Artificial intelligence in business creates value when it moves beyond the pilot phase: it requires measurable use cases, ready-to-use data, widespread expertise, governance that complies with regulations, and an adequate infrastructure.
Key points:
- Global spending on AI is projected to reach $2.7 trillion in 2026
- In Italy, 71% of large companies have active projects, but only one in five has rolled them out across the entire organization
- Projects come to a halt due to vague objectives, incomplete data, and costs spiraling out of control
- Skills and governance are just as important as technology
- A structured adoption process leads from the pilot project to full-scale implementation
Over the past two years, nearly all large Italian organizations have launched at least one artificial intelligence pilot project: virtual assistants for employees, chatbots for customer service, and predictive models for equipment maintenance. There is no shortage of resources. The challenge lies in the ability to move these projects beyond the pilot phase. Many initiatives remain isolated, fail to reach the production environment, or generate benefits that are difficult to translate into measurable economic results. The question today is no longer whether to adopt AI, but how to transform a series of limited trials into a structural capability within the organization. The global landscape, areas of value, competencies, rules, technological foundations, and a phased adoption path are the key elements driving this transformation.
The Global AI Landscape and Italy’s Position
Global spending on AI is expected to grow by nearly 50% in 2026, driven by infrastructure. Italy is picking up the pace: by 2025, the market had reached 1.8 billion euros, and 71% of large companies had active projects.
Artificial intelligence is now the leading driver of growth in global technology spending. The United States and China account for the largest share of investment, computing power, and the development of foundational models. The European Union has responded with a dedicated plan. The InvestAI initiative aims to mobilize 200 billion euros, with a portion earmarked for large shared computing hubs. The competition also extends to chips, energy, data centers, and networks.
AI (Artificial Intelligence): a set of software techniques that analyze large volumes of data to generate predictions, classifications, content, or actions. In business, it is used to support decision-making and automate tasks.
According to Gartner (AI Spending Forecast, September 2026), global spending on AI will reach $2,700 billion in 2026, representing a 49.5% increase. More than half of the total—approximately $1,484 billion—goes toward infrastructure.
Major economic regions are approaching the AI race with different priorities:
- United States: Leading the Way in Investments, Basic Models, and Cloud Platforms
- China: Focusing on Technological Self-Reliance, Open Models, and Industrial Applications
- European Union: Combines public investment, shared computing hubs, and a common regulatory framework
- Italy: Growth Exceeds the Historical Average, with a Wide Gap Between Large Companies and the Rest of the Business Sector
How far along are large Italian companies in adopting AI?
Almost all large Italian companies have moved beyond the exploratory phase, but few have made AI a widespread capability. ISTAT data show that the percentage of companies with at least 250 employees using AI rose from 32.5% in 2024 to 53.1% in 2025.
According to the Digital Innovation Observatories at the Politecnico di Milano (February 2026), 71% of large companies have at least one AI project, but only 20% have AI deployed across multiple business functions. The Italian market is worth 1.8 billion euros, up 50%.
Where AI Creates Value in Large Organizations
AI creates value in areas where processes are repetitive, data-intensive, and tied to a key performance indicator: customer service, marketing, IT and security, finance, manufacturing and the supply chain, as well as knowledge management.
There are many areas where artificial intelligence is used in a business, but its value lies primarily where there is a measurable process. An assistant that reduces response times lowers the cost per contact. A model that predicts a failure reduces downtime. The starting point is the process, not the technology.
The areas in which artificial intelligence is most widely used today are:
- Customer Operations: Virtual Assistants, Request Summaries, Ticket Routing
- IT and Cybersecurity: Log Analysis, Anomaly Detection, and Operator Support
- Finance and Procurement: Reviewing Documents, Consistency Checks, Cash Forecasts
- Manufacturing and Supply Chain: Predictive Maintenance, Visual Quality Control, Workplace Safety, Demand Planning
- Marketing: lead generation, content personalization, campaign performance analysis
- Knowledge management: searching document archives, support for legal and HR departments
A concrete example is workplace safety. Cosafety AI combines the AI-based video analytics software developed by Italtel, which runs on Retelit’s cloud infrastructure. The result is a ready-to-use platform for factories and construction sites that transforms video surveillance into a tool for protecting people. Its value depends as much on the model as on the environment that makes it reliable and scalable.
