The words, without the jargon.

A

Churn

abbandono

The share of customers who stop buying over a given period. In businesses built on repeat purchases, it decides your fate more than any acquisition campaign.

In practiceCustomers who stop buying rarely tell you. They slip away quietly, and the only people who notice are the ones watching buying patterns rather than cancellations.

Training

addestramento

The process by which a model learns its behaviour from a set of examples, adjusting its parameters until the error comes down. It's the expensive part, and it happens once: after that, the model is only queried.

In practiceNo SME trains a model from scratch. The realistic options are two others: give it the right context at the moment you ask, or use fine-tuning for repetitive cases.

AI agent

agente AI

A system that does more than answer. It's given a goal, decides for itself which steps to take and uses outside tools (reading an inbox, writing to a spreadsheet, querying the business software) until the job is done.

What sets it apart from a chatbot isn't cleverness, it's permission to act. That's why an agent needs clear boundaries before it's switched on.

In practiceA chatbot tells you the status of an order. An agent looks it up in the business software, finds it's held at customs, drafts the email to the customer and leaves it waiting until someone reads it.

AI Act

Regulation (EU) 2024/1689

The world's first comprehensive law on artificial intelligence, in force since 1 August 2024 and applying in stages. It doesn't regulate the technology itself: it sorts uses by level of risk, and the obligations follow from that.

The bans on prohibited practices and the staff AI literacy requirement have applied since February 2025; the obligations for high-risk systems come later. For the vast majority of SMEs the real impact is small: replying to a customer or filling in a document isn't a high-risk use. It becomes one when a system selects people, decides access to an essential service or affects someone's rights.

In practiceThe question isn't “do we use AI?” but “where does an automated decision affect a person?”. If the answer is nowhere, the heavy obligations don't apply to you. If there's one such point, they apply to that point.

Algorithm

A defined sequence of steps for solving a problem. In public debate the word has taken on a more sinister meaning than it really has: most of the algorithms that run a business are rules someone wrote down and anyone can read.

In practice“The algorithm decides” isn't an answer. If nobody in the business can explain the rule, it was adopted without anyone reading it.

Hallucination

allucinazione

When a model produces a statement that is fluent, plausible and false, with exactly the same confidence as a true one. An LLM predicts likely text, not correct text: hallucination is the normal result of how it works, not a malfunction.

You can't eliminate it. You contain it, by giving the system a knowledge base to draw on (RAG) and putting a person at the point where a mistake would be expensive.

In practiceThe risk isn't that the system gets things wrong. It's that it gets them wrong confidently, in a reply to a customer that nobody checked.

AI hallucinations: what they are, examples and how to avoid them →

Test environment

staging · sandbox

A copy of the system where you try out a change before applying it to the real one. It costs little and prevents the most expensive kind of mistake: the one you discover in production.

In practiceFor an automation that messages real customers, testing in production means sending the wrong message to a real person.

Master data

anagrafica

The core records that identify customers, suppliers and products, which every other system refers back to. When the master data is messy, every analysis built on it inherits the mess.

In practiceThe same customer spelled four different ways leaves you with their revenue split four ways, and no report will put it back together.

API

application programming interface

A program's back door: the point where another program can ask it for data or actions without going through the interface built for people. It's what makes integration possible.

In practice“Does the business software have an API?” is the first technical question to ask a supplier, and the answer decides whether an integration takes days or months.

Unsupervised learning

apprendimento non supervisionato

Training on unlabelled data, where the system finds recurring patterns by itself: groups of similar customers, unusual behaviour, hidden structure.

In practiceGood for discovering, not for deciding. Any grouping needs checking by someone who knows the business, because the machine finds patterns even where they mean nothing.

Reinforcement learning

apprendimento per rinforzo · RL

Training by trial and reward: the system tries something, gets a score and adjusts its strategy. It's the method behind game-playing systems and, in modified form, behind aligning language models with the behaviour we want.

Supervised learning

apprendimento supervisionato

Training on examples that are already labelled: every input comes with the right answer. It's the most common approach in business applications and the most reliable, because the goal is unambiguous.

In practiceSomeone has to do the labelling. It's the human work that quotes tend to forget to price in.

Attention

attenzione

The mechanism a model uses to decide, word by word, which other parts of the text to look at. It's what lets it connect a reference at the start of a document with the sentence that refers back to it at the end.

Two-factor authentication

2FA · MFA

Proving your identity with two independent things: a password and a one-time code. No other single measure gives you as much protection for as little hassle.

