Business process automation: the complete guide for SMBs (with a lesson from the Venetian Arsenal)

Every company has a process everybody hates: the order that arrives by email and has to be retyped into the management system, the invoice waiting for three signatures, the customer calling to ask where the goods are. Business process automation exists to remove exactly that kind of work, but it only works if you first take an honest look at how work actually gets done. This guide explains where to start, which method to follow, what it costs in effort and which mistakes to avoid.

First, though, a short history lesson. I promise we'll be back at the office within two paragraphs.

Dante at the Arsenal: the first process map in Italian literature

In Canto XXI of the Inferno, written in the early 1300s, Dante has to describe the pit where corrupt officials are plunged into boiling pitch. To do it, he picks an image his readers knew well: the Venetian Arsenal in winter, when ships can't sail and repairs get done.

Quale ne l'arzanà de' Viniziani / bolle l'inverno la tenace pece / a rimpalmare i legni lor non sani, / ché navicar non ponno – in quella vece / chi fa suo legno novo e chi ristoppa / le coste a quel che più vïaggi fece; / chi ribatte da proda e chi da poppa; / altri fa remi e altri volge sarte; / chi terzeruolo e artimon rintoppa – (Inferno XXI, 7-15)

Roughly: as in the Venetians' arsenal the sticky pitch boils in winter to recaulk their damaged ships, since they cannot sail, and meanwhile one builds a new hull, one plugs the sides of a ship that has made many voyages, one hammers at the bow, one at the stern, others make oars or twist ropes, one patches the jib and the mainsail.

Now read it the way an operations consultant would. There's a dead season (winter) used for maintenance. There's a shared raw material (pitch). Above all, there's a crystal-clear division of labor: building new hulls, caulking ships with many voyages behind them, bow work and stern work, oars, ropes, sails. Everyone has a task, the tasks run in parallel, and the result is a fleet ready for spring.

The Arsenal was the Republic's public shipyard, enlarged several times during the fourteenth century; it also turned out the merchant galleys that Venice leased to merchants through a public auction, the incanto (more on that in our piece on AI for logistics). No computers, obviously. But it had what most small businesses that want to "automate" today are missing: precise knowledge of who does what, in which order, with which materials.

Your company today is in that Arsenal. Before you buy any software, you need to be able to describe your own shipyard as precisely as Dante did in nine lines.

What business process automation is (in plain words)

A business process is a sequence of steps that turns an input (a request, a document, an order) into an outcome (a delivery, an invoice, an answer). Automating it means letting a system carry out some of those steps, so that nobody has to retype, forward or check everything by hand each time.

Very different tools live under the same label. It pays to tell them apart, because their costs and risks differ.

One rule of thumb is worth its weight in gold: start from the lowest level that solves the problem. Anthropic, one of the companies building these models, says the same in its technical guide on agents: aim for "finding the simplest solution possible, and only increasing complexity when needed". If a few lines of integration do the job, you don't need an agent. We gave agents their own article: AI agents: what they are and when a small business needs one.

Where Italian businesses stand

The numbers say the game has barely started. According to Istat (ICT usage in enterprises, 2025), the share of Italian companies with at least 10 employees using at least one AI technology rose from 8.2% in 2024 to 16.4% in 2025. The size gap is widening, though: 53.1% of large companies use AI, against 15.7% of SMBs.

Two details from the same report matter to anyone making decisions:

• among use cases, right after marketing and sales, comes the organization of administrative processes: that's where automation pays off first;
• among companies not using AI, 58.6% cite lack of skills as the obstacle. The bottleneck isn't the technology; it's the people who have to choose and use it.

Which processes to automate first

Not every process deserves automation. In We automate what makes sense you'll find the three questions to ask (is it repetitive? how costly is a mistake? who will use it?). Here we turn them into a concrete list, area by area.

Administration and accounting

• Recording supplier invoices: reading PDFs, proposing fields, human check before posting.
• Graduated payment reminders, with different wording by due date and customer.
• Reconciling bank statements against receivables and payables.
• Expense reports: photo of the receipt, data extraction, one-click approval.

