The words, without the jargon around them.

A

Churn

abbandono

The share of customers who stop buying in a given period. In businesses with repeat purchases it is the indicator that decides their fate more than any acquisition campaign.

In practiceA customer who stops rarely says so. They leave in silence, and the only ones who notice are those who look at buying behaviour rather than cancellations.

Training

addestramento

The process by which a model derives its behaviour from a set of examples, adjusting its own parameters until the error goes down. It is the expensive phase, and it happens only once: afterwards, the model is only queried.

In practiceNo SME trains a model from scratch. The realistic operations are two others: giving it the right context at the moment of the question, or fine-tuning on repetitive cases.

AI agent

agente AI

A system that doesn't just answer. It receives a goal, decides for itself which steps to take and uses external tools (reading a mailbox, writing to a spreadsheet, querying management software) until the task is done.

What separates it from a chatbot isn't skill, but permission to act. And that is why an agent has to be given boundaries before it is switched on.

In practiceA chatbot tells you the status of an order. An agent looks it up in the management software, finds it held at customs, prepares the email to the customer and leaves it as a draft 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 with staggered application. It doesn't regulate the technology itself: it classifies uses by level of risk, and derives the obligations from that.

Prohibited practices and the obligation of staff AI literacy have applied since February 2025; the obligations on high-risk systems come later. For the vast majority of SMEs the real impact is modest: answering a customer or filling in a document isn't a high-risk use. It becomes one when a system selects people, assesses access to an essential service or affects a right.

In practiceThe question isn't “do we use AI?”, but “at which point does an automated decision touch a person?”. If the answer is nowhere, the heavy obligations don't concern you. If there is one, they concern that one.

Algorithm

A defined sequence of steps for solving a problem. In public debate the word has taken on a more sinister meaning than the real one: most of the algorithms that run a company are rules written by someone and readable by anyone.

In practice“The algorithm decides” isn't an answer. If nobody in the company can explain the rule, it means it was adopted without being read.

Hallucination

allucinazione

The case in which a model produces a fluent, plausible and false statement, with the same confidence with which it produces a true one. An LLM predicts the probable text, not the correct one: hallucination is the normal product of that mechanism, not a malfunction of it.

It can't be eliminated. It is contained, by giving the system a knowledge base to draw on (RAG) and putting a person at the point where the error would be costly.

In practiceThe risk isn't that the system gets it wrong, but that it gets it wrong with confidence, inside a reply to a customer that nobody reread.

Test environment

staging · sandbox

A copy of the system where a change is tried out before it is applied to the real system. It costs little and prevents the most expensive category of error: the one discovered in production.

In practiceFor an automation that sends messages to real customers, testing in production means sending a wrong communication to a flesh-and-blood person.

Master data

anagrafica

The basic data that identifies customers, suppliers and products, and that every other system refers to. When the master data is dirty, every analysis built on top inherits the dirt.

In practiceThe same customer written in four different ways leaves you with revenue split into four, which no report puts back together.

API

application programming interface

A program's service entrance: the point where another program can ask it for data or actions without going through the interface designed for humans. It is what makes integration possible.

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

Unsupervised learning

apprendimento non supervisionato

Training on unlabelled data, in which the system looks for recurring structures on its own: groups of similar customers, anomalous behaviour, hidden patterns.

In practiceUseful for discovering, not for deciding. A grouping always has to be reread by someone who knows the business, because the machine finds regularities even where they mean nothing.

Reinforcement learning

apprendimento per rinforzo · RL

Training by trial and reward: the system tries, receives a score and corrects its strategy. It is the method behind game-playing systems and, in modified form, behind aligning language models with the desired behaviour.

Supervised learning

apprendimento supervisionato

Training on examples that are already labelled: each input is paired with the right answer. It is the most common case in business applications and the most reliable, because the goal is defined without ambiguity.

In practiceSomeone has to do the labelling. It is the part of human work that quotes forget to price.