In practiceTurn it on for email accounts first, before the business software. Email is where the passwords for everything else get reset.

Automation

Getting a machine to do a step that someone currently does by hand, with the same result and without using up their time. It doesn't need AI: most of the automations that give an SME its hours back are fixed rules, not models.

In practiceBefore asking which AI you need, ask whether the step needs doing at all. Automating a pointless step just gives you a faster pointless step.

Workflow automation

automazione dei flussi

Automating a sequence of steps that runs across different people and systems, rather than a single isolated action. This is the level where real hours get freed up.

In practiceIn an SME, time isn't lost in individual tasks. It's lost in the handovers between them. That's where to measure it: in the wait between one task and the next.

Autonomy

autonomia · livello di autonomia

How much a system can do without asking. Think of it as a ladder, not a switch: suggest, prepare, act and notify, act silently.

It's the most important design decision, and almost always the only one made out of habit rather than by choice.

In practiceYou climb one rung at a time, once the rung below has worked for weeks. Starting at the top rung doesn't mean you've built a more advanced system. It means you skipped the testing.

Self-hosted

autonomo dal fornitore

Software installed and run on your own infrastructure, often built on open-source projects. You lose the subscription and gain the maintenance: it isn't free, you pay for it in hours.

B

Backup

copia di sicurezza

A copy of the data kept somewhere separate. The number that matters isn't how often it's made, but when someone last checked it by actually restoring it.

In practiceA backup that's never been tested is just a hope with an automatic schedule attached.

Knowledge base

base di conoscenza

Everything a business already knows (product sheets, price lists, terms of sale, internal rules, the answers given a thousand times), organised so that a system can search it.

It's the piece almost nobody wants to build, and without it every automated assistant stays generic.

In practiceIt already exists, just scattered: in two people's heads, in a shared folder, in ten years of sent emails. The job isn't creating it, it's gathering it.

Algorithmic bias

distorsione

A model's systematic tendency to favour or penalise certain cases, inherited from the data it was trained on. The system has no opinions. It reproduces, in statistical form, the history it was given.

It becomes a legal problem, not just an ethical one, when the system affects people: recruitment, access to credit, assessing an application.

In practiceA system trained on how you've made decisions so far will repeat the mistakes you've made so far too, only faster and on more cases.

C

Chain of thought

catena di pensiero · CoT

The technique of asking a model to show its intermediate steps before giving its conclusion. It makes multi-step tasks more reliable and makes mistakes easier to spot, because you can see where the reasoning went wrong.

Chatbot

An interface that answers in the form of a conversation. The term describes how a system looks, not how capable it is: behind the same little chat window there could be a tree of canned answers written ten years ago, or a model connected to the business's knowledge base.

In practiceThe useful question isn't “do I need a chatbot?” but “which questions do we answer ten times a day?”. If there aren't any, the chatbot has nothing to do.

Cloud

nuvola

Running software on someone else's servers, paying for what you use. You no longer buy machines, but you take on a dependency: whether you can work depends on a contract and a connection.

Queue

coda

A waiting line between two systems that work at different speeds. It absorbs peaks and makes sure nothing gets lost when one side produces faster than the other can handle.

Long tail

coda lunga

All the rare, highly specific searches which, added together, outnumber the generic ones. There's less competition for them, and they bring visitors who already know what they want.

In practiceEveryone is competing for “business automation”. “How to reduce no-shows in a restaurant” is searched only by people who have that exact problem.

Lead nurturing

coltivazione dei contatti

A sequence of useful contacts with someone who isn't ready to buy yet, so you stay in the picture until they are.

In practiceThe difference between nurturing and nagging isn't how often you get in touch. It's whether every message offers something useful even to someone who will never buy.

Business continuity

continuità operativa · disaster recovery

The plan that sets out how work gets going again after a serious failure, and how quickly. The two questions: how long can we afford to be down, and how much data can we afford to lose?

Copilot

copilota · assistente

A system that works alongside a person rather than replacing them: it suggests, completes and summarises, and the operator has the final say.

In practiceIt's the way AI gets into a business with the least resistance, because it takes nothing away from anyone. The gain is smaller too, and harder to measure.

Customer acquisition cost

costo di acquisizione · CAC

What it costs on average to win a new customer, counting everything: advertising, sales time, tools. Always read it alongside lifetime value: on its own it tells you nothing.

In practiceAn expensive customer isn't a problem if they stay for years. A cheap one who leaves after two months is a loss.