The three types of AI have different levels of maturity, risks, and payback periods. Comparing them helps you decide where to start.
What type of AI should you choose to get started?
Predictive, generative, and agent-based approaches address different needs. Comparing them helps us understand what each one does, what role people still play, and where to start.
AI predittiva, generativa e agentica a confronto
| Criterio | AI predittiva | AI generativa | AI agentica |
|---|---|---|---|
| Cosa fa | Stima cosa potrebbe accadere, partendo da dati passati e presenti | Crea contenuti come testi, immagini e codice su richiesta | Riceve un obiettivo e compie le azioni per raggiungerlo |
| Come lavora | Analizza serie storiche e segnali | Risponde a un input umano | Pianifica i passaggi e usa strumenti e sistemi aziendali |
| Ruolo delle persone | Usano le previsioni per decidere | Validano il risultato prima dell'uso | Fissano obiettivi, limiti di autonomia e controlli |
| Casi d'uso tipici | Previsione di vendite e domanda, rischio di abbandono dei clienti, frodi, guasti | Sintesi di report, analisi di documenti, supporto al codice, materiali formativi | Gestione dei ticket, monitoraggio dei rischi finanziari, workflow di supply chain, alert IT |
| Dati richiesti | Dati storici strutturati | Documenti e basi di conoscenza | Accesso a sistemi e API aziendali |
| Maturità | Alta, in uso da anni | Media, diffusione rapida | Iniziale, pochi casi su larga scala |
| Rischio principale | Qualità e deriva dei dati | Risposte errate, riservatezza | Azioni non controllate, costi |
| Tempo al valore | Medio | Breve per i casi semplici | Lungo, richiede integrazione |
What is the difference between generative AI and agent-based AI in a business setting?
Generative AI responds; agentic AI acts. Generative AI creates content on demand but does not perform actions, even though it can be a component of an agent-based system. Agent-based AI is given a goal and coordinates tools and systems to achieve it: it opens a ticket, queries a management system, or prepares an order. This introduces a new responsibility: defining the boundaries of autonomy.
According to Gartner (August 2025), by the end of 2026, 40% of enterprise applications will incorporate AI agents dedicated to specific tasks, compared with less than 5% in 2025.
What is the difference between generative AI and agent-based AI in a business setting?
Generative AI produces content on demand, while agent-based AI decides on and carries out a sequence of actions to achieve a goal. The former answers a question. The latter opens a ticket, queries a management system, prepares an order, and requests confirmation. Autonomy increases the potential value, but also the need for oversight.
From Pilot Project to Production: Critical Factors for Scaling AI
Projects stall due to unmeasurable objectives, data that isn't ready, underestimated operating costs, difficulties integrating with existing systems, and the absence of a business owner.
A pilot, or proof of concept, operates in a controlled environment: few users, selected data, and no integration with actual processes. Bringing artificial intelligence into the enterprise—from the pilot phase to production—changes the rules of the game. The system must handle thousands of requests, comply with security policies, interface with ERP and CRM systems, and remain reliable over time. This is where the costs and constraints that weren’t apparent during the pilot phase come to light. It’s often discovered too late that the use case wasn’t a priority, or that no one on the business side was accountable for it.
According to Gartner (June 2025), more than 40% of agent-based AI projects will be canceled by the end of 2027 due to rising costs, unclear business value, or inadequate risk controls.
There are five main reasons that most often prevent a project from moving from the pilot phase to production:
- Vague goals: A pilot is meant to explore technology, not to improve a specific metric
- Data Not Ready: Fragmented sources, variable quality, undefined access permissions
- Operating costs: The cost of inference increases with use and must be estimated beforehand, not afterward
- Integration: Legacy systems require interfaces and controls that the pilot did not anticipate
- Accountability: Without a business owner, the project remains an IT experiment
How do you measure the ROI of artificial intelligence?
