MIXED
Glossary
AI and data glossary for business
The terms you come across when talking about artificial intelligence, data and digital, explained in a few lines for people who need to decide, not code. Every entry has an example applied to a business. Use the search or jump to a section.
53 terms
Artificial intelligence 22 terms
From the basics to the terms that show up in suppliers' quotes.
- Artificial intelligence (AI)
- A set of techniques that let software perform tasks that would require human intelligence: understanding text, recognising an image, making a prediction, taking a decision. It isn't "a program" but a family of approaches, from simple to complex.In business: the system that reads customer emails and routes them to the right department uses AI; the Excel sheet with formulas doesn't.
- Machine learning
- Branch of AI in which software learns rules from example data instead of having them written by a programmer. The more good-quality examples it has, the better it works.In business: given three years of orders, the system learns to forecast how much of each item you'll sell next month.
- Deep learning
- Machine learning based on neural networks with many layers. It's the technique behind image and voice recognition and behind language models. It needs lots of data and computing power.In business: the camera that spots a defective part on the line uses deep learning.
- LLM Large Language Model
- AI model trained on huge amounts of text, able to understand and generate natural language: summarise, translate, answer, classify, write. ChatGPT, Claude and Gemini are interfaces to LLMs.In business: an LLM reads a badly written quote request and extracts product, quantity and urgency.
- Generative AI
- AI that produces new content — text, images, code, audio — from a request. It differs from "analytical" AI, which classifies or predicts without creating.In business: a draft reply to a complaint, generated from the customer's history, is generative AI.
- Prompt
- The instruction or question given to a generative AI model. The quality of the answer depends heavily on how precise and contextualised the prompt is, and whether it includes examples.In business: "Summarise this contract in 5 points, highlighting penalties and deadlines" is a prompt; "summarise" is a worse one.
- Token
- Unit of text (word fragments, roughly 4 characters in English) with which language models read and write. Model usage costs are calculated in tokens, both input and output.In business: having a model read 1,000 PDF invoices a month costs in proportion to the tokens on those pages.
- Context window
- Maximum amount of text a model can consider in a single request (question + documents + answer). Beyond the limit, the model "can't see" the rest.In business: to analyse a 300-page tender specification you need to split it or use a RAG system, not paste it all in.
- RAG Retrieval-Augmented Generation
- Technique in which the model, before answering, retrieves the relevant company documents (procedures, contracts, manuals) and uses them as a source, citing them. It reduces hallucinations and lets you use AI on your own data without retraining it.In business: "what warranty did we give customer X in 2023?" — the system finds the contract and answers with the reference.
- Embedding
- Numerical representation of a text (or image) that captures its meaning: similar texts have similar embeddings. It's what makes it possible to search "by meaning" and not just by exact word.In business: searching for "water leak" also finds tickets about "seepage" and "burst pipe".
- Vector database
- Store optimised for finding similar content via embeddings. It's the search engine behind a RAG system.In business: the office's 5,000 technical documents, indexed in a vector database, can be queried in natural language.
- Fine-tuning
- Additional training of an existing model on company data to specialise it (tone, terminology, format). More expensive and delicate than RAG; useful in specific cases, not for "teaching the model our documents".In business: a model tuned to classify 50 complaint types using your internal vocabulary.
- AI agent
- AI system that, given a goal, plans and executes a sequence of actions using tools: reads emails, queries the management system, fills in forms, sends replies. It must be designed with clear limits and human oversight.In business: an agent receives an order by email, checks stock availability, creates the job and proposes the confirmation to send.
- Hallucination
- Invented but plausible answer produced by a generative AI model: dates, figures, legal references that don't exist. It can be reduced with RAG, guardrails and human checks; it can't be eliminated entirely.In business: an assistant citing a non-existent law in a reply to a customer. That's why answers must be checked.
- Guardrail
- Rules and checks that limit what an AI system can say or do: forbidden topics, data not to reveal, actions that require approval.In business: the customer assistant can't promise discounts or discuss competitors, and hands over to a human if the customer is angry.
- Human in the loop
- Approach in which a person checks or approves the AI's proposals before they become actions. It's the safest way to introduce AI into a process and measure its reliability.In business: the AI prepares the draft reply, the employee reads it and clicks "send".
- Computer vision
- AI that interprets images and video: recognises objects, defects, people, text; counts and measures. It needs real photos for training and a calibration period.In business: automatically checking that PPE is worn on site, or counting packages on a pallet.
- OCR Optical Character Recognition
- Turns scans and photos of documents into software-readable text. Today often combined with AI to also understand the document's structure (tables, fields).In business: supplier invoices in PDF become rows in the management system without retyping anything.
