![\"Value/complexity](\"/article-assets/intelligenza-artificiale-pmi/intelligenza-artificiale-pmi-casi-uso-ai.jpg\")
In a business context, artificial intelligence is a set of technologies — not a single one — that automate cognitive tasks: reading documents, classifying requests, generating text, recognizing patterns, forecasting demand. It is not a product to buy; it is an integration to design.
\n8.2% of Italian companies with at least 10 employees report using at least one of the seven artificial intelligence technologies surveyed by ISTAT, Italy's national statistics office, in 2024, compared with 5.0% in 2023 [1]. The figure points to plenty of room for adoption, not to a judgment on the companies that are still waiting.
\nThe following sections cover concrete use cases, typical costs, value indicators, and common mistakes.
\nThe term \"artificial intelligence\" is used to mean very different things: from autocorrect to a system that forecasts demand. ISTAT surveys show that fewer than one in ten small and mid-sized Italian companies reported using at least one AI technology in 2024 [1]. Distinguishing between families of technologies is the first step to avoid investing in the wrong use case.
\nWhich tasks is it reasonable to hand over to AI, and which are better left to people? Giving AI what is creative and keeping for people what is repetitive is a common trade-off, and often not a useful one.
\nBefore going further, it helps to set three boundaries that often get blurred in public discussion.
\nClassic automation executes a defined sequence of operations decided by its designer — following fixed, predictable rules. Artificial intelligence learns from data and produces outputs that are not explicitly coded: it recognizes a document it has never seen, answers a request phrased in a new way, estimates the probability of a future event. In many companies it is reasonable to start with classic automation and introduce AI where a fixed rule is not enough.
\nGenerative AI produces new content (text, images, code). Predictive AI estimates probabilities and classifies (the probability of a customer churning, predictive maintenance of a machine). The two families have different costs, constraints, and value indicators: confusing them leads to misjudging both adoption costs and the expected return.
\nOff-the-shelf models are used through an interface or an API without advanced configuration. Custom models require the company's own data and a more structured investment. For most smaller companies, the most common starting point is the first group.
\nThe three main families of AI technologies that apply to a small or mid-sized company are: language technologies (they understand, generate, and summarize text), vision technologies (they recognize images, documents, production defects), and prediction technologies (they estimate demand, churn, credit risk). The operational rule to keep in mind: AI is not a substitute for human judgment, it is an amplifier of cognitive operations that are worth amplifying.
\nAI use cases do not all deliver the same return: some save cognitive time, others reduce errors, others open up new revenue. ISTAT surveys report that, among Italian companies already using artificial intelligence, the most frequent areas of adoption are marketing and sales (35.7%), organization of administrative processes (28.2%), and research and development (24.6%); among the technologies, extracting knowledge and information from text documents (54.5%) and generating written or spoken language (45.3%) [1]. Recognizing the right use case for your company matters more than choosing the tool.
\nIs it better to start with the most visible use case or with the one with the best value-to-risk ratio? The most visible cases attract resources; the most solid cases attract results.
\nFour families of use cases, with concrete examples for a small or mid-sized company:
\nA useful matrix for choosing the use case to start with crosses two dimensions: potential value (time saved, errors avoided, additional revenue) and adoption complexity (technical integration, training required, available data). The cases in the top-left corner of the matrix (high value, low complexity) are the natural candidates for the first project.
\nFor business processes where classic automation may be enough, it is also worth reading how to automate business processes.
\n\n\n
The costs of an AI project in a small or mid-sized company rarely match the license cost of the tool. The items most often underestimated are integration with the systems already in use and training for the people who will have to use the tool every day. Estimating the total costs well prevents projects that stall halfway.
\nWhich cost item is often discovered only once the project is underway? The time of the business owner and the department heads rarely makes it into the initial estimate.
\nThe four cost categories of an AI project:
\nThe \"cost of non-use\" is an invisible but real item: tools purchased and not adopted, active subscriptions with marginal use, hours invested in configuration that was never completed.
\nBank of Italy surveys indicate that adoption remains largely experimental: among the Italian companies that have introduced artificial intelligence tools, only a minority use them intensively, while the others remain in a testing phase or in limited use [2].
\nThe same study observes that the highest adoption rates are found in companies with more structured management practices, which brings the issue back from technology to organization.
