Method and data
Business process automation with AI: where to start
16.4% of Italian enterprises use AI, but only 2.9% use it to automate workflows. The gap tells you the right question is not "which tool" but "which process". Here we explain how to choose it, how to build the first automation and what to measure, with the numbers and with examples from our own projects.
Written by Maximilian Giurastante. . 13 minute read.
Automating a business process with AI means handing a language model the steps that call for reading, classifying or writing (reading an email, pulling the data out of a document, preparing a draft) inside a flow that stays rule based and stays under human control. It does not replace the ERP: it connects to the existing systems through APIs and works on the repetitive cases, leaving the exceptions to people.
In short
- An AI automation puts a language model into the reading, classifying and writing steps of a flow that stays rule based: that is the shape that holds up in production and that you can keep under control.
- In Italy 16.4% of enterprises with 10 or more employees use AI (ISTAT, 2025), twice the year before; large enterprises are at 53.1%, small ones at 14.2%. AI workflow automation sits at 2.9% of the total.
- A process is a good fit if it is repetitive, has an unstructured input, has rules you can write down, has an error you can check and a volume that pays the work back.
- Pick the first process small and measure it first: time per case, cases per month, errors. Without the before there is no after.
- Order of magnitude for automating one process: from 1,500 to 6,000 euros and from one to three weeks, plus the model calls on a pay per use basis.
- The most common mistakes: starting from the tool, automating a broken process, leaving out the human check, putting personal data into tools with no contract behind them.
What this means in practice
A business process is a sequence of steps that turns an input into a result: a request into a quote, an invoice into a bookkeeping entry, a report into a service call. Automating it with AI means finding the steps that today need a person because the input is not structured, and handing those to a model inside a flow that follows rules the rest of the time.
What sets this apart from classic automation is the input. Rule based automation (including RPA, robotic process automation) works well on structured data: a form, a table, a file with known columns. It stops in front of an email written by hand, a PDF laid out differently every time, a request in plain language. A language model reads those inputs and turns them into structured data that the rules can then use. The model is one step of the flow, not the flow.
If instead the flow needs the model to decide the sequence of steps on its own, picking between different tools depending on the case, then you are talking about an agent. That is a different category, with different costs and risks, and we wrote about it in AI agents for business. Most processes do not need one.
The numbers: where Italian enterprises stand
The official 2025 figures describe adoption that doubles every year but stays concentrated in large enterprises and in the simplest functions. Anyone automating a process today is not following the crowd: they are ahead of it.
Italian enterprises using AI, by employee size class (2025)
Bar chart: 14.2% of enterprises with 10 to 49 employees, 27.6% of those with 50 to 249 employees, 53.1% of those with 250 or more employees, 16.4% overall among enterprises with 10 or more employees.
- From 10 to 49 employees: 14.2%
- From 50 to 249 employees: 27.6%
- 250 employees and over: 53.1%
- All enterprises with 10 or more employees: 16.4%
What the enterprises that use AI actually do with it
| Technology or purpose | Italy, share of the enterprises using AI | European Union, share of all enterprises |
|---|---|---|
| Extracting information from text | 70.8% | 11.8% |
| Generating text, images, video or audio | 59.1% | 9.6% images and media, 8.8% text |
| Speech recognition | 41.3% | 7.2% |
| Machine learning for data analysis | 20.0% | 5.1% |
| Workflow automation (RPA with AI) | about 18% | 5.4% |
| Purpose: marketing and sales | 33.1% | 34.7% among AI users |
| Purpose: administrative processes | 25.7% | 31.1% among AI users |
The two columns rest on different bases and do not compare with each other: ISTAT gives the technologies as a share of the enterprises that already use AI, Eurostat as a share of all enterprises. Read one at a time they say the same thing: the most widespread use is reading and writing text, workflow automation comes last. Across all Italian enterprises, AI workflow automation reaches 2.9% (it was 2.3% in 2024). ISTAT adds one figure that explains a lot: a third of the enterprises that say they use AI report no business purpose at all, and 83% of those have between 10 and 49 employees. Adoption, not yet process.
- €1.8 bn
- the AI market in Italy in 2025, up 50% on 2024 (Osservatorio Artificial Intelligence, Politecnico di Milano, February 2026)
- 71% / 15% / 7%
- large, medium and small enterprises running at least one AI pilot (same source; SMEs account for 18% of the market)
- 20.0%
- EU enterprises using AI in 2025, up from 13.5% in 2024 (Eurostat, December 2025; Digital Decade target for 2030: 75% with cloud, big data or AI)
- 58%
- of a knowledge worker's day spent on "work about work" (Asana, Anatomy of Work 2023, 9,615 respondents in six countries)
Which processes fit, and which do not
Two criteria are enough for a first pass: how structured the input is and how much volume goes through. The input axis decides whether you need a model at all; the volume axis decides whether the work pays for itself.
Where AI pays, where rules are enough, where nothing pays
Two by two matrix. Horizontal axis: volume, from low to high. Vertical axis: input, from structured to unstructured. Low volume and structured input: do not automate, a spreadsheet or the ERP is enough. High volume and structured input: rule based automation or RPA. Low volume and unstructured input: an assistant that prepares the draft and a person who decides. High volume and unstructured input: AI inside the flow, with spot checks.