Encryption

Turning data into a form nobody can read without the key. There are two situations to keep apart: in transit, while the data is moving, and at rest, while it's stored. Many providers mention the first and stay quiet about the second.

CRM

customer relationship management

The single record of every business relationship: who each contact is, what they've been told, what needs to happen next and when. A CRM doesn't make sales. It makes visible what currently lives in different people's heads and separate inboxes.

In practiceThe usual failure isn't technical. A CRM nobody updates becomes a second job and gets dropped within three months: if it doesn't fill itself from what's already happening, it won't last.

CRM: what it is, what it does and when an SME needs one →

Dashboard

cruscotto

The screen that brings all your indicators together in one place. Its value isn't in the charts, it's in someone looking at it and deciding something.

In practiceA dashboard nobody has opened for three weeks has already told you what you needed to know: those numbers weren't feeding any decision.

D

Vector database

database vettoriale

A store that indexes documents by meaning rather than by exact words, using embeddings. It's the infrastructure behind RAG once the number of documents gets large.

In practiceIt solves a volume problem. Bringing it in before you have that volume just adds one more thing to maintain and solves nothing.

Dataset

insieme di dati

The collection of examples a model is trained or tested on. The quality of the dataset shapes the model's behaviour more than any technical choice made later.

In practiceIt holds on a small scale too: if you build a system that sorts customer enquiries, how good it is will depend on the examples you gave it, not on the skill of the person who set it up.

Structured data for the web

schema markup · JSON-LD

Hidden labels in a page that tell search engines what its content is: a business, a frequently asked question, an article, a definition. They don't change what visitors see; they change what machines understand.

Unstructured data

Everything else: emails, PDFs, contracts, voice notes, a crooked photo of a delivery note taken in the warehouse. It's most of the information that moves around a business.

This is exactly why AI has become relevant to businesses: unstructured data, which used to need a person, can now partly be handled by a machine.

In practiceAlmost all admin work is turning unstructured data into structured data by hand: reading a PDF and retyping it into the business software.

Structured data

Anything that fits in rows and columns: an orders table, a customer list, a bank statement. It's the form in which data can be queried, added up and compared directly.

Deduplication

Finding and merging records that describe the same thing written in different ways. It's tedious and unglamorous, and in practice it unlocks more value than many AI projects.

Deep learning

apprendimento profondo

The branch of machine learning built on neural networks with many layers. It's the technique behind image recognition, good machine translation and modern language models.

In return for that power it needs two things: masses of data and masses of computing power. An SME has neither, which is why these models are rented rather than built.

Distillation

distillazione

Using a large model to train a smaller one that copies its behaviour on a narrow task. The result is cheaper and faster, within that task.

E

Batch processing

elaborazione a lotti

Handling many items together at fixed intervals: every night, every hour. It's cheaper and makes the design simpler, at the cost of waiting.

In practiceFine for the accounts, terrible for a customer enquiry. The question is always: who's waiting for this result, and how long can they wait?

Embedding

rappresentazione vettoriale

Turning a piece of text into a string of numbers that represents its meaning. The point is that two sentences saying the same thing in different words end up close together, so you can find one starting from the other.

In practiceIt's why a well-built search finds “the document for shipping outside Europe” even when the document itself says “export declaration”.

ERP and business software

enterprise resource planning · gestionale

The system where a business records what actually happens: orders, stock, production, invoices. In many Italian SMEs it's the oldest piece of the infrastructure and also the most important, because it holds the accounting truth.

In practiceConnect it, don't replace it. Replacing business software that works is a months-long project that creates no new value: you pay to end up back where you started.

ERP: what an ERP system is and how it differs from a CRM →

ETL

extract, transform, load

The process of extracting data from one system, converting it to a common format and loading it into another. It's the plumbing of any data project.

F

Few-shot

apprendimento per esempi

Including a few worked examples in the instructions, so the model picks up the expected format instead of having to guess it. It's the cheapest way to get consistent output.

In practiceThree real examples taken from your own documents are worth more than a page of instructions describing what you'd like them to look like.

Fine-tuning

riaddestramento mirato

Retraining an existing model on your own data so that it reliably sticks to a format, a tone or a repetitive task. It costs money, needs plenty of examples, and has to be redone whenever the underlying model changes.

In practiceIt's needed far less often than people think. Most of the time what's missing isn't a trained model, it's the right information put in front of an ordinary one.

