given back to people, margin. The number of active licences is not a measure. Top management looks at the numbers and decides what is extended, what is stopped and which process comes next.
The most common mistakes
- Writing the strategy and waiting for something to happen.
- Letting things run, and finding out afterwards where the data ended up.
- Banning everything without offering an alternative.
- Buying licences for everyone without training and without an objective.
- Handing everything to whoever manages the IT systems, as if it were a purely technical matter.
- Mistaking the time saved by individuals for the company’s result.
- Not measuring.
How we work
At FT Support we keep the two routes together. With senior management we define objectives, rules and priorities; with the people who do the work we choose the first process; then we build one step at a time, measuring as we go. This is the path of our AI consulting for companies and of our method.
It is also our own story. The CRM with AI agents that we use every day came about this way: the decision to build it was taken by senior management, and AI entered it one task at a time. The filters written in plain language (in Italian in our case), the customer records filled in starting from the VAT number, the checks on quotes and contracts, the reading of orders in PDF.
Sources
- ISTAT (Italy's national statistics institute), Imprese e ICT, year 2025: report of 15 December 2025 (in Italian)
- Eurostat, 20% of EU enterprises use AI technologies (11 December 2025)
- Eurostat, use of artificial intelligence by economic activity (table isoc_eb_ain2, updated on 15 June 2026)
- Artificial Intelligence Observatory, Politecnico di Milano, press release of February 2026 (in Italian)
- HR Innovation Practice Observatory, Politecnico di Milano, press release of 12 May 2026 (in Italian)
- OECD, Generative AI and the SME Workforce: New Survey Evidence (2025)
- ZEW, Discussion Paper 26-001: Low Barriers, High Stakes. Formal and Informal Diffusion of AI in the Workplace (January 2026)
- Regulation (EU) 2024/1689 (AI Act), text in the Official Journal of the European Union
- Regulation (EU) 2026/1744 of 8 July 2026 (the Digital Omnibus on AI), which rewrote Article 4
- Italy's Law No. 132 of 23 September 2025, Article 11, on Normattiva (in Italian)
- Italy's Ministry of Labour and Social Policies, Ministerial Decree No. 180 of 17 December 2025: guidelines for the implementation of AI in the world of work (in Italian)
- Punto Impresa Digitale (the digital support desks of Italy's Chambers of Commerce), digital maturity assessment tools (in Italian)
Checked on 28 October 2026. Rules and deadlines change: always check the official source before you decide.