The software company that created FT Mercati Go to ftmercati.com

Glossary

AI glossary for companies: the terms explained

What AI agent, LLM, RAG, token, shadow AI, deployer, high risk, Official Cash Settlement and premium mean: short definitions for the people who run a company, with the source and the guide that goes into more depth.

In brief

The glossary explains in plain English the terms of artificial intelligence (AI) that a company comes across when it adopts it: large language models (LLMs), agents, RAG, tokens, the words of the AI Act (deployer, high risk, AI literacy) and those of metals, such as the LME Official Cash Settlement and the premium. Each entry is a definition of a few sentences that stands on its own, with an example, the official source and the guide that goes into more depth.

This glossary brings together the definitions of the artificial intelligence (AI) terms that the people who run a company need in order to decide, and of the metals terms that go into contracts and into the ERP, the company's management software. Each entry has a definition that stands on its own, a practical example, the official source and, where there is one, the guide on this site that goes into more depth. There are 54 entries, in alphabetical order.

Action log log · audit trail · traceability

The action log is the record, written automatically by the system, of what an artificial intelligence system has done: which request it received and from whom, which tools and data it used, what it proposed, who confirmed and when. It serves to reconstruct an error, to show who decided what and to comply with the rules: for high-risk AI systems the AI Act requires the automatic recording of events (logs) over the whole lifetime of the system and obliges the deployer to keep them for at least six months. For business agents it is a rule of common sense even when it is not an obligation, in compliance with the rules on the remote monitoring of workers.

In practice. Every order created by the agent has a matching line with the source PDF, the fields extracted, the user who confirmed and the time; if a customer disputes a quantity, it takes a minute to trace who saw what.

Source: AI Act, Article 12: record-keeping; AI Act, Article 26, paragraph 6: deployers keep the logs for at least six months. Read more: AI agents: what they are and what they can do · AI Act: what a manufacturing company must do · Company AI policy: usage rules and an example · ChatGPT, Copilot and Claude: using AI at work · AI assistant on company documents: how it works · Metals hedging: what AI can do · AI inside the ERP: how to connect an agent.

Agent tools (tool use) function calling · tool calling

Tools are the functions an AI agent can call to do something beyond writing text: look up a customer in the CRM, read a file, query a price service, create a draft record. The model does not run them by itself: it decides which tool to use and with which parameters, and the program that hosts it makes the call and returns the result to it (tool use or function calling); MCP (Model Context Protocol) is the open standard for connecting tools to models. The list of tools is closed and decided by whoever designs the system: that is where it is established what the agent can and cannot do, and which actions require human confirmation.

In practice. The CRM agent has four tools: find customer, read order, prepare draft quote, record activity; none of these sends emails or deletes data, and none can be added by the agent itself.

Source: Anthropic, Tool use with Claude; Anthropic, glossary: MCP (Model Context Protocol). Read more: AI agents: what they are and what they can do · CRM with AI: what it does and what it does not · AI inside the ERP: how to connect an agent.

AI agent agentic AI

An AI agent is a program that uses a language model to reach an assigned goal: it interprets the request, chooses among the tools it has been granted (searching an archive, reading a document, filling in a form), uses them one after another, checks the result and replies. It differs from a chatbot, which only produces text, because it carries out actions on the systems it is connected to. For this reason, in a company it works with a closed list of tools, the same permissions as the user who is using it, human confirmation for actions that change data and a log of every step. The form "agentic AI" refers to the same way of working.

In practice. In a company that receives orders as PDFs, the agent reads the document, looks up the customer in the ERP (the company's management software), prepares the draft order and submits it to a person, who confirms it.

Source: Anthropic, Building effective agents (December 2024); AI Act, Article 3, point 1: definition of AI system. Read more: AI agents: what they are and what they can do · AI agent for purchasing: what it is, what it does · ChatGPT, Copilot and Claude: using AI at work · AI assistant on company documents: how it works · Metals hedging: what AI can do · AI inside the ERP: how to connect an agent.

AI assistant on company know-how AI assistant on company documents · internal company chatbot

An AI assistant on company know-how is an artificial intelligence system that answers questions using the company's documents, procedures and data, not the general knowledge the model was trained on. For each question it retrieves the relevant passages from the authorised internal sources (the RAG technique) and passes them to the model, which answers by citing them; it respects the permissions of the person asking and changes nothing. It is the tool through which the experience that today sits in a few people's heads becomes available for everyone to consult, with the source alongside.

In practice. A new employee asks "how do we handle a return for a defect?" and receives the procedure in force, with a link to the document and to its version; if the procedure does not exist, the assistant says so instead of inventing one.

Source: Anthropic, glossary: RAG (retrieval augmented generation). Read more: AI agents: what they are and what they can do · AI assistant on company documents: how it works.

AI consultant artificial intelligence consultant · AI strategist

An AI consultant, that is, an artificial intelligence consultant, is the professional or firm that helps the people who run a company decide where artificial intelligence delivers a measurable result, which process to start from, with which data, rules and spending, and that leaves a step-by-step plan (roadmap) that can also be carried out with others. In Italy it is a profession not organised into ordini or collegi, the bodies that govern the regulated professions (Law No. 4/2013): there is no official register, anyone can practise it and qualification against a technical standard is voluntary. Since 30 April 2026 the Italian standard UNI 11621-8:2026 has described the professional profiles of those who work in artificial intelligence.

