How much does a website cost in Italy in 2026
Why the price depends on scope and how to read a quote without comparing only the total.
Blog · Automation
AI automation for SMEs: 5 processes to start with
AI is not a strategy: it is a tool. SMEs that get results do not start from the model. They start from a repetitive process, measure it, and then decide whether to automate it.
AI is not the starting point
Generative AI made it easy to produce text and images. It did not make it easy to decide which process is worth automating.
The right question is not “which model do we use”, but “which repetitive work can we remove without increasing risk”. Automation that gets a quote or an invoice wrong costs more than the work it replaces. That is why you start from a process, not a tool.
SMEs have an advantage: processes often live in the hands of a few people, so they can be observed and measured. They also have a constraint: there is no IT department to absorb the project. The pilot has to be small, understandable and manageable by the people who do that work every day.
This article does not list “cool things AI can do”. It lists five processes with three characteristics: they repeat, they have a checkable output, and someone in the company can say whether the result is right.
The five processes to start with
For each one you will find what the automation does, where it breaks and how to measure it. The order is not a ranking: it depends on your volume and your data.
Quotes and proposals
Trigger: a request arrives by email or form. Input: description, price list, terms. Output: a draft quote with lines, quantities and notes. A person approves prices and discounts before it is sent.
Where it breaks: unstructured price lists and ambiguous requests. How to measure it: time from request to quote and the share of lines correct on the first pass.
Supplier invoices and documents
Trigger: an invoice or delivery note arrives. Input: PDF, email, scan. Output: extracted data and a proposed entry in the ERP. Confirmation stays with a person.
Where it breaks: different formats, skewed scans, suppliers with incomplete data. How to measure it: documents processed without correction and average entry time.
Inbound email and requests
Trigger: a message arrives. Output: classification, owner, extracted data and a draft reply. Sending can stay manual, at least in the first phase.
Where it breaks: requests that mix several topics and customers writing from different addresses. How to measure it: time to first reply and the share of messages routed correctly.
Customer support
Trigger: a question about a product, order or service. Input: approved documentation, FAQs, order status. Output: an answer with its source, and a handover to a person when confidence is low.
Where it breaks: outdated documentation and cases outside the scope. How to measure it: requests resolved without escalation, escalation quality and customer satisfaction.
Reporting and control
Trigger: periodic deadlines. Input: data from the ERP, CRM and spreadsheets. Output: a summary, anomalies and questions to check. The decision stays human.
Where it breaks: scattered data and different definitions of the same metric. How to measure it: preparation time and the number of anomalies found before they become a problem.
How to choose the first process
Do not choose the one that sounds most innovative. Choose the one you can observe, measure and correct.
- FrequencyDoes it happen often enough to justify the project? A weekly, complex process can be worth more than a daily, trivial one, but volume has to be measured.
- StandardisationAre the steps the same every time? If every case is different, automation produces exceptions instead of reducing them.
- Cost of errorIf it gets something wrong, can you correct it before the damage is irreversible? Start with reversible, checkable errors.
- Available dataIs the information digital, current and accessible? A process with dirty data needs data work first.
- A process ownerIs there a person accountable for the result who can stop the pilot? Without an owner, nobody corrects the errors.
Data, privacy and responsibility
Automation moves data and decisions. Before writing code, the rules have to be clear.
- Legal basis and minimisation. Process only the data the workflow needs and keep it for the stated period. Do not add fields “because they might be useful”.
- Suppliers and contracts. If you use an external service, ask where the data runs, whether it is used to train models, what security measures exist and who answers if there is an incident.
- Impact assessment. Some processing requires an assessment before it starts. It is not paperwork: it is where you decide what not to automate.
- Human oversight. Define who checks, how often and what happens when the system is not confident. “The AI decided” is not an acceptable answer.
- Logs and security. Operation logs, limited access, managed secrets and instructions against prompt injection. Security is not added afterwards.
This is not legal advice. Before processing personal or sensitive data, involve a privacy adviser and check the obligations that apply to your business.
A four-week pilot
You do not need a year-long project to find out whether a process can be automated. You need a narrow scope and a stop criterion.
Week 1: map and baseline
Observe the real process, not the written procedure. Measure volume, time per case, errors and people involved. Pick one output to improve.
Week 2: narrow prototype
Build the minimum path with real, authorised data. No complete integrations yet: first prove that the output is useful and checkable.
Week 3: evaluation
Compare the prototype with the baseline on a fixed set of cases. Look at the errors, not just the average. Ask the people who did the work whether they would use it.
Week 4: decision
Scale, change or stop. If the pilot does not beat the baseline, do not add features: close it and move to another process. The cost of a stopped pilot is much lower than a failed project.
What not to automate
There are processes where AI can assist but must not decide. And others where it is better not to start at all.
- Decisions about people: hiring, evaluation, dismissal. The cost of an error is not only financial.
- Credit, insurance and access to services: high-impact decisions that require transparency, fairness and oversight.
- Legal, medical or tax advice: AI can prepare material, not take responsibility for the answer.
- Processes without an owner: if nobody can say what a correct result looks like, there is no way to evaluate the automation.
- Dirty data with no source of truth: automating a mess produces faster errors. Fix the data first.
The practical rule: if you cannot explain to a person how to recognise a good result, you cannot automate it reliably.
A simple time return example
A simple sum on a repetitive process.
| Item | Value |
|---|---|
| Emails handled per month | 400 |
| Minutes per email | 3 |
| Hours saved per month | 20 |
Formula. Emails × minutes / 60 = hours saved.
Frequently asked questions
Do I need ChatGPT to automate a process?
No. Models are interchangeable; data, integration and control matter. A more powerful model on an undefined process only produces faster errors.
How much does an automation project cost?
It depends on the process, the integrations, the volume and how much human control is needed. A quote is built after measuring the process and the available data. A narrow pilot is the cheapest way to find out whether it is worth it.
Does my data end up in the models?
It depends on the supplier and the contract. Ask whether data is used to train models, where it is stored, for how long and who can access it. Prefer enterprise terms, European hosting and data processing agreements.
Do I have to replace my ERP?
In most cases, no. You integrate what already exists if it exposes APIs or import/export. Replacing the ERP is a much larger project and should be decided separately.
How long does a pilot take?
Four weeks is a good starting point for a narrow process. Processes with many exceptions or complex integrations take longer. A small, finished pilot beats a broad, inconclusive project.
How do I measure the return?
Start with a baseline: hours per case, errors, cycle time, number of steps. Compare the after with that baseline and add the total cost: tools, integration, maintenance and review time. If the saving does not cover the cost, the project is not working.
Pick a concrete process
Tell us where time is lost today: email, quotes, invoices, reporting. We can measure the process and work out whether a pilot makes sense.