Attention

attenzione

The mechanism by which a model decides, word by word, which other parts of the text to look at. It is what makes it possible to hold together a reference at the start of a document and the sentence that recalls it at the end.

Two-factor authentication

2FA · MFA

Verifying identity with two independent elements: a password and a temporary code. It is the single measure with the best ratio between protection gained and hassle introduced.

In practiceIt should go on the email accounts first, not on the management software. Email is where the passwords for everything else get reset.

Automation

Having a machine carry out a step that someone does by hand today, with the same outcome and without their time. It doesn't require artificial intelligence: most of the automations that give hours back in an SME are deterministic rules, not models.

In practiceBefore asking which AI you need, it is worth asking whether the step is needed. Automating a useless step makes it a faster useless step.

Workflow automation

automazione dei flussi

The automation of a sequence of steps that runs across different people and systems, not of a single isolated action. It is the level at which real hours are freed up.

In practiceIn an SME time isn't lost in the single gestures: it is lost in the handovers between one gesture and the next. That is where it is measured, in the wait between one task and the following one.

Autonomy

autonomia · livello di autonomia

How much a system can do without asking. It should be thought of as a scale, not a switch: propose, prepare, act and notify, act silently.

It is the most important design decision and almost always the only one taken out of habit rather than by choice.

In practiceYou go up one step at a time, after the previous step has worked for weeks. Whoever starts from the top step hasn't built a more advanced system: they have skipped the test.

Self-hosted

autonomo dal fornitore

Software installed and run on your own infrastructure, often starting from open-source projects. It removes the subscription and adds maintenance: it isn't free, it is paid for in hours.

B

Backup

copia di sicurezza

A copy of the data kept separately. The number that counts isn't how often it is made, but when it was last verified by actually restoring it.

In practiceA backup that has never been tested is worth as much as a hope, plus an automatic schedule.

Knowledge base

base di conoscenza

The organised body of what a company already knows (product sheets, price lists, terms of sale, internal rules, the answers given a thousand times) put into a form a system can query.

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

In practiceIt already exists, but scattered: in two people's heads, in a shared folder, in ten years of sent emails. The work isn't creating it, it's collecting 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, as well as an ethical one, when the system affects people: staff recruitment, access to credit, the assessment of an application.

In practiceA system trained on how you have decided so far will also repeat the mistakes you have made so far, only faster and on more cases.

C

Chain of thought

catena di pensiero · CoT

The technique of asking the model to set out the intermediate steps before the conclusion. It improves reliability on multi-step tasks and makes the error traceable, because you can see where the reasoning went off course.

Chatbot

An interface that answers in the form of a conversation. The term describes how a system presents itself, not how capable it is: behind the same little window there can be a tree of fixed answers written ten years ago or a model connected to the company'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 are none, the chatbot has no work to do.

Cloud

nuvola

Running software on third-party servers, paid for by usage. It removes the purchase of machines and introduces a dependence: the availability of your work depends on a contract and a connection.

Queue

coda

A waiting list between two systems that work at different speeds. It absorbs peaks and ensures nothing is lost when one produces faster than the other consumes.

Long tail

coda lunga

The set of infrequent, very specific searches which, added together, outnumber the generic ones. They are less contested and bring visitors who already know what they want.

In practice“Business automation” is contested by everyone. “How to reduce no-shows in a restaurant” is searched only by people who have that problem.

Lead nurturing

coltivazione dei contatti

The sequence of useful contacts towards someone who isn't ready to buy yet, to stay present until the moment they are.

In practiceThe difference between nurturing and pestering isn't in the frequency: it's in whether each message contains something useful even to someone who will never buy.

Business continuity

continuità operativa · disaster recovery

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

Copilot

copilota · assistente

A system that works alongside a person instead of replacing them: it proposes, completes, summarises, and the last word stays with the operator.

In practiceIt is the form in which AI enters a company with the least resistance, because it takes nothing away from anyone. The gain is also smaller and harder to measure.

Customer acquisition cost

costo di acquisizione · CAC

How much it costs on average to win a new customer, counting everything: advertising, sales time, tools. It should always be read together with lifetime value: on its own it says nothing.