Sales and customer service

• Quote requests arriving via email and web forms: routing, collecting missing data, drafting the offer.
• Answers to frequent questions (opening hours, order status, documents), with a hand-off to a person for everything else.
• Updating the CRM after every call or meeting.

Operations and logistics

• Entering incoming orders into the system without retyping.
• Automatic alerts on minimum stock and expiry dates.
• Telling the customer where the shipment is.

There's a dedicated deep dive: AI for logistics and warehouses. For a sector example, see also our piece on pharmacies, orders and stock.

Documents and archives

• Automatic classification of incoming documents (contracts, delivery notes, certificates).
• Natural-language search across the company archive.
• A calendar of contract, warranty and certification deadlines.

Details in AI document management.

People and organization

• Onboarding new hires: accounts, documents, training, all in a checklist that ticks itself.
• Leave requests with approval and calendar updates.

The six-step method

This is the heart of the guide. It's the same pattern we use in Penzai's process automation projects, and it works for a ten-person workshop as well as a professional firm.

1. Map the process as it is, not as it should be

Take a sheet of paper (that's perfectly fine) and write down every step of a real process, from the first email to the last click. For each step note who does it, with which program, how long it takes, and what happens when something goes wrong. Do it with the people who run it every day: they're the only ones who know the exceptions, which are exactly where automations break.

2. Measure before you touch anything

Three numbers are enough: how often it happens each month, how long it takes each time, how many errors or delays it produces. Without that baseline you'll never know whether automation worked. If a process happens twice a month, it's probably not worth it.

3. Pick the right level of technology

Go back to the four-level table. Ask yourself: does the data already arrive structured? Do the programs involved have APIs? Do you need to read free text? Is judgment required? The more "yes" answers to the last two, the more AI makes sense; otherwise rules and integrations will do.

4. Prototype on a small slice

You don't automate "administration". You automate "recording invoices from supplier X", for one month, alongside the manual work. That way you compare results with no risk and find the odd cases before they do damage.

5. Keep a person at the right checkpoint

Every automation needs a human checkpoint wherever mistakes are expensive: before paying, before sending an offer, before deleting data. "Human in the loop" isn't a slowdown; it's insurance. It's also what the rules ask for, as we'll see shortly.

6. Maintain it, like winter at the Arsenal

The Venetians used winter to put their ships back in shape. Automations need the same care: a supplier changes its invoice format, a portal redesigns its pages, an AI model gets updated. Put a quarterly review on the calendar and name the person responsible.

How much it costs to automate a process

Giving one figure for everyone would be dishonest: it depends on your software, the quality of your data and the number of exceptions. What we can tell you is which cost items to consider, so no quote catches you off guard:

• analysis and mapping: your time and the consultant's for steps 1 and 2;
• licenses: integration platforms, RPA tools, pay-per-use AI services (often billed per operation or document);
• development and integration: connecting existing programs, often the most uncertain item;
• training: whoever uses the new process has to understand it, or they'll work around it;
• maintenance: updates, checks, small changes over time.

Return is calculated with the step-2 numbers: hours saved per month, errors avoided, shorter response times. If those numbers don't move after the prototype, stopping is a good decision, not a failure.

A sample calculation, to redo with your own numbers

A typical scenario: a process that happens 400 times a month and takes 6 minutes each time eats up 40 hours a month, roughly one person's working week. If automation handles three cases out of four well and the person only needs to check them (say one minute each), hours drop to about 15. The 25 hours freed up are your margin: subtract licenses and maintenance and see whether the math works. Note that the calculation needs no vendor promises, only the numbers you measured yourself.