Context window

finestra di contesto

The most text a model can hold in view at once: the question, any attached documents, the conversation so far. Go past that limit and you don't get an error, you get silent forgetting.

In practiceIt's why an assistant that worked perfectly for twenty exchanges starts contradicting itself by the fortieth.

Workflow

flusso di lavoro

The actual sequence of steps by which something comes into a business, gets worked on and goes out again: who receives it, who decides, who does it, who checks it.

In practiceAlmost nobody has written it down. Automating before you've written it down means speeding up the mess, and a fast mess is worse than a slow one: it does damage before anyone notices.

Workflow: meaning, examples and automation in a business →

Single source of truth

fonte unica di verità

The principle that every piece of information has one official home, and everyone else reads it from there rather than keeping their own copy. It's the only structural defence against things drifting out of sync.

In practiceThe sign it's missing: two people in a meeting with two different numbers for the same thing, both taken from a company system.

G

GDPR

Regulation (EU) 2016/679

The EU rules on the protection of personal data. They apply to any system that processes data relating to an identifiable person, and that includes AI systems: automated processing doesn't reduce the obligations, if anything it adds some on transparency.

In practiceThe two points that matter most in practice: where the data you send to an AI provider physically ends up, and whether that provider can use it to train its own models. Both are questions about the contract, not matters of opinion.

Guardrail

vincolo di sicurezza

The limits built into an automated system: what it mustn't say, what it mustn't do, and the point at which it has to stop and hand over to a person. They're not a later add-on: they're as much part of the design as the main function.

In practiceThe question that tests them: what happens if the system gets an absurd request? If the answer is “we hadn't thought about that”, there's no guardrail.

I

Idempotence

The property whereby running the same operation twice gives the same result as running it once. It sounds like a textbook detail, but it's what stops the same invoice going out twice when a system retries.

In practiceIt's the difference between an automation that restarts by itself after a failure, and one that does the damage twice when it restarts.

Inference

inferenza

Using a model that has already been trained to get an answer. It happens every time the system does any work, and it's the recurring cost: training is paid for once, inference every day.

Data entry

inserimento dati

Copying information by hand from one place to another. It's the most repetitive, least visible and most easily eliminated task in any business.

In practiceIt never shows up as a cost line in the accounts, which is why it survives for years: nobody has ever added it up.

Integration

Connecting two systems so that data entered once shows up wherever it's needed, without being retyped. It's the least glamorous work, and the kind that gives back the most hours.

In practiceEvery bit of retyping is two costs in one: the time of the person doing it, and the mistake that someone makes sooner or later. You always find out about the second one after the first.

Artificial intelligence

intelligenza artificiale · AI

The field that builds systems able to do things we associate with human intelligence: recognise, classify, predict, generate, decide. It's a broad label, and that's the first reason sales conversations about “AI” go round in circles.

The same word covers things that differ hugely in cost, risk and reliability: a classifier that sorts emails and a model that writes text have almost nothing in common.

In practiceWhen a supplier says “we use AI”, they haven't said anything you can check. The useful question is: which task, on which data, with what acceptable error rate?

Generative AI

AI generativa · GenAI

The family of systems that create new content (text, images, code, translations) rather than just classifying or predicting. It's the part of AI that, since 2023, has made the subject visible outside IT departments, and it's only one part.

In practicePlenty of operational problems don't need anything generated. They need things read, sorted and routed: duller jobs, more reliable, and often more profitable.

J

Jailbreak

aggiramento dei limiti

A prompt crafted specifically to get a model to break the rules it has been given. The business risk is less about embarrassing output and much more about an assistant that, pushed the right way, reveals how your processes work or what's in its knowledge base.

In practiceA public-facing assistant shouldn't have access to anything you wouldn't be happy to publish.

JSON

The format two systems use to swap structured data: name-value pairs that a person can read too. It's the common language of modern integrations.

K

KPI

key performance indicator · indicatore chiave

The measure you pick to tell whether something is going the way it should. A KPI is only useful if someone can act on it: if nobody can change it, it's not an indicator, it's just news.

In practiceThree indicators checked every week beat twenty on a dashboard nobody opens.

KPIs: what they are, examples and how to choose them →

L

Latency

The time between asking a system for something and getting its response. For an overnight job it doesn't matter at all; for a customer who's waiting, it matters more than how good the answer is.

In practiceA perfect answer that takes thirty seconds is worth less than a good one that takes three.