\nA survey reported by Il Sole 24 Ore points in the same direction: among Italian small and mid-sized companies that use AI tools, a share close to 42% has not yet seen a significant impact on day-to-day operations [3].
\nAn indicative order of magnitude for an AI pilot project on a use case of medium complexity in a company with 15-50 employees: 3,000-15,000 euros in direct costs (technology + basic integration), plus an equivalent amount in internal person-hours for training and oversight. These ranges are operating assumptions to be validated case by case.
\nA useful AI pilot project, stripped to the essentials, follows a five-step sequence: choosing a low-risk use case, defining success metrics, minimal data preparation, running the pilot within a narrow scope, deciding whether to scale or stop. Skipping one of the steps is a common route to endless pilots that never lead to a decision. Order matters: defining success metrics after the pilot is a common and harmful practice.
\nWhat scope is reasonable for a first AI pilot? Too broad a scope never gets finished; too narrow a scope teaches nothing.
\nThe five steps, with timing and roles:
\nStep 1 — Choose the use case (weeks 1-2). Select a use case with low operational risk, with data already available, and with a measurable effect within 8-12 weeks. The person responsible is the business owner or the operations director. Output: a use-case template describing the current process and the goal of the pilot.
\nStep 2 — Define the success metrics (week 2). Before the pilot, establish which metrics will tell you whether the project is worth scaling. Examples: reducing the average email response time from 4 hours to 1 hour; reducing invoice classification errors from 12% to 3%; increasing the rate of same-day order fulfillment. Output: a shared document with metrics and thresholds.
\nStep 3 — Prepare the data (weeks 2-4). Collect and clean the minimum data the tool needs. For most use cases in smaller companies, this step takes between 20 and 80 hours of internal work. Output: a dataset in a format the chosen tool can use.
\nStep 4 — Run the pilot (weeks 4-10). Launch the pilot within a narrow scope (one department, one process, one customer segment). Monitor the metrics defined in step 2 every week. Involve the people who work in the process in gathering operational feedback. Output: a weekly log of metrics and observations.
\nStep 5 — Decide whether to scale or stop (weeks 10-12). Compare the pilot data with the thresholds you defined. If the thresholds are met, plan the scale-up; if not, analyze why and decide whether to correct the approach or abandon it. Output: a decision document with supporting data.
\nA concrete end-to-end example: automatic classification of incoming emails in a company with 20 employees. In the current process, one person spends 90 minutes a day reading and sorting emails. The pilot uses a general-purpose AI tool to classify emails into three categories (urgent, standard, archive) and suggest how to route them. The success metric is reducing manual sorting time to under 30 minutes a day, while keeping classification accuracy above 85%.
\nMeasuring only adoption (how many people use the tool) is late and ambiguous: use is not value. You need indicators that capture the effect on the process — time saved, errors reduced, output quality, user satisfaction — and that allow before/after comparisons. The distinction between adoption indicators and outcome indicators is central to avoid confusing use with benefit.
\nWhich metrics are reasonable for judging the value of an AI tool after the first 90 days? A dashboard showing the number of daily prompts says nothing about the return on investment.
\nThe three families of indicators, with examples by area:
\nThree concrete examples of indicators by area:
\nTo build a coherent KPI system around AI projects, see the guide to business KPIs.
\nIntroducing AI in a company raises legitimate questions about personal data, trade secrets, and compliance with the GDPR and the European AI Act. Addressing them in advance, in proportion to the use case, avoids both paralysis and carelessness. The operating principle is: first clarify which data can leave the company's perimeter, which cannot, and with which tools.
\nHow much data oversight is \"enough\" for a smaller company? Neither the huge legal templates of multinationals nor the absence of rules — what you need is a documented middle ground.
\nThree operational questions to answer before launching any AI project:
\nWhat data goes into the AI system? Customer data, contracts, financial information, and trade secrets should not be sent to general-purpose AI systems without checking the provider's contractual terms. Many general-purpose tools state that they do not use user data for training, but this must be verified for each specific provider.
\nIs the data processed in Europe? The GDPR requires that the personal data of European citizens be processed in compliance with European rules, even when processed by providers outside Europe. It is worth checking where the provider's servers are located and whether adequate contractual clauses (Standard Contractual Clauses) are in place.