- Volume of cases: low: a few a month / high: dozens a day. Type of input: structured: forms, tables / unstructured: email, PDF, voice.
- Assistant plus person: the model prepares the draft or the extraction, a person decides. Low volume, but every case eats time.
- AI inside the flow: the model reads and structures, the rules do the rest, a person runs spot checks. This is the case that pays back first.
- Do not automate: a spreadsheet or the ERP is enough. The project would cost more than the time it saves.
- Rules or RPA: no model: a rule based flow written once and tested on real data.
The five signs of a process that fits
- It is repetitive. It happens every day or every week, always the same way, and whoever does it describes it with a sigh.
- The input is not structured. Email, PDFs, messages, transcripts. If the input is already a table, rules are enough.
- The rules can be written down. Whoever does the job can say when a case goes one way and when it goes another. If they cannot say it, fix the process first.
- The error shows up and can be corrected. A bad extraction in a ticket gets noticed and put right; a wrong payment does not. The cost of an error decides how much checking you need.
- The volume pays back. Thirty cases a day at ten minutes each is five hours a day. Three cases a month do not even justify the analysis.
The processes we see most often, and what sits inside them
| Process | Step handed to the model | Stays with the rules | Stays with people |
|---|---|---|---|
| Incoming support requests | Reading, classifying by type and urgency, extracting the data | Opening the ticket, assigning it, answering the known cases | Uncertain cases, replies that commit to something |
| Supplier documents (invoices, delivery notes, quotes) | Extracting fields and line items from different formats | Matching against the order, bookkeeping entry | Discrepancies above threshold |
| Recurring quotes | Reading the request and proposing the configuration | Pricing from the price lists, generating the document | Approval and sending |
| Periodic reports | Writing the commentary on the aggregated figures | Data extraction, scheduled delivery | Reading and decisions |
| Content to publish | Drafts, summaries, adaptations per channel | Scheduling, formats, publishing | Review and editorial responsibility |
Three examples from our projects, including the ones without AI
The most useful part of an example is knowing where the model earns its place and where it does not. In two cases out of three, the decisive step was rule based.
The emails that turn into tickets
In SYNAPSE the reports that arrive from an external system come in a stable format: a rule based flow turns them into jobs, spots the follow ups on a case already open and creates no duplicates. No model, because on an input that is already structured it would cost more and be less reliable. A model would only come into play for reports people write by hand, as a single extraction step inside the same flow, with a confidence score that decides when a person has to check.
The spreadsheets from three countries that turn into an archive
In Omnia Leads Manager the automation is the import: Excel files with sixteen columns and different local formats, checked row by row, with duplicates caught and notifications going out on their own. No model here either, because the columns are known. The time saved is the manual cleanup nobody used to do, which is why the archive did not exist.
The monthly report written by the model, on the numbers and not the names
In LeavePilot the absence report carries a commentary written by a language model. The model only receives figures aggregated by department, never the names: it is the most effective rule for staying inside GDPR and the AI Act, because a model that never receives a piece of data cannot use it. This is a case where the model really does earn its place, because writing a sensible commentary on numbers is a writing step, not a calculation.
How it is done, in six steps
The method is the same for a company of ten people and for one of two hundred: one process at a time, measured before and after, with a person checking until the numbers say they can stop.
- Map the process and measure it as it is Who does what, with which tools, how long per case, how many cases a month, how many errors. Two weeks of counting by hand is enough. Without that number you will never be able to say whether the automation worked.
- Pick a process that is small and annoying Not the most important one: the most repetitive, with the least structured input and the most recoverable error. The first project is there to teach you how working with a model goes, so pick one where a mistake costs nothing.
- Write down the input, the output and the exception cases What goes in, what has to come out, and the list of cases where the flow has to stop and hand over to a person. Thirty real cases with the expected result become the set you test every change against.
- Build the flow with a human check in it The model does its step, the rules do the rest, and a person confirms before anything has an effect outside the company. The flow connects to the existing systems through APIs or exports: nothing gets replaced.
- Measure for a month Cases handled without intervention, time per case, errors caught in the check, cost of the model calls. Compare them with the numbers from step one. If they hold, the check goes from every case to spot checks.
- Extend, and get the paperwork in order You choose the second process on the same criteria. In the meantime you document: anyone using AI in a company has to be able to show AI literacy among staff (Article 4 of the AI Act) and, if the flow interacts with people outside the company, to declare it (Article 50, from 2 August 2026). Personal data going into the prompts calls for a contract with the model provider. Our AI Act guide (in Italian) lists the documents to produce.
What it costs and how to work out whether it pays
An automation has a build cost, a running cost and a return you count in hours. You do the sums before you start, with the numbers from step one, and you do them again after a month with the real ones.
| Item | Order of magnitude | What pushes it up |
|---|---|---|
| Building an automation on one narrow process | from 1,500 to 6,000 euros, from one to three weeks | Number of systems to connect, volume of data, need for a human check on every step |
| Model calls | pay per use: from a few euros to a few tens of euros a month at SME volumes | Length of the documents, number of cases, model chosen |
| Infrastructure | from a few tens of euros a month | Queues, run logging, separate environments |
| Maintenance | inside the agreed retainer or billed as used | Changes in the connected systems, new exception cases |
These orders of magnitude are the ones on our how we work and what it costs page. For the service in detail, from the analysis to the release, see the AI automation page.