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

Encryption

Turning data into a form that can't be read without the key. It should be split into two moments: in transit, while the data travels, and at rest, while it is stored. Many providers state the first and say nothing about the second.

CRM

customer relationship management

The single register of commercial relationships: who each contact is, what they have been told, what needs to happen next and when. A CRM doesn't produce sales: it makes visible what today lives in different heads and separate mailboxes.

In practiceThe typical failure isn't technical. A CRM that nobody updates becomes a second job and is abandoned within three months: if it doesn't fill itself with what already happens, it won't last.

Dashboard

cruscotto

The screen that gathers the indicators in a single place. Its value isn't in the charts, but in the fact that someone looks at it and decides something.

In practiceA dashboard nobody has opened for three weeks has already said what there was to know: those numbers didn't serve any decision.

D

Vector database

database vettoriale

An archive that indexes documents by meaning rather than by exact word, using embeddings. It is the infrastructure RAG rests on when documents become numerous.

In practiceIt is an answer to a volume problem. Introducing it before you have that volume adds one more piece to maintain and solves nothing.

Dataset

insieme di dati

The collection of examples on which a model is trained or evaluated. The quality of the dataset determines the model's behaviour more than any later technical choice.

In practiceIt holds on a small scale too: if you build a system that classifies customer requests, its quality will depend on the examples you gave it, not on the skill of whoever configures it.

Structured data for the web

schema markup · JSON-LD

The annotations hidden in a page that tell search engines what its contents are: a company, a frequently asked question, an article, a definition. They don't change what the user sees; they change what the machine understands.

Unstructured data

Everything else: emails, PDFs, contracts, voice notes, the photo of a delivery note taken crooked in the warehouse. It is most of the information that circulates in a company.

The reason AI has become relevant to businesses is precisely this: unstructured data, which used to require a person, can now partly be processed by machine.

In practiceAlmost all administrative work consists of turning unstructured data into structured data by hand: reading a PDF and copying it into the management software.

Structured data

What sits in rows and columns: an order table, a customer master file, a bank statement. It is the form in which data can be directly queried, added up and compared.

Deduplication

Finding and merging records that represent the same entity written in different ways. It is tedious, it isn't spectacular, and in practice it unlocks more value than many AI projects.

Deep learning

apprendimento profondo

The branch of machine learning based on neural networks with many layers. It is the technique that made image recognition, high-quality machine translation and modern language models possible.

In exchange for the power it asks for two things: quantity of data and computing power. Neither is found inside an SME, which is why these models are rented rather than built.

Distillation

distillazione

The technique by which a large model is used to train a smaller one that imitates its behaviour on a narrow task. The result is cheaper and faster, within that task.

E

Batch processing

elaborazione a lotti

Processing many items all together at fixed intervals: every night, every hour. It costs less and simplifies the design, at the price of waiting.

In practiceIt is perfect for accounting and terrible for a customer request. The question is always: who is waiting for this result, and how long can they wait?

Embedding

rappresentazione vettoriale

Translating a text into a sequence of numbers that represents its meaning. It serves so that two sentences that say the same thing in different words come out close, and can therefore be found one from the other.

In practiceIt is why a well-built search finds “the document for shipping outside Europe” even if the document says “free export declaration”.

ERP and management software

enterprise resource planning · gestionale

The system in which the company records what really happens: orders, stock, production, invoices. In many Italian SMEs it is the oldest piece of the infrastructure and at the same time the most important, because it holds the accounting truth.

In practiceIt should be connected, not replaced. Replacing management software that works is a months-long project that produces zero new value: you pay to get back to where you started.

ETL

extract, transform, load

The sequence by which data is extracted from one system, brought to a common format and loaded into another. It is the plumbing of any data project.

F

Few-shot

apprendimento per esempi

The technique of including a few worked examples in the instruction, so that the model infers the expected format instead of having to imagine it. It is the cheapest way to make an output consistent.