\nDoes the use case fall among the high-risk ones under the AI Act? The European AI Regulation (AI Act, which entered into force in 2024) classifies some uses of AI as high-risk (e.g., AI in hiring, credit, and safety contexts). Additional obligations apply to these uses. Most typical use cases in smaller companies (document classification, customer support, internal data analysis) do not fall into the high-risk categories.
\nA minimum checklist in proportion to a smaller company: check the provider's privacy terms, define an internal policy on the use of AI tools (which data can and cannot be used), and train the people who use the tools on correct behavior. For higher-impact projects, it is reasonable to involve a lawyer who specializes in privacy.
\nFor the change management issues that come with adopting new technologies, see the guide to change management.
\nThe most common mistakes in adopting AI in small and mid-sized Italian companies are not technical, they are about posture: buying the tool before choosing the use case, handing adoption over to a single person, measuring activity rather than value. Recognizing them is more useful than memorizing them, because they show up in different forms depending on the company's digital maturity. Dealing with them takes honesty more than method.
\nWhich mistake is more costly: waiting too long or starting without a plan? Both extremes produce the same outcome — a company that spends energy on technologies that make no difference.
\nThe six most common mistakes, with an operational micro-fix:
\nStarting from the tool instead of the use case. Buying an AI license and then looking for an application produces projects that are not aligned with real processes. Micro-fix: first identify the process you want to improve, then look for the right tool.
\nUnderestimating integration and training. The cost of the tool is visible; the cost of technical integration and training people almost never makes it into the estimate. Micro-fix: mentally triple the license cost to get a more realistic estimate of the project's total cost.
\nDelegating adoption to a single person. When AI is \"Marco's thing\" or \"the IT project\", it does not become an organizational integration but a side note. Micro-fix: involve at least two people from the process from the pilot onward and develop an \"internal champion\" for each department.
\nMeasuring activity instead of outcome. \"We used AI 200 times this week\" is not a value indicator. Micro-fix: define an outcome metric before the pilot (e.g., reduced cycle time) and measure it with a before/after comparison.
\nNot managing the data. AI models produce better results the cleaner and more structured the input data is. Launching an AI project without checking the quality of the available data is one of the most frequent causes of failed pilots. Micro-fix: dedicate an explicit phase to cleaning and structuring the data before the pilot.
\nReplicating use cases from different sectors without adapting them. A successful use case in a large manufacturing company does not automatically transfer to a small service company. Micro-fix: read case studies as inspiration for the use case, not as a blueprint to copy; always adapt to your own specific process.
\nThe operational guidance in this article is based on the evidence available as of 2024 in the Italian and European context. Some conditions limit how far the conclusions can be transferred:
\nThis is an editorial analysis intended for orientation: it does not replace technical and legal assessment for AI adoption projects of significance to the business.
\nWhere should you start if you have never used AI in your company?\nThe most accessible starting point is a use case in document management or in support for written communication, with general-purpose tools that are already mature and inexpensive. Before buying any tool, it is useful to map the process you want to improve and identify the metric that will tell you whether the project worked.
\nHow many people does it take to manage an AI pilot project?\nA pilot project on a single use case can be managed by a team of two people: a process owner (who knows the process to be improved) and a technical or digital lead (who handles the configuration of the tool). A dedicated team is not necessary for low-complexity projects.
\nCan AI replace staff in a small company?\nThe available evidence suggests that in smaller companies AI tends to free up time on repetitive and cognitive tasks, not to replace entire roles. Replacing a role requires AI to be able to carry out all of that role's activities — a condition that is hard to meet in the short term in small companies where activities vary. This is an operating hypothesis that has yet to be validated at scale.
\nDo you need in-house technical skills to use AI?\nFor low-complexity use cases (general-purpose tools, document management, writing support) advanced technical skills are not necessary. For medium- and high-complexity cases (integration with management systems, custom models, predictive analytics on structured data) it is advisable to involve a technical professional, even an external one.
\nHow do you assess whether an AI tool provider is reliable?\nThe main criteria to check: where the data is processed (Europe or outside the EU), what contractual guarantees are offered on privacy, whether the provider states that it uses user data to train its models, how long the product has been on the market, and what references are available.