In practice. Before engaging an AI consultant, check the company register extract (in Italy, the visura camerale), references for systems in production, independence from providers declared in writing, and the contract: scope, data, ownership of what remains, exit. Whoever resells a product is not a neutral third party; Italy's AI law (Law No. 132 of 23 September 2025, Article 13) requires consultants, like anyone in the liberal professions, to tell the client which AI systems they use in their work.

Source: Law No. 4 of 14 January 2013 (Normattiva, in Italian); UNI 11621-8:2026, professional role profiles in the artificial intelligence sector (in Italian); Law No. 132 of 23 September 2025, Article 13 (Normattiva, in Italian). Read more: AI consultant: what they do and how to choose one · How to bring AI into your company: where to start.

AI literacy

Artificial intelligence (AI) literacy is the set of skills, knowledge and understanding that allows those who provide or use AI systems to deploy them in an informed way and to be aware of their opportunities, risks and possible harm. Article 4 of the AI Act, which has applied since 2 February 2025 and was rewritten in July 2026, asks providers and deployers to take measures to support its development among their staff, in proportion to skills, roles and context of use; it does not impose courses, hours or certificates, nor a minimum level to be guaranteed. It also applies to those who only use general-purpose tools such as ChatGPT. In Italy, Legislative Decree No. 179 of 7 October 2026 sets no fine for this obligation, but requires employers to provide information and training on the specific risks of AI systems that affect work organisation and workplace safety, and requires the ordini professionali (the bodies that govern Italy's regulated professions) to provide compulsory courses with a minimum number of hours.

In practice. A company whose sales department uses ChatGPT to translate quotes must tell those who use it which data not to enter and that the results must be checked, and keep a record of what it has done: an internal log is enough, no certificate is required.

Source: AI Act, Article 3, point 56 and Article 4; European Commission, questions and answers on AI literacy; Legislative Decree No. 179 of 7 October 2026, Articles 23, 42 and 47 (in Italian). Read more: AI literacy: what the AI Act requires of companies · AI Act: what a manufacturing company must do · AI consultant: what they do and how to choose one · ChatGPT, Copilot and Claude: using AI at work · AI assistant on company documents: how it works.

API application programming interface

An API is the access point through which one program uses the functions of another program or service, according to defined rules on what is asked for and what is returned. In artificial intelligence it is the channel through which company software (the ERP, the company's management software; the CRM; the website) sends text or documents to a model and receives the answer without going through a chat; it is also the channel through which an ERP receives market prices every day from an authorised source. The APIs of the main model providers have contract terms that differ from those of personal accounts, in particular on the use of data for training.

In practice. The ERP asks via API for the monthly average of the LME price and writes it into the price lists; the CRM agent asks the model via API to summarise a customer's history.

Source: NIST, glossary: Application Programming Interface; AI Act, recital 97: general-purpose AI models are also placed on the market through APIs. Read more: AI in the cloud or on-premise: how to choose · How to connect LME prices to your ERP · AI inside the ERP: how to connect an agent.

Back-to-back hedging back-to-back pricing · physical hedging

Back-to-back hedging, or physical hedging, is the practice of aligning the pricing formulas of metal purchases with those of product sales: same reference (for example the monthly average of the LME Official Cash Settlement), same quotation period, same quantities. If the formulas match, the price risk passes from one contract to the other without derivatives; if mismatches of timing, quantity or reference remain, an exposure remains to be managed. It is the first hedge of an SME that works with metals, often done without calling it that.

In practice. A metal processor that sells at the LME average of the month of delivery and buys the metal at the same average has hedged the metal price, but not the premium or the euro-dollar exchange rate; if a customer asks for a fixed price for six months, the back-to-back breaks and that contract has to be handled separately.

Source: London Metal Exchange, LME Digest: Physical and financial hedging, beginners guide (back-to-back pricing). Read more: Metals hedging: what AI can do · Metal price indexation clauses: a practical guide.

Basis risk

Basis risk is the risk that the reference price of the physical contract and that of the hedging instrument move differently: the hedge then offsets only in part. For those who work with metals in Italy it arises from three differences: the metal premium, which the exchange contract does not cover; a contractual reference other than the LME Official price, such as a Milan Index, a formula of one's own or a single day instead of the average; the euro-dollar exchange rate, because LME contracts are in dollars. It is reduced by using the same reference and the same period in the physical contract and in the hedge, and by hedging the exchange rate separately.

In practice. A company that sells on an Italian copper index and hedges with a contract on the LME Official price takes on a basis risk that a monthly check can measure: the difference between the two references over the quotation period, recorded with source and date, shows how much the hedge really offset.

Source: London Metal Exchange, LME Digest: Physical and financial hedging, beginners guide (basis risk, price references); London Metal Exchange, Contract types (contracts in dollars); FT Mercati, Euro-dollar exchange rate and metal prices: which rate to use. Read more: Metals hedging: what AI can do.

Chunk chunking

A chunk is the portion of a document, usually a few hundred tokens, into which the archive is split for indexing by an artificial intelligence assistant; each chunk remembers which document and which page it comes from, so that the answer can cite them. The size and the way of cutting affect the quality of the answers: Microsoft suggests starting from chunks of about 512 tokens with a 25% overlap and following headings and paragraphs, while Amazon uses about 300 tokens by default. A chunk without context, which does not say which document or which revision it is about, is a frequent cause of wrong answers.