In practiceThree real examples taken from your documents are worth more than a page of instructions on how you would like them to be.

Fine-tuning

riaddestramento mirato

Retraining an existing model on your own data, so that it consistently adopts a format, a tone or a repetitive task. It costs money, requires examples in quantity and has to be redone when the underlying model changes.

In practiceIt is needed much less often than people think. In most cases what's missing isn't a trained model: it's the right information put in front of an ordinary model.

Context window

finestra di contesto

The maximum amount of text a model can keep in front of it at once: the question, the attached documents, the conversation so far. Beyond that limit there is no error, there is silent forgetting.

In practiceIt is why an assistant that worked perfectly for twenty exchanges starts contradicting itself at the fortieth.

Workflow

flusso di lavoro

The real sequence of steps by which something comes into the company, is worked on and goes out: who receives it, who decides, who carries it out, who checks.

In practiceAlmost nobody has written it down. Automating before writing it down means putting speed into disorder, and fast disorder is worse than slow disorder: it does damage before anyone notices.

Single source of truth

fonte unica di verità

The principle by which every piece of information has only one official place, and everyone else reads it there instead of keeping a copy. It is the only structural defence against misalignment.

In practiceThe symptom that 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 European rules on the protection of personal data. They apply to every system that processes data relating to an identifiable person, so to an AI system too: the fact that processing is automated 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. They are two contractual questions, not two opinions.

Guardrail

vincolo di sicurezza

The limits written into an automated system: what it can't say, what it can't do, beyond which threshold it must stop and call a person. They aren't a later add-on: they are as much part of the design as the main function.

In practiceThe question that puts them to the test: what happens if the system receives an absurd request? If the answer is “we didn't think about it”, there is no guardrail.

I

Idempotence

The property by which running the same operation twice produces the same result as running it once. It sounds like a textbook detail; it is what stops the same invoice being sent twice when a system retries.

In practiceIt is the difference between an automation that restarts on its own after a failure and one that, in restarting, does the damage twice.

Inference

inferenza

Using an already trained model to get an answer. It is what happens every time the system works, and it is the recurring cost item: training is paid for once, inference every day.

Data entry

inserimento dati

Manually copying information from one medium to another. It is the most repetitive, least visible and most easily eliminated activity there is in a company.

In practiceIt appears in no profit and loss account as a cost item, and that is why it survives for years: nobody has ever added it up.

Integration

The connection between two systems so that data entered once appears wherever it is needed, without being copied. It is the least spectacular work and the one that gives back the most hours.

In practiceEvery manual copy is two costs in one: the time of whoever does it and the error of whoever sooner or later gets it wrong. The second is always discovered later than the first.

Artificial intelligence

intelligenza artificiale · AI

The field that builds systems capable of tasks we associate with human intelligence: recognising, classifying, predicting, generating, deciding. It is a broad label, and that is the first reason why sales conversations about “AI” go round in circles.

Under the same word sit things that are very different 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 verifiable. The useful question is which task, on which data, with what accepted margin of error.

Generative artificial intelligence

AI generativa · GenAI

The family of systems that produce new content (text, images, code, translations) instead of just classifying or predicting. It is the part of AI that since 2023 has made the subject visible outside technical departments, and it is only one part.

In practiceMany operational problems don't need to generate anything: they need to read, sort and route. Duller tasks, more reliable and often more profitable.

J

Jailbreak

aggiramento dei limiti

A wording built on purpose to get a model to break out of the rules it has been given. The business risk has little to do with embarrassing content, and a lot to do with the assistant that, prompted well, reveals how your processes work or what is in its knowledge base.

In practiceA public assistant shouldn't have access to anything you aren't prepared to publish.

JSON

The format in which two systems exchange structured data: name-value pairs, readable by a human too. It is the lingua franca of modern integrations.

K

KPI

key performance indicator · indicatore chiave

The measure chosen to know whether something is going as it should. A KPI is only useful if someone can act on it: if nobody can change it, it isn't an indicator, it's news.

In practiceBetter three indicators looked at every week than twenty on a dashboard nobody opens.