\nAdopting artificial intelligence in a small or mid-sized company is a path that starts with choosing the right use case — not with buying the tool. The use cases with the best value-to-complexity ratio are concentrated in document management, customer support, and basic predictive analytics. The real cost of an AI project is typically 2-3 times the cost of the technology alone, because it includes integration, training, and oversight.
\nA five-step pilot project — from choosing the use case to deciding whether to scale — makes it possible to contain adoption risk within 8-12 weeks. The value indicators that matter are not those of use (how many people use the tool) but those of outcome (how the process changes). Privacy, security, and compliance issues should be addressed proportionately: neither ignored nor blown out of proportion.
\nThe most common mistake in Italian companies is not choosing the wrong tool, but following the wrong sequence: starting from the tool instead of the process, measuring activity instead of value, delegating adoption to a single person. The fix is within reach: identify the process first, define the outcome metrics, involve people from day one.
\nIn a business, artificial intelligence is not a product to buy but an integration to design: it sets the criteria by which you will decide whether a use case brings value, whether it is worth the business owner's time, and whether the company is ready to measure outcomes instead of activity. Building an adoption path requires choosing the right use case, estimating costs realistically, overseeing the data, and using value indicators.
\nThe thread that ties these steps together is consistency with the way the company already works. When AI is grafted onto unclear processes, it amplifies the confusion instead of reducing it. To frame AI within a coherent operating system, it is also worth reading the guide to business innovation and, on the process side, how to automate business processes.
\nA company that truly adopts artificial intelligence stops experiencing technological pressure as a fad. It chooses where to put its energy, knows the real costs of the project, and knows how to measure the value generated. It is a calmer way of working with more solid results — within reach of organizations of any size, as long as AI remains a tool in the service of processes, not the other way around.
\n[1] ISTAT, \"Imprese e ICT — Anno 2024: l'adozione di intelligenza artificiale nelle imprese italiane\", ISTAT, 2024. Available at: https://www.istat.it/comunicato-stampa/imprese-e-ict-anno-2024/
\n[2] Banca d'Italia, \"La transizione digitale nelle imprese italiane: l'adozione del cloud computing e dell'intelligenza artificiale\", Questioni di Economia e Finanza no. 946, by L. Bencivelli, S. Formai, E. Mattevi and T. Padellini, June 2025. Available at: https://www.bancaditalia.it/pubblicazioni/qef/2025-0946/index.html
\n[3] Il Sole 24 Ore, \"Le Pmi accelerano su digitale ma c'è ancora diffidenza per l'AI\", by G. Brazzioli, July 13, 2026. Available at: https://www.ilsole24ore.com/art/le-pmi-accelerano-digitale-ma-c-e-ancora-diffidenza-l-ai-AJ4eOmH
","headings":[{"level":2,"text":"Defining what artificial intelligence can really do in a company","id":"defining-what-artificial-intelligence-can-really-do-in-a-company"},{"level":2,"text":"Recognizing the AI use cases with concrete value for the business","id":"recognizing-the-ai-use-cases-with-concrete-value-for-the-business"},{"level":2,"text":"Estimating the real costs and adoption time of an AI project","id":"estimating-the-real-costs-and-adoption-time-of-an-ai-project"},{"level":2,"text":"Building an AI pilot project in 5 practical steps","id":"building-an-ai-pilot-project-in-5-practical-steps"},{"level":2,"text":"Choosing indicators of AI value, not just of adoption","id":"choosing-indicators-of-ai-value-not-just-of-adoption"},{"level":2,"text":"Dealing with data, security, and compliance without getting paralyzed","id":"dealing-with-data-security-and-compliance-without-getting-paralyzed"},{"level":2,"text":"Common mistakes in adopting AI in Italian companies","id":"common-mistakes-in-adopting-ai-in-italian-companies"},{"level":2,"text":"Limits and conditions of applicability","id":"limits-and-conditions-of-applicability"},{"level":2,"text":"FAQ","id":"faq"},{"level":2,"text":"Operational summary","id":"operational-summary"},{"level":2,"text":"Conclusion","id":"conclusion"},{"level":2,"text":"Sources and references","id":"sources-and-references"}],"tldr":"Is it better to wait for artificial intelligence to become more mature and accessible, or to start now with a few simple applications and learn along the way? There is no single answer: it depends on the value the technology can bring to the company's specific processes, on the cost of a failed adoption, and on how much time the business owner has to oversee its introduction.","tldrItems":null}