In practice. "Tolerance ±0.05 mm" in a chunk that does not say which drawing and which revision it refers to leads to a confident and wrong answer; this is why the best systems add a line of context (document, section, date) to each chunk before indexing it, and documents are cut along their structure, not at a fixed number of characters.

Source: Microsoft Learn, Chunk large documents for vector search solutions in Azure AI Search (512 tokens, 25% overlap); Anthropic, Introducing Contextual Retrieval (chunks, context, exact search); AWS, Content chunking and parsing for Bedrock knowledge bases (chunks of about 300 tokens). Read more: AI assistant on company documents: how it works.

Company AI policy AI usage rules · AI acceptable use policy

A company AI policy, that is, the company's rules on artificial intelligence, is the internal document that sets out which artificial intelligence tools may be used in the company, with which data, with which checks and who is responsible for what. No law requires it under this name, but it is the simplest way to comply with the AI Act (AI literacy among staff, transparency) and the GDPR, and to govern the spontaneous use of tools (shadow AI); in Italy, the Ministry of Labour's guidelines of December 2025 recommend it. For a small or medium-sized enterprise one page is enough: permitted tools, data that must never be entered, the obligation to check results, the person to ask, how a new tool is approved.

In practice. "Use only the company accounts of these two tools; do not paste in personal data, quotes, contracts or drawings; reread every text produced before sending it; for a new tool, ask the person in charge."

Source: AI Act, Article 4; Italy's Ministry of Labour and Social Policies, Ministerial Decree No. 180 of 17 December 2025: guidelines for the implementation of AI in the world of work (in Italian). Read more: Company AI policy: usage rules and an example · AI literacy: what the AI Act requires of companies · AI adoption in companies: top-down or bottom-up · AI consultant: what they do and how to choose one · ChatGPT, Copilot and Claude: using AI at work.

Connector Copilot connector

A connector is the ready-made component that links an artificial intelligence assistant or an automation platform to an external system (an ERP, the company's management software; a CRM; an archive) without writing code. There are two types: those that copy and index the data into the assistant's system, and those that query it in real time and leave it where it is, with the user's identity and permissions; the difference matters for confidentiality and for how up to date the data is. Connectors to the most widely used ERP systems are published by the providers or their partners; those that can also write do so, in the products we have reviewed, only with the user's approval and with the activity logged.

In practice. A federated connector reads the customer's balance from the ERP at the moment a person asks the assistant for it, without copying it into any index; a synced connector instead copies the product data sheets every night to make them searchable. Before switching one on, ask the provider which of the two types it is and what it logs.

Source: Microsoft Learn, Microsoft 365 Copilot connectors overview (synced and federated); Microsoft Learn, Federated connectors overview. Read more: AI inside the ERP: how to connect an agent · AI assistant on company documents: how it works.

Context window context length

The context window is the amount of text a language model can keep in view in a single run: the request, the attached documents, the previous exchange and the answer itself, measured in tokens. It is the model's working memory, distinct from the data it was trained on: what is not in the window does not exist for the model. A larger window makes it possible to work on long documents, but more context does not automatically give better answers, and it costs more.

In practice. A forty-page contract with its three addenda fits into a single request; the complete archive of contracts does not, and one turns to targeted retrieval of the passages that are needed (RAG).

Source: Anthropic, glossary: context window; Anthropic, documentation on context windows. Read more: AI in the cloud or on-premise: how to choose · What an AI project costs: what it depends on · AI assistant on company documents: how it works.

Data processing agreement (DPA) data processing addendum

A data processing agreement (DPA), also called a data processing addendum, is the contract, provided for in Article 28 of the GDPR, under which a provider that processes personal data on behalf of a company undertakes to follow the company's instructions, to guarantee security, to use only authorised sub-processors, to help the company with data subjects' rights and to delete or return the data at the end of the service. For artificial intelligence assistants it exists in the plans for organisations (ChatGPT Business, Enterprise and API, Microsoft Copilot with a work account, Claude Team, Enterprise and API, Gemini in Google Workspace) and not in the personal ones, where the provider also processes the content for its own purposes.

In practice. Before switching on an AI assistant for its staff, the company asks for or signs the provider's DPA and reads three lines: whether the content is used for training, where the data is processed and how long it is kept. It is the document that turns the promises on the sales page into commitments; without it, the personal data of customers and employees does not go into the tool.

Source: Regulation (EU) 2016/679 (GDPR), Article 28: data processor (EUR-Lex); Microsoft Learn, Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat; OpenAI, Enterprise privacy at OpenAI; Anthropic, Commercial Terms of Service. Read more: ChatGPT, Copilot and Claude: using AI at work · AI in the cloud or on-premise: how to choose · AI assistant on company documents: how it works · AI inside the ERP: how to connect an agent.