L

Latency

The time between a request to a system and its response. On an overnight process it counts for nothing; in front of a customer who is waiting it counts for more than the quality of the answer.

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

Lead

contatto commerciale

A contact who has shown interest but hasn't bought yet: a request from the website, a WhatsApp message, a business card collected at a trade fair. It is the most fragile moment of the whole sales cycle, and almost always the least watched.

In practiceA lead isn't lost to a no. It is lost to silence: nobody answers quickly enough and, by the time someone does, they have already bought elsewhere.

Rate limit

limite di frequenza

The ceiling on the requests a service accepts in a given interval of time. Exceeding it doesn't produce an obvious error: it produces rejected requests which, if nobody logs them, become work that vanishes.

LLM

large language model · modello linguistico di grandi dimensioni

The engine underneath conversational AI tools. Trained on enormous quantities of text, it predicts the most probable continuation of what it has in front of it.

From this mechanism come both its usefulness and its structural flaw: it predicts the probable, it doesn't verify the true. See hallucination.

In practiceIt should be treated as a very fast, very well-read collaborator with no access at all to your data, until you give it that access in a controlled way.

llms.txt

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

In practiceIt is the equivalent of robots.txt for an era in which a growing share of people ask an assistant instead of searching.

M

Machine learning

apprendimento automatico · ML

The approach in which a system's behaviour isn't written by hand rule by rule, but derived from past examples. It is the broadest family, which also contains generative AI.

In practiceIt requires history. If the company hasn't recorded past outcomes there is nothing to learn from, and that, not cost, is the real reason certain projects don't get started.

No-show

mancata presentazione

The appointment or booking that doesn't turn up and doesn't cancel. It costs twice: the place stays empty and nobody else could take it.

In practiceWhat weighs isn't the absence itself, but the fact that it is discovered too late to resell that place.

Metadata

Data that describes other data: when a document was created, by whom, which file it belongs to. It is what makes an archive searchable instead of merely full.

Vanity metric

metrica di vanità

A number that grows, feels good and guides no decision: visits, impressions, total contacts. It can be recognised by a single question: if this number doubled tomorrow, what would we do differently?

In practiceIf the answer is “nothing”, it isn't an indicator worth reporting.

Middleware

strato intermedio

The software that sits between two systems and makes them talk, translating formats and rules. It exists because neither of the two was designed with the other in mind.

Migration

Moving data and processes from one system to another. It is the category of project that overruns most often, for a recurring reason: the real work isn't moving the data, it's discovering in how many different ways it had been entered.

Data minimisation

The principle by which only the data needed for that purpose is processed, and not one item more. It is the most effective remedy there is, because data you don't have can't be taken from you.

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

Baseline

misura di partenza

The value of an indicator before intervening. Without it, there is no honest way to say whether something worked.

In practiceIt is the only step that can't be recovered afterwards. Whoever doesn't measure before won't be able to prove anything after, not even to themselves.

Model

modello

The product of training: a mathematical object that, given an input, produces an output. “Model” refers to the specific thing you are 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 the outcome of a flow that used to work.

In practiceA system that goes into production states which model it uses and in which version. If it doesn't, the day the behaviour changes nobody will know why.

Monitoring

monitoraggio

The continuous watch over a running system, with an alert when something moves outside the expected values. The question that defines the level: who gets alerted, and how long before they notice?

In practiceThe typical failure isn't the system that breaks, but the one that breaks and still looks switched on.

Multimodal

multimodale

A model that processes several types of input: text, images, audio, documents. It is the property that makes the crooked photo of a delivery note workable, without going through a separate recognition step.

In practiceIt is the leap that makes document automation realistic for businesses that receive paper, photos and scanned PDFs instead of tidy files.

N

No-code and low-code

Tools that let you build automations by connecting blocks on a screen instead of writing code. They lower the entry threshold; they don't lower the need to know what you are building.

In practiceThe typical risk is the flow built by a single person, never documented, which stops working when that person changes role. The fragility isn't in the tool, but in the fact that nobody knows any more what it does.