The mistakes we see most often
There are five of them and they come back every time. None is technical: they are all mistakes of choice or of method, and you avoid them before writing a line of code.
- Starting from the tool. "We want to use AI" is not a process. You start from a list of repetitive tasks with the times written next to them, and the tool comes after.
- Automating a broken process. If the process today rests on one person's memory and every case is different, automation breaks it faster. Fix it first, then automate it.
- No human check. A model gets things wrong. A flow with no checkpoint finds its errors when the customer finds them.
- Personal data in tools with no contract. Pasting customer emails into a consumer tool breaks GDPR even if nobody notices. You need a provider with a data processing agreement and training on your data switched off.
- Never measuring. Without the time per case before and after, the automation stays an opinion. And you will not know where to put the second one.
Frequently asked questions
Which business processes can be automated with AI?
The repetitive ones with an unstructured input: incoming support requests, supplier documents in different formats, recurring quotes, periodic reports, content to adapt per channel. The model reads and structures, the rules do the rest, a person checks. Processes whose input is already structured are better automated without AI.
Which process should you start from?
A small, repetitive one, with a recoverable error and a volume that pays back: classifying incoming requests, for instance. Not the most important one. The first project is there to teach you how working with a model goes, so pick one where a mistake costs nothing. First you measure how it works today: time per case, cases per month, errors.
What does it cost to automate a process with AI?
The order of magnitude for one narrow process is from 1,500 to 6,000 euros and from one to three weeks, plus the model calls on a pay per use basis, which at SME volumes stay within a few euros or a few tens of euros a month. The variable that weighs most is the number of systems to connect. The return is counted in hours freed up, against the numbers measured beforehand.
Do we have to change the ERP?
No. The automation connects to the existing systems through APIs, webhooks or scheduled exports, and works alongside them. In most cases the ERP notices nothing: it receives structured data that a person used to type in.
How many Italian enterprises already use AI?
According to ISTAT, in 2025 16.4% of enterprises with 10 or more employees use at least one AI technology, against 8.2% in 2024. Large enterprises are at 53.1%, small ones at 14.2%. AI workflow automation, though, reaches only 2.9% of the total: the most widespread use is reading and writing text.
What do I have to do about GDPR and the AI Act?
Three things. Use a model provider with a data processing agreement and training on your data switched off, if personal data goes into the prompts. Document AI literacy among staff (Article 4 of the AI Act). Declare the use of AI if the flow interacts with people outside the company (Article 50, from 2 August 2026). The AI Act guide on our site lists the documents to produce.
Sources
The figures quoted in the text, each linked to the original document. Numbers are reported as published by the source, with the year they refer to.
- ISTAT, Enterprises and ICT, 2025: Press release of 15 December 2025 and the attached report: 16.4% of enterprises with 10 or more employees use AI (8.2% in 2024); large enterprises 53.1%; technologies and purposes as a share of the users; 33.4% of the users with no stated purpose.
- Eurostat, Use of artificial intelligence in enterprises: 2025 data: 20.0% of EU enterprises use AI (13.5% in 2024); Italy 16.4%, with 14.2% among enterprises with 10 to 49 employees and 27.6% among those with 50 to 249; workflow automation 5.4% in the EU and 2.9% in Italy across all enterprises.
- Osservatorio Artificial Intelligence, Politecnico di Milano: Press release of 5 February 2026: the AI market in Italy worth 1.8 billion euros in 2025, up 50% on 2024; 71% of large enterprises, 15% of medium ones and 7% of small ones running at least one pilot.
- European Commission, 2026 State of the Digital Decade report: Report of 17 June 2026: almost 20% of EU enterprises use AI in 2025, up 48% in one year. The 2030 target set by Decision (EU) 2022/2481 is 75% of enterprises using cloud, big data or AI.
- Asana, Anatomy of Work Global Index 2023: Press release of 8 March 2023: 58% of a knowledge worker's day goes on "work about work", coordination instead of skilled work. 9,615 respondents in six countries.
- McKinsey Global Institute, A future that works: Automation, employment, and productivity: January 2017: about 60% of occupations have at least 30% of their activities open to automation with technologies already demonstrated; fewer than 5% can be automated in full.
- Regulation (EU) 2024/1689 (AI Act), Articles 4 and 50: AI literacy among staff and transparency obligations. The application dates as updated by the Digital Omnibus are in the guide on our site.
Written by Maximilian Giurastante
Founder and software developer, MAXYMIZE. Designs and builds SaaS platforms, custom business software and automations with language models. Over twenty years as a project leader in technology, audio and video systems integration.
Do you have a repetitive process you want to measure?
Bring us the list of the tasks that keep coming back, with the times if you have them. In thirty minutes we pick out the first process to automate and tell you what it would cost.