Data residency

Data residency is the provider's commitment to store, and in some cases process, the customer's data within a defined geographical area, for example the European Union. It is not the same as "the data never leaves": almost all providers allow for exceptions (system data, web searches, some third-party models) and the coverage depends on the plan. For artificial intelligence assistants, as of October 2026, Microsoft keeps European tenants within the EU Data Boundary, except for web searches and Anthropic's models; OpenAI offers residency in Europe for ChatGPT Enterprise, Edu and the API, and the Business plan is not among those eligible; Anthropic stores data in the United States, and Claude is used in Europe through Amazon Bedrock or Google Cloud; Google has data regions only in some editions of Workspace.

In practice. A company that wants its data in Europe writes this in the request for quotation and checks it in the contract, not in the advertising: for ChatGPT that means Enterprise or the API, for Copilot a tenant registered in Europe, for Claude a cloud with a European region, for Gemini an edition of Workspace with data regions. Providers' pages change often: the check has to be repeated at every renewal.

Source: Microsoft Learn, What is the EU Data Boundary?; OpenAI Help Center, Data residency and inference residency for ChatGPT; Anthropic Privacy Center, Where are your servers located?; Anthropic, Data residency (platform documentation). Read more: ChatGPT, Copilot and Claude: using AI at work · AI in the cloud or on-premise: how to choose · AI assistant on company documents: how it works.

Data retention retention policy

Data retention is the length of time a provider keeps conversations, uploaded files and results after use, and the rule by which it then deletes them. In the plans that artificial intelligence assistants offer to organisations, it is normally decided by the administrator: in Microsoft 365 with Purview retention policies; in Google Workspace by choosing between three, eighteen or thirty-six months; in Claude Team and Enterprise the data is kept indefinitely unless a custom period is set, which only Enterprise can do, with a minimum of thirty days; in ChatGPT Enterprise with the workspace settings. In personal plans the provider's rule applies: in ChatGPT, deleted chats are removed within thirty days; in Claude, retention is thirty days or five years depending on the choice made about training; in Gemini, automatic deletion is set to eighteen months.

In practice. A company that uses an assistant to prepare quotes decides, before switching it on, how long conversations should be kept: long enough to reread them if a quote is disputed, no longer than necessary for the personal data they contain. The rule chosen is written into the usage rules (AI policy) and, where needed, into the privacy notice for customers.

Source: Microsoft Learn, Data, Privacy, and Security for Microsoft Copilot (retention and Microsoft Purview); Google Workspace, Generative AI in Google Workspace Privacy Hub (retention of conversations); Anthropic, Updates to Consumer Terms and Privacy Policy (retention of 30 days or 5 years); OpenAI Help Center, Data controls in ChatGPT; Anthropic Help Center, Configure custom data retention controls for Enterprise plans. Read more: ChatGPT, Copilot and Claude: using AI at work.

Deployer and provider provider and deployer

In the AI Act, the provider is whoever develops an artificial intelligence system, or has one developed, and places it on the market or puts it into service under its own name or trademark; the deployer (in plain terms, the user) is whoever uses an AI system under its own authority, except for personal non-professional use. The provider's obligations are much heavier than the deployer's, especially for high-risk AI systems. A company that uses tools made by others is almost always a deployer, but it becomes a provider if it puts into service under its own name a system it has had developed, if it makes a substantial modification to a high-risk AI system, or if it changes the intended purpose of a system so that it becomes high-risk (Article 25). In Italy both are supervised by the National Cybersecurity Agency (ACN) under Legislative Decree No. 179 of 7 October 2026.

In practice. A company that uses an assistant inside its ERP (the company's management software) is a deployer; the software company that supplies it under its own trademark is the provider; if the company takes a general-purpose assistant and uses it to rank CVs, it takes on the obligations of the provider of a high-risk AI system.

Source: AI Act, Article 3, points 3 and 4: definitions of provider and deployer; AI Act, Article 25: when a distributor, importer or deployer becomes a provider; Legislative Decree No. 179 of 7 October 2026: market surveillance authority and penalties in Italy (in Italian). Read more: AI Act: what a manufacturing company must do · AI literacy: what the AI Act requires of companies · AI consultant: what they do and how to choose one · ChatGPT, Copilot and Claude: using AI at work.

Digital Innovation Hub (DIH) European Digital Innovation Hub · EDIH

A Digital Innovation Hub (DIH) is a guidance and support organisation that helps businesses assess and adopt digital technologies, artificial intelligence included, with services for assessment, training, technology testing and access to funding. The European Digital Innovation Hubs (EDIHs) are those selected and funded by the European Commission's Digital Europe Programme; alongside them operate the DIHs of trade associations and of other initiatives. In Italy, the Ministry of Labour's Guidelines on artificial intelligence in the world of work (Ministerial Decree No. 180/2025) mention a free digital assessment delivered through the DIH network.

In practice. Before paying for consultancy, a company looks in the European Commission's catalogue (as of 12 October 2026: 57 hubs in Italy, with a few possible duplicates, 13 funded by the Digital Europe Programme) for the hub that lists its sector and artificial intelligence among its technologies, reads the list of services, which vary from hub to hub, and asks for an assessment: it comes away with a picture of where it stands. The hub will not, however, carry out the project in its place.

Source: European Commission, catalogue of European Digital Innovation Hubs; Italy's Ministry of Labour and Social Policies, Ministerial Decree No. 180 of 17 December 2025, Guidelines for the implementation of AI in the world of work (in Italian); Legislative Decree No. 179 of 7 October 2026, Article 29: the European Digital Innovation Hubs in the Spazio di sperimentazione italiano per l'IA, Italy's AI regulatory sandbox (in Italian). Read more: AI consultant: what they do and how to choose one.