Lead

contatto commerciale

A contact who has shown interest but hasn't bought yet: an enquiry through the website, a WhatsApp message, a business card picked up at a trade fair. It's the most fragile moment in the whole sales cycle, and almost always the one nobody's watching.

In practiceLeads aren't lost to a no. They're lost to silence: nobody replies fast enough, and by the time someone does, they've already bought elsewhere.

Leads: what the word means, and cold, warm and qualified leads →

Rate limit

limite di frequenza

The maximum number of requests a service will accept in a given period. Going over it doesn't throw an obvious error: it produces rejected requests which, if nobody logs them, turn into work that quietly disappears.

LLM

large language model · modello linguistico di grandi dimensioni

The engine underneath conversational AI tools. Trained on vast amounts of text, it predicts the most likely continuation of whatever is in front of it.

That mechanism gives it both its usefulness and its built-in flaw: it predicts what's likely, it doesn't check what's true. See hallucination.

In practiceTreat it like a very fast, very well-read colleague with no access whatsoever to your data, until you give it access in a controlled way.

LLMs: what a large language model is and how it works →

llms.txt

A text file at the root of a website that describes what the site contains in a tidy form, designed to be read by language models rather than traditional search engines.

In practiceIt's the robots.txt of an era in which more and more people ask an assistant instead of searching.

M

Machine learning

apprendimento automatico · ML

The approach where a system's behaviour isn't written by hand, rule by rule, but learned from past examples. It's the broadest family, and generative AI sits inside it.

In practiceIt needs history. If the business hasn't recorded past outcomes, there's nothing to learn from, and that, not cost, is the real reason some projects never get off the ground.

No-show

mancata presentazione

An appointment or booking where nobody turns up and nobody cancels. It costs you twice: the slot stays empty, and nobody else could book it.

In practiceThe damage isn't the absence itself, it's finding out too late to fill the slot again.

No-shows: what they are, what they cost and how to reduce them →

Metadata

Data about other data: when a document was created, by whom, which case it belongs to. It's what makes an archive searchable rather than just full.

Vanity metric

metrica di vanità

A number that goes up, feels good and doesn't inform any decision: visits, impressions, total contacts. One question gives it away: if this number doubled tomorrow, what would we do differently?

In practiceIf the answer is “nothing”, it's not worth reporting.

Middleware

strato intermedio

Software that sits between two systems and lets them talk, translating formats and rules. It exists because neither system was designed with the other in mind.

Migration

Moving data and processes from one system to another. It's the kind of project that overruns most often, and for the same reason every time: the real work isn't moving the data, it's discovering how many different ways it was entered.

Data minimisation

The principle that you only process the data needed for a given purpose, and not a single item more. It's the most effective safeguard there is, because data you don't hold can't be stolen from you.

In practiceIt applies fully to AI: sending a whole document to an outside model when three fields would have done is the most common way of creating a problem that wasn't there before.

Baseline

misura di partenza

The value of an indicator before you change anything. Without it, there's no honest way to say whether something worked.

In practiceIt's the one step you can't go back and do later. If you don't measure before, you can't prove anything afterwards, not even to yourself.

Model

modello

The output of training: a mathematical object that takes an input and produces an output. “Model” means the specific thing you're using, with its own version, cost and behaviour.

Different models from the same provider behave differently on the same task, and an update can change what comes out of a flow that used to work.

In practiceA system that goes live should state which model it uses, and which version. If it doesn't, the day its behaviour changes, nobody will know why.

Monitoring

monitoraggio

Keeping a constant watch on a live system, with an alert when something strays outside the expected range. The question that tells you how good it is: who gets the alert, and how long before they notice?

In practiceThe usual failure isn't a system that breaks. It's one that breaks and still looks like it's running.

Multimodal

multimodale

A model that handles more than one kind of input: text, images, audio, documents. It's what makes a crooked photo of a delivery note usable, without a separate recognition step.

In practiceIt's the leap that makes document automation realistic for businesses that receive paper, photos and scanned PDFs rather than tidy files.

N

No-code and low-code

Tools for building automations by connecting blocks on screen instead of writing code. They lower the barrier to entry; they don't lower the need to understand what you're building.

In practiceThe classic risk is a flow built by one person, never documented, that stops working when that person moves to another role. The weak point isn't the tool, it's that nobody knows what the flow does any more.

Normalisation

Bringing mixed data into a common format: dates, units of measure, company names, currencies. You can't compare anything without it.