- NLP Natural Language Processing
- Techniques with which software analyses and understands human language: classifying, extracting information, measuring tone. LLMs are the latest evolution of NLP.In business: automatic analysis of reviews to understand what customers complain about most.
- Inference
- Phase in which an already-trained model is used to produce answers. It's the part you pay for per use (per token or per call) and that determines a project's recurring costs.In business: each email classified by the assistant is one inference; 3,000 emails a month are 3,000 inferences.
- Open-source vs proprietary model
- Open-source models can be downloaded and installed on your own servers (full control of data, infrastructure costs); proprietary ones are used through the provider's service (simpler and often more capable, with data leaving the company under the contract's terms).In business: to analyse medical records or trade-secret data, an on-premise model is worth considering.
- Predictive analytics
- Use of historical data and statistical or machine learning models to estimate future events: demand, failures, late payments, customer churn. It's only as good as the data behind it.In business: predicting which customers will pay late so reminders can go out earlier.
Data 14 terms
The terms of data strategy: how data is collected, organised and turned into decisions.
- Data strategy
- Decisions about which data to collect, where to keep it, who is accountable and how to use it to decide. For an SME it isn't a document: it's one source per data point, a few KPIs, an automatic dashboard and minimal quality rules.In business: deciding that attendance originates from site clock-ins and nobody retypes it anywhere else.
- Structured and unstructured data
- Structured: organised in tables with defined fields (orders, customer records, attendance). Unstructured: emails, PDFs, photos, recordings. Around 80% of a company's data is unstructured, and that's where generative AI makes the difference.In business: the management system holds structured data; the certified-email inbox holds unstructured data.
- Database
- System that stores data in an organised, shared and reliable way, with rules on what can go in and who can change it. It's the right alternative to Excel used as an archive.In business: customer records in a database: one version, updated by everyone, no diverging copies.
- Data warehouse
- Central database that collects data from several company systems (management system, CRM, e-commerce, production), organised for analysis and reporting. It's the foundation of management dashboards.In business: sales, costs and hours worked in the same place, to calculate margin per customer.
- Data lake
- Store that keeps raw data of every kind, including unstructured, for future analysis. Suited to large volumes; for most SMEs a well-built data warehouse is enough.In business: machine logs, site photos and documents kept "as they are" for later analysis.
- ETL / ELT Extract, Transform, Load
- Process that extracts data from sources, transforms it (cleans, standardises) and loads it where it will be analysed. It's the "invisible" work that weighs most in data projects.In business: every night orders are extracted from the management system, reconciled with customer records and loaded into the data warehouse.
- Data pipeline
- Automatic chain of steps that takes data from source to dashboard or model, with no manual intervention. If a step requires "someone to export the spreadsheet", it isn't a pipeline.In business: clock-in → hours per site → labour cost → job margin, updated every hour.
- Single source of truth
- Principle that each data point has one official source everyone refers to. The opposite is the same number having three different values depending on who calculates it.In business: the official price list is in the management system; salespeople's Excel copies don't count.
- Master data
- Core, relatively stable company data — customers, suppliers, items, employees, sites — that all systems must share. If every system has its own records, integrations break.In business: customer "Rossi Ltd" must have the same code in the management system, the CRM and invoicing.
- KPI Key Performance Indicator
- A number that measures whether a goal is being met. A good KPI has a written, agreed formula, a source and a frequency.In business: "estimated margin per job" with a defined formula, updated weekly by the system.
- Dashboard
- Screen showing KPIs updated automatically, with charts and tables. A dashboard someone has to "prepare" every month is a report, not a dashboard.In business: the owner opens the dashboard on Monday morning and sees late jobs, deadlines at risk and forecast cash.
- Business intelligence (BI)
- Tools and practices to turn company data into reports, analysis and dashboards that support decisions (Power BI, Looker and Metabase are BI tools).In business: analysing which items have negative margin after discounts.
- Data quality
- Degree of completeness, correctness, consistency and timeliness of data. Poor quality = wrong decisions and failed AI projects. It's maintained with mandatory fields, checks and an owner per source.In business: 30% of customers without a VAT number in the CRM is a data quality problem.
- Data governance
- Rules and responsibilities on who can create, change and use data, and how. In large companies it's a function; in an SME it's a checklist with a name next to each source.In business: only purchasing creates new suppliers; salespeople can propose them, not enter them.
Software and automation 11 terms
The terms you find in quotes and in incentive guidelines.