Draft action

A draft action is the way of putting an artificial intelligence agent to work in which whatever it produces (an order, a purchase invoice, a posting, a payment reminder) is saved as a document that is not yet valid, which a person reviews and confirms. It is the rule that providers of ERP systems (company management software) such as Microsoft have chosen for their agents: the quote before the order, because it does not touch planning and availability; the unposted invoice before posting; the new supplier blocked until someone verifies it. The log keeps a record of who proposed and who confirmed.

In practice. The agent reads the customer's PDF order and creates a draft order with customer, items, quantities and dates; the salesperson opens it, checks it, corrects it and confirms. If an item code cannot be found, the draft stays on hold and flagged: never "the closest match" in place of the right code.

Source: Microsoft Learn, Sales Order Agent overview (the quote as a safe document, approved messages); Microsoft Learn, Payables Agent overview (unposted invoices, new supplier blocked). Read more: AI inside the ERP: how to connect an agent · AI agents: what they are and what they can do.

Embedding vector embedding · text embedding

An embedding is the representation of a text (a sentence, a paragraph, a document) as a series of numbers calculated by a model, so that texts with similar meanings have numbers that are close together. It is used to search "by meaning": the question is transformed in the same way and compared with the document chunks already indexed, and relevant passages are found even when they share no words with the question. It does not replace the exact search for codes and abbreviations, which an embedding model can miss: serious systems run the two searches side by side.

In practice. The query "thickness tolerance" finds the paragraph that talks about "permitted deviation in thickness", but the error code "E-47" in the manual has to be searched for literally: that is why an assistant on company documents uses semantic search together with exact-word search.

Source: AWS, What is Retrieval-Augmented Generation? (embeddings and vector store); Anthropic, Introducing Contextual Retrieval (embeddings, exact search, the error code example); OpenAI, platform Retrieval guide (semantic and hybrid search). Read more: AI assistant on company documents: how it works.

EMIR (for a non-financial counterparty) European Market Infrastructure Regulation · Regulation (EU) No 648/2012

EMIR is the European regulation on over-the-counter (OTC) derivatives, central counterparties and trade repositories (Regulation (EU) No 648/2012, amended by EMIR Refit in 2019 and by EMIR 3 in 2024). An industrial company that enters into a hedging derivative with a bank is a "non-financial counterparty". It may calculate its positions every twelve months: if it stays below the clearing thresholds it has no clearing obligation; if it exceeds them, or does not calculate, it must notify ESMA and its national competent authority and clear; contracts that objectively reduce the risks of its commercial activity are not counted. It must apply timely confirmation and reconciliation to contracts that are not cleared and give the bank accurate data, LEI code included, because for derivatives between a bank and a company below the thresholds, reporting to the trade repository falls to the bank. In Italy the authority for notifications on the thresholds is Consob, the financial markets regulator.

In practice. An SME that hedges copper with a forward contract from its bank keeps a register of derivatives with the linked physical contract, quantity, period and reference: it serves to show that the hedge reduces a commercial risk, to answer the bank with accurate data and to redo the annual calculation of positions without hunting for documents at the last minute.

Source: Regulation (EU) No 648/2012 (EMIR), consolidated version of 24 December 2024, Articles 2, 9, 10 and 11 (EUR-Lex); Delegated Regulation (EU) No 149/2013, Article 10: contracts that objectively reduce commercial risks (EUR-Lex); Consob, EMIR (in Italian); ESMA, EMIR Reporting. Read more: Metals hedging: what AI can do.

ERP (enterprise resource planning) ERP system · business management software

An ERP (enterprise resource planning) system is the management software with which a company manages its resources by sharing information across its functional areas: accounting, purchasing, warehouse, production, sales. It can be a standard product, a customised product or software built in-house. In Italy it is used by 49.5% of businesses with at least ten employees and by 59.9% of manufacturing businesses (ISTAT, Italy's national statistics institute, December 2025). For an artificial intelligence agent it is the system where the company's data sits and where, with permissions and confirmations, it can read, propose in draft and, by exception, write.

In practice. The agent reads from the ERP the month's copper requirement and the open orders, prepares the draft order to the supplier with quantity and quotation period, and leaves it with purchasing for confirmation: the ERP remains the system of record, and the agent works within its permissions.

Source: ISTAT, Imprese e ICT 2025 report (in Italian): definition of ERP in the report's glossary, share of businesses; Eurostat, E-business integration (isoc_eb_iip): enterprises with ERP software, 2025. Read more: AI inside the ERP: how to connect an agent · How to connect LME prices to your ERP.

Fine-tuning

Fine-tuning is the additional training of an already trained language model, with your own examples, so that it absorbs their style, vocabulary or type of task. It produces a modified version of the model, which has to be maintained and updated when the base model changes. For most office tasks it is not needed: well-written instructions (prompts) and documents retrieved at the time of the question (RAG) achieve the same result without touching the model, and they can be corrected quickly instead of through new training.

In practice. A company that wants the answers in its own quote format does not train a model: it gives it the document template and three examples in the instructions.