O

OCR

optical character recognition · riconoscimento ottico

Automatically reading the text in an image or a scanned PDF. The technology is decades old, but recent models have made it far more reliable, especially on crooked, faded or handwritten documents.

In practiceIt's the first rung of document processing: without it, every photographed delivery note stays an image that only a human eye can read.

OCR: what it is, how it reads a PDF and when you actually need it →

Omnichannel

omnicanale

Handling every channel (phone, email, WhatsApp, social media, website) as one conversation, with a single history for each customer.

In practiceThe sign it's missing: customers have to explain their situation from scratch depending on which channel they use.

On-premise

in sede

Running on machines inside the business. It costs more to manage, but it's still the right answer when data can't leave the company network.

Orchestration

The layer that decides which system acts, in what order, with what data and, above all, what happens when a step fails. It's the difference between a handful of automations and actual infrastructure.

In practiceThe question that tells them apart: when something breaks at two in the morning, does the system notice and tell someone, or do you find out from an angry customer?

Optimisation for AI assistants

AEO · answer engine optimization

Shaping content so that a conversational assistant can quote it, not just so it can rank in a list of results. It rewards clear structure, verifiable claims and named sources.

In practiceIt's why a source quoted in full is worth more than a catchy line: a machine can repeat the first, but not the second.

Overfitting

sovradattamento

The flaw in a model that has memorised its training examples instead of learning the general rule: excellent on cases it has seen, unreliable on new ones.

In practiceIt's the statistical equivalent of a colleague who handles every case they've seen before perfectly, and freezes at the first one that's different.

P

Parameters

parametri · pesi

The internal values a model adjusts during training, which encode how it behaves. Their number, which runs into billions in large models, is used as a measure of power.

It's a rough measure: more parameters means more potential capability, not necessarily better results on your task, and always more cost.

Sales pipeline

imbuto di vendita

All the open deals, sorted by stage. It answers a question almost no SME can answer with a number: how much work is coming, and how much of it will fall through.

Least privilege

privilegio minimo

The principle that every user and every system gets only the permissions its job requires. It applies to people and, increasingly, to automated agents.

In practiceAn automation that reads email shouldn't be able to delete it. An automation that prepares a payment shouldn't be able to make it.

Prompt

istruzione

The instruction given to a model. In a properly built system the user doesn't rewrite it each time: it's part of the system, written once, tested, version-controlled and changeable without touching anything else.

In practiceIf the quality of the answer depends on how the person on shift happened to word the request, you don't have a system. You have a tool, and a different result for every person.

System prompt

prompt di sistema

The standing instruction that sets an assistant's role, limits and tone, applied to every conversation before the user asks anything. It's where the business's rules get written down: what it may say, what it must refuse, when it has to hand over to a person.

In practiceIt's the most important document behind any automated assistant, and almost always the only one nobody has ever read from start to finish.

Prompt engineering

The craft of wording instructions so they reliably produce the result you want. The term sounds grander than it is: mostly it's about spelling out what you'd been taking for granted.

Prompt injection

iniezione di istruzioni

An attack where a hostile instruction is hidden in content the system will read (the body of an email, a PDF, a web page) to make it do things nobody asked it to. It's the signature weakness of any system that reads material from outside.

In practiceIt's the technical reason why a system that reads incoming email mustn't also be able to send, delete or pay without approval.

Lead scoring

punteggio dei contatti

Giving each contact a priority based on signals you can observe: sector, size, source, behaviour. The point is to send sales time, your scarcest resource, where it's most likely to pay off.

In practiceIt's not for throwing away low-scoring contacts. It's for deciding the order you call them in, a decision you're making anyway, currently by gut feeling.

Touchpoint

punto di contatto

Every moment a customer comes into contact with the business: the website, a phone call, a quote, an invoice, a message after the sale.

In practiceCustomers judge the whole experience by the worst touchpoint, not the average. It pays to find that one, rather than polishing the ones that already work.

Q

Data quality

qualità del dato

How complete, accurate, up to date and consistent the data is. It's the constraint that makes or breaks every analysis or automation project, and it's almost always assessed after the project has started.

In practiceNo model fixes bad input data. It passes it on, faster and more confidently.

Quantisation

quantizzazione

Lowering a model's numerical precision so it fits in less memory and runs on modest hardware, at the cost of a small loss in quality.

In practiceIt's what makes it possible to run a model in-house rather than on someone else's servers, when confidentiality matters more than top-end quality.