The ROI of artificial intelligence is measured by comparing the value generated by a process with the total cost of the system over its entire lifecycle. The value should be expressed in metrics already used by the company: time to process a case, cost per contact, avoided downtime, and incremental revenue. The cost includes licenses or models, inference, infrastructure, integration, security, training, and ongoing management. Measurement begins before the pilot phase, with a clear baseline.
ROI (Return on Investment): the ratio of net profit to the cost of an investment. For AI projects, it is calculated over a 24- to 36-month period.
An AI project goes into production only if it has a key performance indicator, a business owner, and an estimate of operating costs defined prior to the pilot.
People and Skills: The Decisive Factor in Adoption
Skills are the primary barrier to adoption. We need role-specific training, clear guidelines on how to use the tools, and a collaborative model in which people remain in charge of their own decisions.
The technology is available. The skills needed to make the most of it still need to be developed. According to ISTAT, the lack of skills is the top obstacle cited by companies evaluating AI, mentioned by 58.6%. This is followed by regulatory uncertainty and data quality. There is also a phenomenon that directly affects IT: shadow AI. Researchers at the Politecnico di Milano note that 81% of workers who use AI also rely on tools not provided by their company. This poses a risk to confidentiality and signals a demand that needs to be addressed.
Shadow AI: the use of artificial intelligence tools not authorized by the organization, often via personal accounts, in violation of IT and security policies.
An effective skills development plan consists of four levels:
- Basic Literacy: For All Employees—On Limits, Risks, and Proper Use
- Role-Based Training: Specific Programs for Functions Such as Customer Care, Finance, Legal, and IT
- Specialized expertise: data, security, integration, and model management
- Internal Certifications: Training Programs with Final Exams to Assess Competency Levels
- Change Management: New Processes, New Roles, and Adoption Metrics
How should you organize employee training on artificial intelligence?
Employee training on artificial intelligence is effective when it is tied to the processes specific to each role and the tools the company actually makes available. It’s best to start with the teams involved in the first use cases and measure adoption over time, while keeping human oversight central.
According to the Capgemini Research Institute (July 2025), among 1,500 managers at companies with more than $1 billion in revenue, confidence in fully autonomous AI agents fell from 43% to 27% in one year. Only 2% of organizations are using these agents on a large scale.
How is the organizational model changing with artificial intelligence?
With AI, the organizational model shifts from roles to skills. In a skills-based approach, skills become visible and transferable, teams are formed around priorities, and collaboration transcends functional boundaries. AI supports this transition: it processes data, automates repetitive tasks, organizes information, and helps build scenarios. Tasks are assigned to those with the most suitable skills, and career paths also become cross-functional.
The way we lead people is also changing:
- From Control to Orchestration
- From Individual Knowledge to Shared Knowledge
- From task supervision to employee development
Curiosity and a willingness to experiment become drivers of change: the success of adoption depends on the development of new skills.
Governance, Rules, and Risks: the AI Act and Law No. 132/2025
Corporate AI governance defines who approves, oversees, and is accountable for AI systems. The European AI Act and Italian Law 132/2025 establish requirements with deadlines spread out between 2025 and 2028.
The AI Act classifies systems by risk level and assigns different obligations to their developers and users. In July 2026, the AI Omnibus came into effect, revising the timeline for the most onerous obligations. In Italy, Law 132/2025, effective as of October 10, 2025, adds national regulations. Among these is the requirement for employers to inform employees when they use AI systems.
AI Act (EU Regulation 2024/1689): a European regulation governing the development and use of artificial intelligence systems based on a risk-based approach, with increasing obligations for high-risk systems.
Following the AI Omnibus, the AI Act’s implementation schedule includes four key deadlines:
- Effective August 2, 2025: Requirements for General-Purpose AI (GPAI) Models
- Effective August 2, 2026: Transparency Requirements for User-Generated Content and Interactions with AI Systems
- Effective December 2, 2027: Requirements for high-risk systems listed in Annex III, such as staff selection and credit assessment
- Effective August 2, 2028: Requirements for high-risk AI integrated into regulated products
What obligations does the AI Act impose on companies that use AI?
Companies that use AI must know which systems they are using, what risk level they fall under, and what obligations arise from this. High-risk systems require human supervision, log recording, use in accordance with the provider’s instructions, and notification of affected individuals. Transparency and support for staff training remain essential for all systems.