Source: Anthropic, glossary: fine-tuning; AI Act, recital 97: models further modified or fine-tuned. Read more: AI in the cloud or on-premise: how to choose · What an AI project costs: what it depends on.

Fixed-scope pilot AI pilot project · pilot project

A fixed-scope pilot is a company's first artificial intelligence project, limited to a single process, with written boundaries (what it does and what it does not do), an internal contact person, the necessary data accessible and a measure of the result decided before starting. It lasts a few weeks and serves to check, on real work, time saved, errors and user satisfaction before spending to extend it; whatever is left outside the scope goes on the list for later. It is the opposite both of "let's do everything at once" and of the pilot that never ends because nobody decided how to measure it. It should not be confused with the "AI regulatory sandboxes" of the AI Act (Article 57), which the authorities set up for providers.

In practice. Three weeks on a single task, "read the PDF orders of the ten main customers and fill in the draft in the ERP" (the company's management software), counting the minutes saved per order and the errors corrected by the member of staff.

Source: AI Act, Article 57: AI regulatory sandboxes, not to be confused with a company pilot; in Italy it is the Spazio di sperimentazione italiano per l'IA, the national sandbox run by AgID and the ACN (Italy's digital and cybersecurity agencies), with priority for SMEs (Legislative Decree No. 179/2026, Article 26). Read more: How to bring AI into your company: where to start · What an AI project costs: what it depends on · AI adoption in companies: top-down or bottom-up · AI consultant: what they do and how to choose one · AI assistant on company documents: how it works · AI inside the ERP: how to connect an agent.

Generative AI generative artificial intelligence · GenAI

Generative artificial intelligence (generative AI) is the class of models that produce new content, such as text, images, audio, video and code, starting from a request, having learned the structure and characteristics of large amounts of data. The language models behind ChatGPT, Claude, Copilot or Gemini are the most widespread example; the AI Act treats them as general-purpose AI models and requires synthetic content to be recognisable as such. In a company it is used to write, summarise, translate, extract data from documents and answer questions, always with a person who checks the result.

In practice. From the PDF of an order the model extracts items, quantities and delivery dates and fills in the draft in the ERP (the company's management software); a member of staff checks it and confirms it.

Source: NIST AI 600-1 (2024): definition of generative AI; AI Act, Recital 99: generative models as general-purpose AI models; AI Act, Article 50: transparency and marking of generated content. Read more: Company AI policy: usage rules and an example · AI adoption in companies: top-down or bottom-up · ChatGPT, Copilot and Claude: using AI at work.

Hallucination confabulation

A hallucination is a response from a generative artificial intelligence system that contains wrong or invented information, presented with confidence: a regulation, a date, a figure or a citation that does not exist. It is not an occasional fault but a consequence of the way language models produce text, word after word, on the basis of probability. It can be reduced by giving the model the documents to answer from (RAG), by asking it to cite the source and by having a person check what matters, but it cannot be eliminated.

In practice. An assistant asked for the payment term in a contract must answer by citing the paragraph of the contract, not "from memory"; and whoever uses the answer rereads it before sending it to the customer.

Source: NIST AI 600-1, Generative AI Profile (2024), section 2.2 "Confabulation"; European Commission, questions and answers on Article 4 (hallucination as a risk staff should be informed about). Read more: Company AI policy: usage rules and an example · AI literacy: what the AI Act requires of companies · ChatGPT, Copilot and Claude: using AI at work · AI assistant on company documents: how it works.

Hedging (covering price risk) hedge · price hedging

Hedging is the operation by which a company reduces or removes the price risk of an open position: for metals, offsetting the risk of a physical contract with an opposite contract on the exchange or with a financial intermediary, or aligning the pricing formulas of purchases and sales. A company that has to buy metal hedges against a rise (long hedge); one that holds stocks or sells at a floating price hedges against a fall (short hedge). It is the opposite of speculation: a company that hedges also gives up the gains of a favourable move in order to have a known cost or revenue, and decides for itself how much risk to keep. The reference price of the hedge must be the same as that of the physical contract, normally the LME Official Settlement price.

In practice. A company that accepts a fixed-price order and will buy the copper in three months can agree with its intermediary a contract that fixes the price of that copper today: if the market rises, the contract offsets the higher cost; if it falls, the company pays the difference and gives up the saving. Whether, when and how much to hedge is decided by the company with its intermediary, not by software.

Source: Borsa Italiana, financial glossary: Hedging (in Italian); London Metal Exchange, Hedging (Physical market benefits). Read more: Metals hedging: what AI can do · Metal price indexation clauses: a practical guide.

High-risk AI system (AI Act) high-risk AI

Under the AI Act an artificial intelligence system is high-risk in two cases: when it is a safety component of a regulated product, or is itself such a product, subject to third-party conformity assessment (medical devices, lifts, toys; for machinery, after the 2026 Omnibus, the requirements come through the Machinery Regulation); or when it falls within the uses listed in Annex III, including recruitment, the evaluation and monitoring of workers, access to credit, education and critical infrastructure. The provider bears heavy requirements (risk management, data, documentation, logs, human oversight) and the deployer has obligations of its own; an Annex III system is not high-risk if it poses no significant risk to health, safety or rights and does not materially influence the decision, unless it profiles people. After the postponement of July 2026 the obligations apply from 2 December 2027 for the uses in Annex III and from 2 August 2028 for AI embedded in products.