R

RAG

retrieval-augmented generation · generazione aumentata dal recupero

The technique where the system, before answering, searches the business's own documents for the relevant passages and then builds its answer from them. It cuts down hallucinations, lets you cite the source and keeps knowledge up to date without retraining anything.

In practiceYou don't always need it. When there are only a few documents and they rarely change, putting them straight in front of the model is simpler, cheaper and just as effective. RAG earns its place when the volume outgrows the context window.

Event log

registro eventi · log

The written record of what a system has done: when, to what, and with what result. It's the only thing that turns a failure into something you can reconstruct rather than argue about.

In practiceA system without a log isn't simpler, just harder to fix, and you pay the difference all at once, the first day something goes wrong.

Data residency

residenza dei dati

The physical location where data is stored and processed. It matters for the rules on data transfers and, in practice, for the answer you give a customer who asks.

Data processor

responsabile del trattamento

Anyone who processes personal data on the controller's behalf and under its instructions: the business software provider, the web host, the email service. The relationship must be set out in writing, under Art. 28 GDPR.

REST

The most common convention for building APIs on the web. It's not a technology but an agreement on how to ask and how to answer, and it's why systems written by different companies can talk to each other.

Neural network

rete neurale

A computing structure very loosely inspired by neurons: layers of connected units, each with a weight that gets adjusted during training until the system produces the expected output.

The name suggests a brain and sets the wrong expectations: there's no understanding involved, just statistical optimisation on a huge scale.

Reactivation

riattivazione · win-back

Getting back in touch, systematically, with people who used to buy and have stopped. It's the cheapest pool of potential sales a business has, because those people already know you.

In practiceAll you need is the list of customers who've gone quiet. Almost no SME has one, because no system ever noticed the silence.

Retry

ritentativo

Automatically repeating an operation that failed, usually with longer and longer waits in between. It's needed because most failures are temporary: a slow network, a service that's briefly busy.

In practiceA retry without idempotence doesn't protect anything. It multiplies the mistakes.

RPA

robotic process automation

Automation that mimics what a user does on screen: opening windows, copying fields, clicking buttons. It was invented to work with systems that have no API.

It works, but it's fragile: all it takes is an interface moving a button and it breaks without a sound.

In practiceIt's the last resort, not the first. If there's an API, use that: it's cheaper to maintain and doesn't break every time the supplier releases an update.

S

SaaS

software as a service

Software sold on subscription and hosted by the provider: you pay to use it, not to own it. The upside: no infrastructure to manage. The downside: the subscription never ends and your data lives on someone else's premises.

In practiceThe question to ask before signing: if we wanted to leave in two years, would our data come out in a format we could use? If the answer is vague, the real price isn't the subscription.

Scalability

The ability to cope with more load without redesigning everything. It's a real virtue and a common excuse: designing for volumes you don't have costs money today for a problem you may never have.

Black box

scatola nera

A system where you can see what goes in and what comes out, but not the reasoning in between. It comes from how deep models are built, not from any choice made by the people using them, and you manage it with controls around it rather than by solving it.

In practiceThe practical fix isn't opening the box. It's recording what went in and what came out, so that at least a mistake can be traced.

SEO

ottimizzazione per i motori di ricerca

The set of choices that make a website understandable to search engines and useful to the people searching. The technical side is the bare minimum: beyond that, what counts is answering a question better than the pages already ranking for it.

Time series

serie storica

A sequence of measurements over time. It lets you tell a real change from a normal fluctuation, and mixing the two up is the most expensive mistake in business decisions.

In practiceOne bad month isn't a trend. But without the history there's no way to know that, and you end up fixing things that weren't broken.

Legacy system

sistema preesistente

An old system, often clunky, that still carries real work and holds years of the business's history. Techies use the term as an insult; in an SME it's nearly always the only place where the data is complete.

In practiceAnyone who suggests replacing it as step one is selling you a migration, not a solution. Connect it first, and replace it last, if at all.

Multi-agent system

sistema multi-agente

A setup where several specialised agents split a job between them and pass the results along. It raises what's possible in theory, and multiplies the places where something can break without anyone noticing.

In practiceAlmost never the right first step. A single agent that does one thing well covers the vast majority of an SME's real cases at a tenth of the complexity.

SLA

service level agreement · livello di servizio

The written commitment on response times, availability and remedies. Without an SLA, “support included” means whatever the supplier decides it means on the day you need it.

In practiceThe line that matters isn't the response time. It's what happens if the response time isn't met.