Normalisation

Bringing heterogeneous data back to a common format: dates, units of measure, company names, currencies. It is the precondition for any comparison.

O

OCR

optical character recognition · riconoscimento ottico

Automatically reading the text inside an image or a scanned PDF. A decades-old technology, made much more reliable by recent models, especially on crooked, faded or handwritten documents.

In practiceIt is the first step of document processing: without it, every photographed delivery note remains an image only a human eye can read.

Omnichannel

omnicanale

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

In practiceThe sign that it's missing: the customer has to explain their situation all over again depending on the channel they write from.

On-premise

in sede

Running on machines inside the company. It costs more to manage and remains the right answer when data can't leave the company network.

Orchestration

The layer that decides which system steps in, in what order, with which data and, above all, what happens when a step fails. It is the difference between a series of automations and an infrastructure.

In practiceThe question that separates the two: when a piece breaks at two in the morning, does the system notice and tell someone, or do you find out from the customer who complains?

Optimisation for AI assistants

AEO · answer engine optimization

Adapting content so that a conversational assistant can cite it, not just so it can rank in a list of results. It rewards clear structure, verifiable statements and declared sources.

In practiceIt is why a source cited in full is worth more than a catchy phrase: the first can be repeated by a machine, the second can't.

Overfitting

sovradattamento

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

In practiceIt is the statistical equivalent of the colleague who handles the files they have already seen perfectly and freezes on the first one that is different.

P

Parameters

parametri · pesi

The internal values a model adjusts during training, which encode its behaviour. Their number, which in large models is counted in billions, is used as a measure of power.

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

Sales pipeline

imbuto di vendita

The set of open deals, ordered by stage. It serves to answer a question almost no SME can answer with a number: how much work lies ahead, and how much of it will fall by the wayside.

Least privilege

privilegio minimo

The principle by which every user and every system receives only the permissions its task requires. It applies to people and, more and more, to automated agents.

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

Prompt

istruzione

The instruction given to a model. In a properly built system the user doesn't rewrite it every time: it is part of the system, written once, tested, versioned and changeable without touching the rest.

In practiceIf the quality of the answer depends on how the employee on shift wrote 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 underlying instruction that defines an assistant's role, limits and tone, applied to every conversation before any question from the user. It is where the company rules are written: what it can say, what it must refuse, when it must hand over to a person.

In practiceIt is the most important document of an automated assistant, and almost always the only one nobody has ever read in full.

Prompt engineering

The practice of wording instructions that consistently produce the desired result. The term sounds grander than it is: for the most part it consists of being explicit about what was being taken for granted.

Prompt injection

iniezione di istruzioni

The attack in which a hostile instruction is hidden inside content the system will read (the body of an email, a PDF, a web page) to make it carry out actions nobody asked for. It is the characteristic vulnerability of systems that read material coming from outside.

In practiceIt is the technical reason why a system that reads incoming email shouldn't also have the power to send, delete or pay without approval.

Lead scoring

punteggio dei contatti

Assigning a priority to each contact on the basis of observable signals: sector, size, source, behaviour. It serves to send sales time, the scarcest resource, where it is most likely to produce results.

In practiceIt isn't for discarding low-scoring contacts. It is for deciding the order in which to call, which is a decision taken anyway, today by instinct.

Touchpoint

punto di contatto

Every moment in which a customer meets the company: the website, a phone call, a quote, an invoice, a message after the sale.

In practiceThe perceived experience matches the worst of the touchpoints, not their average. It pays to look for that one, rather than improving the ones that already work.

Q

Data quality

qualità del dato

How complete, correct, up to date and consistent the data is. It is the constraint that decides the fate of every analysis or automation project, and it is almost always assessed after the project has started.

In practiceNo model corrects wrong input data. It propagates it, faster and with more confidence.

Quantisation

quantizzazione

Reducing a model's numerical precision so that it fits in less memory and runs on modest hardware, accepting a small loss of quality.