In practice. Reading orders, answering questions about procedures, preparing quotes or checking invoices is not high-risk; using AI to rank CVs, evaluate workers or assign shifts on the basis of personal characteristics is. In Italy the supervisor is the National Cybersecurity Agency (ACN), and whoever uses an Annex III system for decisions about people must be able to explain them to those affected (Legislative Decree No. 179/2026).

Source: AI Act, Article 6: classification rules; AI Act, Annex III: high-risk AI systems; European Commission, timeline for the application of the AI Act; Legislative Decree No. 179 of 7 October 2026, Articles 5, 15 and 23 (in Italian). Read more: AI Act: what a manufacturing company must do · CRM with AI: what it does and what it does not · AI consultant: what they do and how to choose one · AI assistant on company documents: how it works · AI inside the ERP: how to connect an agent.

Human in the loop HITL · human oversight · people in the loop

"Human in the loop" describes a way of using artificial intelligence in which a person stays inside the process: the system prepares, proposes or flags, and a person with the competence and the authority to correct or stop it checks and confirms before the action takes effect. It differs from "human on the loop", where the person supervises and steps in only if needed. It is the design rule of well-built business agents, where records are created as drafts and actions that change data require confirmation; for high-risk AI systems it is an obligation under the AI Act: the provider designs them so that they can be overseen by people (Article 14), and the deployer assigns human oversight to people who are competent, trained and have the authority to intervene (Article 26). In Italy, for decisions on the employment relationship, it is law: the final decision rests with a person who holds effective and autonomous power (Legislative Decree No. 179 of 7 October 2026, Article 41).

In practice. The agent prepares the purchase order starting from the requirement and the day's prices; the buyer opens it, checks it and sends it. No order goes out without that click.

Source: AI Act, Article 14: human oversight; AI Act, Article 26, paragraph 2: the deployer's obligations on human oversight; Legislative Decree No. 179 of 7 October 2026, Article 41: decisions on the employment relationship reserved to a person (in Italian). Read more: AI agents: what they are and what they can do · CRM with AI: what it does and what it does not · AI Act: what a manufacturing company must do · Metals hedging: what AI can do · AI inside the ERP: how to connect an agent.

Idempotency idempotent · idempotence

Idempotency is the property of an operation whereby repeating it several times has the same effect as doing it once. In the connections between an artificial intelligence agent and the ERP (the company's management software), it is the property that prevents duplicates when a request is retried after a network error or an interruption: creating "order number 118" twice must not produce two orders. In the HTTP protocol some operations are idempotent by definition and others are not; for the latter, each request carries a unique key, which the receiving system recognises, together with checks for duplicates in the ERP.

In practice. The agent sends the draft order to the ERP and the connection drops before the reply arrives; the agent tries again with the same key and the ERP, recognising it, does not create a second order. Without this rule, every automatic retry carries a risk of duplicate documents.

Source: IETF, RFC 9110 HTTP Semantics, 9.2.2 Idempotent Methods; IETF, The Idempotency-Key HTTP Header Field (expired Internet-Draft, last revised 15 October 2025). Read more: AI inside the ERP: how to connect an agent.

Inherited permissions security trimming · permission trimming

Inherited permissions are the rule by which an artificial intelligence assistant shows each person only the content that person already has the right to read in the source archive: the index neither widens nor narrows access, which remains that of SharePoint, of the document management system or of the connected application. It is the rule of all the large providers for business assistants, and it applies to the individual person asking the question, whatever the group or role from which they inherit the permission. The practical consequence is that the assistant makes the disorder in permissions visible: a folder shared with the whole company for convenience is found by everyone, even if nobody had ever opened it.

In practice. Before switching on an assistant on its documents, the company checks who can read what in the archive, closes the folders left open out of habit (purchase price lists, customers' drawings, personnel files) and only then connects the sources; projects and notebooks, by contrast, share all their knowledge with whoever opens them, and should be used only for dossiers that all participants are allowed to see.

Source: Microsoft Learn, Data, Privacy, and Security for Microsoft Copilot (only content with read permission); Microsoft Learn, Semantic index for Microsoft Copilot (the index does not change permissions); OpenAI Help Center, Company knowledge in ChatGPT (permissions of the connected sources); Google Cloud, Set up data source access control (Agent Search). Read more: AI assistant on company documents: how it works · ChatGPT, Copilot and Claude: using AI at work · AI inside the ERP: how to connect an agent.

Knowledge base KB · knowledge source

A knowledge base is the set of documents and data that an artificial intelligence assistant is authorised to use to answer: manuals, procedures, quotes, certificates, tables from the ERP (the company's management software), prepared and indexed so that the system can retrieve the relevant passages for each question. It is worth what the documents are worth: if they are old, incomplete or contradictory, the answers will be too. In the providers' products it goes by different names: the "project" in ChatGPT and Claude, the "notebook" in Gemini, the "knowledge source" of an agent in Copilot Studio.

In practice. A company that wants an assistant for its technical department starts from a small, well-kept knowledge base, for example the machine manuals and the quality procedures in force, with one person who keeps it up to date; the questions the assistant cannot answer become the list of documents to add or to write.

Source: AWS, How Amazon Bedrock knowledge bases work; Anthropic, glossary: RAG (the quality of the knowledge base); Microsoft Learn, Knowledge sources summary (Copilot Studio). Read more: AI assistant on company documents: how it works.