How to choose tools

The market is crowded and names change every six months. Brands matter less than a few questions to ask anyone pitching you a solution:

• Where does the data end up? EU-based servers, a data processing agreement (DPA), the option to exclude your data from model training.
• Does it connect to what you already use? ERP, email, CRM, archive. Automation that forces you to change everything else is rarely a bargain.
• Who changes it when something changes? You, through an understandable interface, or only the vendor, for a fee?
• What happens if the tool stops working? There must always be a written, well-known manual fallback procedure.
• How can you see what it did? Every automation must leave a readable trail: who, what, when. No log, no control.

Be wary of anyone promising to automate "the whole company" in one go. Gartner, the technology research firm, predicts that over 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value or inadequate risk controls. The companies that make it to the end almost always started small.

The rules you need to know

Automating with AI in Italy and the EU means working inside three frameworks.

• GDPR: if the process handles personal data (customers, employees, sole-trader suppliers) you need a legal basis, a privacy notice and adequate security measures. Watch out for tools that send data outside the European Union.
• The EU AI Act (Regulation (EU) 2024/1689): since February 2, 2025, Article 4 on AI literacy applies, requiring organizations that use AI systems to ensure a sufficient level of AI skills among their staff.
• Italy's AI law (Law no. 132 of September 23, 2025, in force since October 10, 2025): among other things, Article 11 requires employers to inform workers when they use AI systems, in the cases set out by the rules on transparent working conditions.

In practice: document which tools you use, for what, with which data, and train the people using them. It's good organization before it's an obligation.

The mistakes we see most often

1. Automating chaos. If the manual process is confusing, automation makes it confusing and faster. Simplify first, then automate.
2. Starting from the tool. "We bought the platform, now what do we do with it?" is the wrong question. Start from the problem.
3. Forgetting exceptions. 90% of cases are easy; the remaining 10% generate 90% of the phone calls. Decide upfront who handles exceptions and how they're flagged.
4. No owner. An automation without a name next to it breaks silently, and you find out from an angry customer.
5. Neglecting security. Every connection between programs is also a door. Shared credentials, overly broad permissions and no two-factor authentication are an open invitation. See our deep dive on phishing, AI and cybersecurity for SMBs.
6. Not talking to people. If staff fear automation is there to replace them, they'll find a thousand reasons not to use it. Explain what changes in their jobs (see AI and work: which tasks change) and involve them from day one.

People before software

There's a reason Dante, describing the Arsenal, talks about people rather than tools: "one builds", "one plugs", "one hammers". A process is made of human gestures. Good automation removes the repetitive ones and leaves people the ones that need judgment, relationships and experience. That's the idea behind our Human Tech approach, also explained in why we start from people.

A typical scenario: a distribution company with fifteen employees receives dozens of orders by email every day, in different formats. Two people spend their mornings retyping them. Mapping reveals that half the customers use the same form: for them a simple integration is enough. For the rest, an AI system proposes the order already filled in and a person confirms it. The two people don't disappear: they move from retyping to checking exceptions and calling back customers with unusual orders. Which is exactly the work an algorithm can't do.

Frequently asked questions about business process automation

Which process should we start with?

A frequent one, based on clear rules, with data arriving in digital form and where an error is easy to catch. Recording supplier invoices, entering orders and answering frequent questions are the most common candidates in small businesses.

Do we really need artificial intelligence?

No. Many processes can be automated with rules and integrations, no AI involved. AI is useful when you need to read unstructured text or documents, or when cases vary too much to write fixed rules.

Will automation replace employees?

In most small businesses the effect is a change in tasks more than a cut in jobs: less retyping, more checking, customer relationships and exception handling. The ILO's 2025 study on generative AI indeed points to job transformation rather than replacement. We dig deeper in AI and work: which tasks change.

How long before we see results?

A prototype on a single process can be up and running in a few weeks, if the data is accessible. Results must be measured against the numbers collected before you started, which is why "measure before you touch anything" can't be skipped.

Want to automate a process without losing control?

At Penzai we always start from the map of your "Arsenal": how people work today, where time gets lost, what's worth automating and what isn't. If you run an SMB or you're a business owner wondering where to begin, book a free intro call: in half an hour we'll pick the first process to tackle together.