According to the European Commission, the AI Omnibus has been in effect since July 27, 2026. The postponement buys time, but it does not eliminate the obligations. A robust governance model comprises four elements: a committee that approves use cases; a registry of AI systems in use; a risk assessment prior to release; Rules on data sovereignty and localization. The framework complements existing obligations for many organizations, such as the NIS2 Directive (Network and Information Security 2) and, in the financial sector, the DORA Regulation (Digital Operational Resilience Act) on ICT providers.
The Technological Foundation: Data, the Cloud, and the Network
AI in production requires accessible and governed data, computing capacity scaled to the workload, and a network capable of handling large volumes with controlled latency and security.
An enterprise AI project is only as valuable as the data it receives and the infrastructure that supports it. Decisions must be made early on. Where should models be trained and run? How should on-premises systems and the cloud be connected? How should data be protected? It’s no coincidence that infrastructure accounts for more than half of the global AI spending estimated by Gartner.
The technological foundation of a production-ready AI project rests on four components:
- Data platform: integrated sources, quality-controlled data, defined access permissions
- Computing capacity: public cloud, dedicated cloud, or data center colocation, based on costs, volumes, and sovereignty requirements
- Connectivity: Dedicated connections to the cloud and data centers, with bandwidth and latency guaranteed by an SLA (Service Level Agreement)
- Security: Protection of models, data, and access throughout the entire chain
The network is often the last constraint to emerge. A network infrastructure for AI must provide symmetrical bandwidth, low latency, and controlled data paths between offices, data centers, and the cloud, with different requirements for training and inference.
Getting Started: The Main Steps in the Adoption Process
An effective process begins with selecting use cases, preparing data and governance, validating a pilot with measurable objectives, scaling up the solution, and then expanding it while continuously measuring its performance.
There is no one-size-fits-all sequence, but organizations that bring AI into production generally follow the same pattern. Each phase has a verifiable outcome and a decision to proceed or stop. In this way, artificial intelligence within a company grows as a coherent portfolio, without spreading the budget too thinly across too many parallel experiments.
The adoption process consists of several main steps:
- Assessment and Priorities: Map Processes, Estimate Value and Feasibility, Select Two or Three Use Cases
- Foundations: Data Preparation, Security, Usage Policies, and Governance Roles
- Pilot with clear objectives: Define baselines, metrics, and success thresholds before getting started
- Industrialization: Integrate with business systems, monitor quality and costs, and manage the model lifecycle
- Scale and measurement: Extend to other functions and report the results continuously
For use cases involving interactions with customers and citizens, a good starting point is AI-based chatbot solutions and the development of customized virtual assistants, which allow you to measure the impact on service time and costs within a few weeks.
To generate value over time, the organization must focus on a limited portfolio of measurable use cases, supported by shared data, governance, and infrastructure, rather than multiplying isolated experiments.
Frequently Asked Questions About Artificial Intelligence in Business for Enterprise Organizations
The process begins with mapping the processes and selecting a few use cases with clear economic value. Next, usage policies, reliable data, and a pilot project with predefined metrics are needed before scaling out the solution.
Common causes include unmeasurable objectives, fragmented data, underestimated operating costs, difficulties integrating with existing systems, and the absence of a business manager. Production requires reliability and economic sustainability based on actual volumes.
It depends on the project. Use cases involving generative AI for simple processes show results within a few months. Predictive and agent-based projects integrated with business systems typically take 18–36 months, starting from a measured baseline.
An AI strategy works when it combines technological expertise with business objectives. Coordination is typically handled by a cross-functional committee that approves use cases, monitors risks, and measures results.
Effective July 27, 2026, the AI Omnibus Act postpones the requirements for high-risk systems listed in Annex III until December 2, 2027, and for AI integrated into regulated products until August 2, 2028. Transparency and modeling requirements for general-purpose AI remain in effect.
In most cases, it makes sense to start with off-the-shelf models that have been adapted to the company's data. Developing custom models makes sense when dealing with highly specific data, large volumes, or strict requirements for data sovereignty and confidentiality.
With a contribution from Paolo Annoni, Digital Marketing and AI Adoption, Retelit