In practiceIt is what makes it possible to run a model inside the company instead of on someone else's servers, when confidentiality weighs more than absolute quality.

R

RAG

retrieval-augmented generation · generazione aumentata dal recupero

The technique by which the system, before answering, searches the company's documents for the relevant passages and then formulates the answer using them. It reduces hallucinations, makes it possible to cite the source and keeps knowledge updatable without retraining anything.

In practiceIt isn't always needed. When the documents are few and stable, putting them directly in front of the model is simpler, cheaper and just as effective. RAG is justified when the volume exceeds the context window.

Event log

registro eventi · log

The written trace of what a system has done: when, on what, with what outcome. It is the only thing that makes a failure reconstructable rather than debatable.

In practiceA system without a log isn't simpler, only harder to repair, and the difference is paid all at once on the first day something goes wrong.

Data residency

residenza dei dati

The physical place where data is stored and processed. It matters for transfer obligations and, in practice, for the answer you will give a customer who asks.

Data processor

responsabile del trattamento

Whoever processes personal data on behalf of the controller, within its instructions: the management software provider, the hosting, the email service. The relationship has to be governed in writing, under Art. 28 GDPR.

REST

The most widespread convention for building APIs on the web. It isn't a technology, it's an agreement on how to ask and how to answer, and it is why systems written by different companies manage to talk to each other.

Neural network

rete neurale

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

The term evokes the brain and creates the wrong expectation: there is no understanding, there is statistical optimisation on a large scale.

Reactivation

riattivazione · win-back

Systematically contacting again those who bought in the past and no longer buy. It is the cheapest commercial pool a company has, because those people already know you.

In practiceAll it takes is the list of those who have gone cold. Almost no SME has it, because no system ever noticed the silence.

Retry

ritentativo

Automatically repeating a failed operation, usually with increasing waits. It is needed because most failures are temporary: a slow network, a service that is momentarily busy.

In practiceA retry without idempotence protects nothing. It multiplies the errors.

RPA

robotic process automation

Automation that imitates a user's gestures on screen: it opens windows, copies fields, presses buttons. It was born to talk to systems that have no APIs.

It works and it is fragile: all it takes is an interface moving a button for it to break silently.

In practiceIt is the last resort, not the first. If there is an API, you use that: it costs less to maintain and doesn't break at every update from the supplier.

S

SaaS

software as a service

Software sold as a subscription and hosted by the provider: you pay for use, not ownership. Advantage: no infrastructure to manage. Disadvantage: the subscription never ends and the data lives at someone else's place.

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

Scalability

The ability to handle an increase in load without redesigning everything. It is a real virtue and a frequent excuse: designing for volumes you don't have costs money today for a problem you might never have.

Black box

scatola nera

A system whose inputs and outputs can be observed but not its internal reasoning. It comes from how deep models are built, not from a choice by whoever adopts them, and it is governed with the controls around it rather than solved.

In practiceThe practical remedy isn't opening the box, but recording what went in and what came out, so that an error at least remains reconstructable.

SEO

ottimizzazione per i motori di ricerca

The set of choices that make a website understandable to search engines and useful to people searching. The technical part is a minimum condition: above it, what counts is answering a question better than whoever already ranks.

Time series

serie storica

A succession of measurements over time. It serves to tell a real change apart from a normal fluctuation, which is the most expensive confusion in business decisions.

In practiceOne bad month isn't a trend. Without the history, though, there is no way of knowing, and you end up fixing what wasn't broken.

Legacy system

sistema preesistente

An old system, often awkward, which nonetheless carries real work and contains years of company history. The term is used as a technical insult; in an SME it is almost always the only place where the data is complete.

In practiceWhoever proposes replacing it as the first intervention is selling a migration, not a solution. It should be connected first and, if anything, replaced last.

Multi-agent system

sistema multi-agente

An architecture in which several specialised agents split a job between them and pass the result along. It increases theoretical capability and multiplies the points where something can break without anyone noticing.

In practiceAlmost never the right first step. A single agent that does one thing and does it well solves 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 counts isn't the response time, but what happens if it isn't met.