SLM

small language model · modello linguistico compatto

A small language model, which can run on modest hardware and sometimes inside the business itself. Less capable overall, but often good enough for narrow, repetitive tasks.

In practiceIt's the technical answer to a confidentiality requirement: when data can't leave the company network, a compact model in-house beats a powerful one outside.

Explainability

spiegabilità · XAI

The ability to work out why a system produced a particular output. For some kinds of model it's straightforward; for deep models it's partial, and pieced together after the fact.

In practiceYou need it when something is challenged, not when it's being built. The day a customer asks why their case was handled that way, you either have a record or you have a problem.

Human oversight

supervisione umana · human in the loop

A setup where the system prepares, suggests and puts the case together, but a person approves it before the action has any effect outside the business. It doesn't come from distrusting the machine. It comes from keeping responsibility where the law and common sense put it.

In practiceThe working rule: anything involving a customer's name, money or an outgoing message goes through an approval. Everything else can run by itself.

Human in the loop: what it means, the three levels and the AI Act →

T

Conversion rate

tasso di conversione

The share of contacts who take the next step: from visit to enquiry, from enquiry to quote, from quote to order. Measure it step by step, because an overall rate hides exactly the point where you're losing people.

Temperature

temperatura

The setting that controls how far a model strays from the most likely continuation. Low: predictable, repeatable answers. High: more varied, less controllable ones.

In practiceA bit of variety helps with sales copy. To pull a figure out of an invoice you want the lowest setting possible: the last thing you need is creativity on the taxable amount.

Uptime

tempo di attività

The percentage of time a service is actually available. Read it alongside what happens in the remaining percentage: an outage at night and one during working hours don't carry the same weight.

Time to first response

tempo di prima risposta

The minutes between an enquiry arriving and the first reply, human or automated. Of all sales indicators, it gives you the most impact for the least effort to measure.

In practiceAlmost no business measures it, and almost every business thinks it's quick. When someone finally looks, the gap between perception and reality is routinely measured in hours.

Real time

tempo reale

Processing at the moment the event happens, usually triggered by a webhook. It adds complexity, and it's only worth it where waiting has a cost.

Ticket

richiesta di assistenza

A single request, tracked from the moment it's opened until it's closed. What matters isn't the numbering, it's that no request can go missing because someone was on holiday.

Data controller

titolare del trattamento

The person or organisation that decides the purposes and means of processing personal data, and is accountable for it. Between a business and its software provider, the controller is nearly always the business.

In practiceIt's why you can't outsource responsibility by buying a tool: the work moves, the obligation stays put.

Token

The unit a model uses to measure text: fragments of words, shorter than a word and longer than a letter. It matters because it's the unit you pay by, and the unit the context window is measured in.

In practiceItalian uses more tokens than English to say the same thing. At an SME's volumes the difference is negligible; on a system handling thousands of documents a month, it isn't.

Tokenisation

Splitting text into tokens before the model processes it. It's why models fail at seemingly trivial tasks like counting the letters in a word: they don't see letters, they see fragments.

Audit trail

traccia di controllo

A tamper-proof record of who did what, and when. It's what lets you reconstruct a decision months later, in front of a customer or an inspector.

In practiceIn a system with automated parts, it's the difference between “the system decided” and an actual explanation.

Transformer

The architecture, introduced in 2017, that almost every modern language model is built on. Its central idea, the attention mechanism, lets the model weigh which parts of the text matter for predicting what comes next.

Trigger

innesco

The event that sets an automated flow going: an email arriving, a form being submitted, a record changing status in the business software, a time of day.

In practiceChoosing the right trigger is half the project. A reminder tied to the wrong moment stops being a reminder and becomes a nuisance, which people learn to ignore.

U

Tool use

uso di strumenti · function calling

The mechanism by which a model, instead of answering in words, calls a real function: looking up a price list, creating a record, sending a request to another system. It's what turns a language model into an agent.

In practiceThis is where the nature of the risk changes. While the model is only talking, the worst mistake is a wrong sentence; from here on, a mistake is an action.

V

Customer lifetime value

valore del cliente nel tempo · LTV

What a customer is worth in total over the whole relationship, not on a single order. It's the number that makes spending money to win them make sense.

W

Webhook

The mechanism by which one system alerts another the moment something happens, instead of being checked over and over. It's the difference between getting a phone call and checking your phone every two minutes.

In practiceIt's what decides whether an enquiry gets handled in ten seconds or in the next overnight run.