- API Application Programming Interface
- Interface through which two pieces of software exchange data automatically, without manual exports. "Does it have an API?" is the first question to ask before buying software that will need to talk to others.In business: the online shop creates the order in the management system via API, without anyone retyping it.
- Integration
- Connection between different systems (management system, CRM, e-commerce, payroll) so they share data without retyping. It's the item that most often inflates a project's time and cost.In business: hours clocked on site arrive automatically in the payroll provider's software.
- Cloud and SaaS Software as a Service
- Software and infrastructure used over the internet, paid by subscription, with no servers on premises; updates and backups are the provider's job. Check where the data resides and how it can be exported.In business: a SaaS management system is used from the browser and the app, from the office and the site.
- On-premise
- Software and data installed on company-owned servers, at its own site. Maximum control, but hardware, maintenance and security costs are yours.In business: a typical choice for particularly sensitive data or contractual constraints with certain customers.
- ERP / Management system
- Software that runs the company's core processes: orders, inventory, accounting, production, staff, jobs. "Management system" is the everyday term; ERP usually refers to broader systems.In business: V-Site is the management system for companies that run construction sites: attendance, deadlines, vehicles, costs.
- CRM Customer Relationship Management
- Software to manage customers, contacts, deals, quotes and support. If salespeople keep customers "in their phone", a CRM is the first step of any data or AI project on sales.In business: a customer's complete history, available before a visit.
- RPA Robotic Process Automation
- Software that repeats actions on other programs in place of a person: opening a file, copying fields, clicking. Useful with old systems without APIs; fragile if screens change.In business: a "robot" that downloads bank statements from the bank's portal every morning and loads them into accounting.
- Workflow
- Defined sequence of steps and owners through which a process is carried out — who does what, after what, with which approvals. Digitising a workflow means the system moves it forward on its own. See how to automate a process.In business: leave request → manager approval → attendance calendar update, with no emails.
- Low-code / No-code
- Tools to build applications and automations with little or no programming. Great for prototypes and simple processes; for core systems, watch licence costs and exit constraints.In business: a form to collect fault reports from the field, built in an afternoon.
- 4.0 interconnection
- Requirement of Italy's Transizione 4.0 and 5.0 incentives: the asset or software must exchange data with company systems in a documented way (expert or technical report). Without it, the tax credit isn't granted.In business: production data-collection software qualifies if connected to the management system and the machine.
- MVP / Pilot Minimum Viable Product
- Minimum version of a project built to quickly check whether it works before investing more. The best way to reduce the risk of an AI or data project.In business: one month of AI assistant on a single inbox, measuring the hours saved, before extending it.
Security and regulation 6 terms
What you need to know before giving data to software or to an AI.
- GDPR
- EU regulation on the protection of personal data: what can be collected, for what purposes, for how long, with what safeguards and rights for individuals. It applies to data on customers, employees and suppliers, including when an AI processes it.In business: before having customer emails analysed by an AI service, you need a legal basis and a provider with a compliant contract.
- Anonymisation and pseudonymisation
- Techniques to remove or mask people's identifying data before analysing it or passing it to an AI. Anonymised = no longer traceable; pseudonymised = traceable only with a separate key.In business: analysing complaints after replacing names and tax codes with neutral identifiers.
- AI Act
- EU regulation on artificial intelligence, being applied progressively, that classifies systems by risk level and imposes increasing obligations (transparency, oversight, documentation). Most SME uses are minimal or limited risk; uses on people (recruitment, employee evaluation) are high risk.In business: an assistant that sorts emails has minimal obligations; a system that assesses job candidates has many.
- Backup and disaster recovery
- Safety copies of data, kept separately, and a plan to restore operations after a failure, a mistake or an attack. The right question isn't "do we back up?" but "how fast do we get going again, and how much do we lose?".In business: ransomware encrypts the server on Friday; with tested backups, you're working on Monday.
- MFA multi-factor authentication
- Login that requires, besides the password, a second element (phone code, app, physical key). It's the single security measure with the best cost/benefit ratio.In business: even if the owner's password is stolen, nobody gets into the email without the phone.
- Phishing
- Email or message scam that imitates trusted senders (bank, supplier, the boss) to steal credentials or trigger payments. With generative AI the messages have become far more convincing.In business: a "change of bank details" from a real supplier, perfectly written: always verify by phone.
No term found. Missing an entry? Write to us and we'll add it.
Can't find a term? The glossary grows with our clients' questions. Let us know and we'll add it in the next revision.
Want to know which of these tools your company needs?
Free 30-minute consultation: we start from your process, not from the technology.
Learn more