SLM

small language model · modello linguistico compatto

A small language model, which can run on modest hardware and sometimes inside the company. Less capable in absolute terms, often sufficient on narrow, repetitive tasks.

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

Explainability

spiegabilità · XAI

The ability to reconstruct why a system produced a certain output. On some families of models it is simple; on deep models it is partial and reconstructed after the fact.

In practiceIt is needed at the moment of a dispute, not of construction. The day a customer asks why their file was handled that way, either there is a trace or there is a problem.

Human oversight

supervisione umana · human in the loop

The architecture in which the system prepares, proposes and puts together the file, but a person approves before the action has effects outside the company. It doesn't come from mistrust of the machine, but from the need to keep responsibility where the law and common sense put it.

In practiceThe practical rule: everything that touches a customer's name, money or an outgoing communication goes through an approval. Everything else can run on its own.

T

Conversion rate

tasso di conversione

The share of contacts who take the next step: from visit to request, from request to quote, from quote to order. It should be measured step by step, because an overall rate hides precisely the point where things are lost.

Temperature

temperatura

The parameter that governs how far a model moves away from the most likely continuation. Low: predictable, repeatable answers. High: more varied and less controllable answers.

In practiceFor sales copy a little variety helps. To extract a number from an invoice you need the lowest possible value: the last thing you want is creativity on the taxable amount.

Uptime

tempo di attività

The percentage of time in which a service is actually available. It should be read together with what happens in the remaining percentage: an outage at night and one during working hours don't weigh the same.

Time to first response

tempo di prima risposta

The minutes between the arrival of a request and the first human or automatic reply. It is the sales indicator with the highest ratio between impact and cost of measurement.

In practiceAlmost no business measures it, and almost all think they are fast. The gap between perception and data, when you finally look, is regularly a matter of hours.

Real time

tempo reale

Processing at the instant the event happens, usually triggered by a webhook. It costs more in complexity and is only justified where waiting has a price.

Ticket

richiesta di assistenza

The single request tracked from opening to closing. What counts isn't the numbering, but the fact that no request can disappear because someone was on holiday.

Data controller

titolare del trattamento

Whoever decides the purposes and means of processing personal data, and is answerable for it. In the relationship between a business and a software provider, the controller is almost always the business.

In practiceIt is why responsibility isn't outsourced by buying a tool: the work moves, the obligation doesn't.

Token

The unit by which a model measures text: pieces of words, shorter than a word and longer than a letter. It matters because it is the unit you pay by and by which the context window is measured.

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

Tokenisation

Splitting a text into tokens before the model processes it. It is why models get apparently trivial tasks wrong, such as counting the letters in a word: they don't see letters, they see pieces.

Audit trail

traccia di controllo

The unalterable record of who did what and when. It is what makes it possible to reconstruct a decision months later, in front of a customer or an inspection.

In practiceIn a system with automated parts it is the difference between “the system decided it” and an explanation.

Transformer

The architecture introduced in 2017 on which almost all modern language models rest. Its central idea, the attention mechanism, lets the model weigh which parts of the text matter for predicting the next one.

Trigger

innesco

The event that starts an automated flow: an email arriving, a form being filled in, a row changing status in the management software, a time of day.

In practiceChoosing the right trigger is half the project. A reminder hooked to the wrong moment stops working as 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: checking a price list, creating a row, sending a request to another system. It is what turns a language model into an agent.

In practiceIt is the point where the nature of the risk changes. As long as the model talks, the worst error is a wrong sentence; from here on the error becomes an action.

V

Customer lifetime value

valore del cliente nel tempo · LTV

How much a customer is worth overall for the duration of the relationship, not on a single order. It is the number that makes it sensible to spend on acquiring them.

W

Webhook

The mechanism by which one system notifies another the instant something happens, instead of being polled continuously. The difference is between receiving a phone call and checking the phone every two minutes.

In practiceIt is what decides whether a request is processed in ten seconds or at the next overnight run.