Generative AI for business delivers the best results when it works inside the software you use every day, such as your ERP or CRM, with your company's own data. ChatGPT open in a browser tab helps the individual, but it does not know your customers, your orders or your price lists. Integrated into the ERP, generative AI can answer questions about your data, prepare draft documents, route tickets and read incoming documents, while people remain responsible for whatever goes out to the customer.
Below you will find the concrete uses inside business software, figures on Italian companies and the limits to understand before you start: privacy, wrong answers and costs.
What generative AI is and why it matters to companies
Generative AI is the family of models that produce new content from a written request: text, summaries, translations, code, images. The best-known language models are trained on huge amounts of text, understand natural language and respond fluently.
In a company this ability matters because a lot of office work consists of reading one text and writing another. A customer's email has to become a reply, an enquiry has to become a quote, ten job reports have to become a summary for the technical manager. These are the steps where most time is lost today.
Italian companies are trying it out. According to ISTAT, Italy's national statistics office, in 2025 16.4% of companies with at least 10 employees used at least one artificial intelligence technology and, among these, 59.1% used generative AI for text, speech or images (ISTAT, Imprese e ICT 2025). The same report notes that a third of companies using AI cannot say for what business purpose. Many are therefore using it experimentally, disconnected from their processes. The work to be done is bringing it into those processes.
Generative AI: concrete examples inside an ERP
The uses below are the ones that, in custom projects, give the clearest results. In all of them the AI works on your company's data, not on generic information taken from the web.
An assistant that queries data in natural language
Today, to find out how much a customer has been invoiced in the last quarter compared with the same period last year, someone has to open a report, set the filters or ask the colleague who knows how. With an assistant built into the ERP, the sales manager types the question as they would say it out loud: "Which customers in Tuscany have ordered less than in 2025 over the last six months?".
The system translates the question into a database query, applies the user's permissions (a sales agent sees only their own customers) and returns a table with two lines of commentary. The technical detail that matters is that the numbers come from the database, not from the model. The AI interprets the question and explains the result, but the calculation stays deterministic and verifiable. An assistant that invents a customer's revenue is worse than no assistant at all.
Draft documents: quotes, replies, reports
Another very practical application is preparing drafts. When the sales agent selects the customer and the items, the ERP proposes the cover text for the quote and the standard terms to review before sending. For customer replies, the operator opens the incoming email and finds next to it a reply already written with the actual shipment status, taken from the order. On Friday the technical manager receives a summary of the week's job reports with the anomalies at the top.
In our projects the AI prepares the draft and sending it is always a person's decision. That way you save time without risking a quote with the wrong price.
Ticket classification and routing
If you receive support requests by email, through a website form or a customer portal, someone has to read them, understand what they are about and pass them to the right person. A generative model can do this as they arrive: it assigns a category (fault, accounts, spare parts), estimates urgency, links the request to the correct customer and installation and extracts useful data such as serial number or address.
The technician receives the ticket already sorted. Requests the system cannot classify with confidence go into a manual review queue, where an operator routes them as before.
Reading incoming documents
Supplier invoices, customer orders in PDF, delivery notes, technical data sheets: for extracting data from documents whose layout keeps changing, a generative model is far more flexible than traditional OCR. The ERP receives the fields already filled in and the operator confirms or corrects them.
Contextual help for software users
Every ERP has screens that are opened once a month and that nobody remembers, such as year-end closing or a price list change. An assistant connected to the internal documentation explains step by step how to complete the operation on the screen the user is looking at. Questions to the expert colleague drop and onboarding for new hires gets shorter.
The limits of generative AI to know before you start
There are three limits that need to be handled from the design stage. Vendors who play them down during the sale usually discover them in production.
Privacy and company data
When the ERP calls an AI model, the data sent (customer names, amounts, email text) leaves your system and reaches the model provider. You need to know where it is processed, whether the provider stores it or uses it to train its models and what the contract says. The risk goes down if you send only the data that is strictly necessary and choose services with processing in Europe and clear contractual terms. The GDPR still applies: if you process personal data, the model provider must be assessed like any other data processor.
Among Italian companies that considered AI without adopting it, 43.2% cite concerns about privacy and data protection (ISTAT). Selli holds ISO 27001 certification and addresses these points during the analysis phase, before writing any code.
Hallucinations: confident but wrong answers
Generative models can produce incorrect answers in a convincing tone. These are called "hallucinations", and no serious vendor promises to eliminate them entirely. They are handled through design: the model reads only the company's data and not its generic memory, numbers always come from the database, the user sees the source of the answer and a person confirms everything that goes to the customer. Before release you need tests with real questions, collected from the people who will use the system.
Usage-based costs
Most models are paid for according to the amount of text processed. At small volumes the cost is low, but it grows with use. An assistant used by fifty people all day, or reading thousands of documents a month, needs to be estimated before you start. Costs are kept under control by choosing the right model for each task (classifying a ticket does not need the most powerful one), limiting the text sent and monitoring consumption every month.
There is also a case where generative AI does not pay off at all: if the task follows fixed rules, such as copying a field from one system to another, a traditional integration costs less and makes fewer mistakes.
How we integrate generative AI into custom software
At Selli, generative AI is a feature to add where it is needed in the custom business software we develop, not a separate product. We start from a process and a number to improve: minutes per ticket, time to prepare a quote, documents entered by hand. Then we work in two-week sprints. At the end of each sprint there is a version to try out in staging, which your team uses with real data, and together we see where the AI helps and where it gets things wrong. The code remains your property, including the integration with the models, so you can change AI provider without starting from scratch.
If you are working out where to start, the page on our AI projects for companies describes the services, and the guide to AI for SMEs with examples and a checklist helps you choose the first project. The impact on roles within the company is covered in the article on AI, jobs and intelligent automation, while the examples of AI in large companies show what can be reused on a small scale.
Let's talk about generative AI in your ERP
Generative AI pays off most in processes where people spend their time reading information and rewriting it somewhere else. In a 30-minute call we look at your workflow and tell you whether integrating it makes sense, with what limits and at what order of cost. Book a call with Selli.