Large language model (LLM) language model

A large language model (LLM) is a program trained on enormous amounts of text to predict the most likely continuation of a sequence of words; from this ability come writing, summarising, translating, answering questions and using tools. It is the component at the heart of ChatGPT, Claude, Copilot and Gemini and of business agents, but on its own it is not a complete system: it needs an interface, data and rules. The AI Act classifies it as a general-purpose AI model, with obligations on those who produce it.

In practice. The same model, connected to the CRM, summarises a customer's history; connected to the document archive, it answers questions about procedures; with no connections, it knows only what it read during training, and knows nothing about the company.

Source: Anthropic, glossary: LLM; AI Act, Article 3, point 63: general-purpose AI model. Read more: AI agents: what they are and what they can do · AI in the cloud or on-premise: how to choose · ChatGPT, Copilot and Claude: using AI at work.

LME Official Cash Settlement LME Official Settlement Price · LME cash settlement price

The Official Cash Settlement, or LME Official Settlement Price, is the world reference price for non-ferrous metals: it is the last offer price (the ask, the seller's side) for cash (delivery in two days) quoted in the second Ring session of the morning at the London Metal Exchange, published at midday London time, in dollars per tonne. The LME Cash Settlement price of copper, nickel or aluminium is the number that ends up in physical contracts and invoices, almost always as the average of the Official prices of the delivery month or of the month before, to which the premium is added; it is different from the end-of-day Closing Price, used for margins and valuations. The data is subject to an LME licence: to use it in ERP systems (company management software) and price lists, it has to be taken from an authorised distributor.

In practice. The clause "LME Cash Settlement average of the delivery month plus premium" is calculated by adding up the Official Cash Settlement prices of every trading day of the month and dividing by the number of days; the ERP does this by itself if it receives the series from an authorised source.

Source: London Metal Exchange, LME Official Prices explained; FT Mercati, guide to the LME price. Read more: How to connect LME prices to your ERP · Metal price indexation clauses: a practical guide · Metals hedging: what AI can do.

Metal premium (physical premium) LME premium

The premium is the surcharge, expressed in dollars per tonne, paid on top of the LME price to receive the metal in a given form, in a given place and at a given time: it covers transport from the warehouse, duties, financing, handling, insurance and the availability of the metal in that regional market. The "all-in" price of a supply is therefore the LME price plus the premium, plus the cost of processing; the premium is negotiated separately, normally for a quarter, a half-year or a year, and it can rise even when the exchange price falls. Ingot, billet, slab and wire rod have different premiums, so the form must always be stated in the contract.

In practice. In a contract for aluminium billets delivered in Europe, with the metal already cleared through customs (duty paid), the premium is higher than that of primary ingot and has to be written alongside the quotation period and the exchange rate.

Source: London Metal Exchange, LME Aluminium premiums; FT Mercati, guide to LME premiums. Read more: Metal price indexation clauses: a practical guide · AI agent for purchasing: what it is, what it does · How to connect LME prices to your ERP · Metals hedging: what AI can do.

Metal price indexation clause price adjustment clause · LME price clause

A metal price indexation clause is the part of a supply contract that links the price of a product to an official price of the metal it contains, usually the monthly average of the London Metal Exchange Official Cash Settlement or a market index, according to a written formula: reference, quotation period, premium, exchange rate, rounding and revision rules. It serves to share between supplier and customer the risk of a price that changes every day, and it makes the price determinable even if it is not fixed at signing. The calculation can be entrusted to the ERP (the company's management software), with data from an authorised source and fixed rules, instead of an Excel sheet updated by hand.

In practice. A cable manufacturer invoices the copper content at the previous month's LME average plus the agreed premium, converted into euros at the agreed exchange rate; the ERP recalculates the price list on the first day of each month and produces the document that justifies the adjustment.

Source: Italian Civil Code (in Italian), Articles 1346 (subject matter determined or determinable) and 1474 (price taken from price lists and market bulletins); FT Mercati, guide to LME premiums; FT Mercati, guide to the LME price. Read more: Metal price indexation clauses: a practical guide · How to connect LME prices to your ERP · AI agent for purchasing: what it is, what it does · Metals hedging: what AI can do.

National AI authorities national competent authorities · AI Act market surveillance authority · ACN and AgID (Italy)

Under Article 70 of the AI Act, every EU member state designates its own national authorities: at least one notifying authority and one market surveillance authority. In Italy they are the National Cybersecurity Agency (ACN) and the Agency for Digital Italy (AgID), designated by Law No. 132 of 23 September 2025 and given powers by Legislative Decree No. 179 of 7 October 2026, in force since 23 October 2026. The ACN is the market surveillance authority and single point of contact: it receives complaints, requests information, carries out inspections, orders corrections and imposes the AI Act's fines on providers and deployers; for minor infringements it may simply order a remedy by a set deadline. AgID is the notifying authority: it designates and monitors the bodies that certify high-risk AI systems. Banks, intermediaries and insurers answer to Banca d'Italia, CONSOB and IVASS; the Garante (Italy's data protection authority) remains competent for personal data. The two agencies jointly run Italy's AI regulatory sandbox.

In practice. A company that receives from the ACN a request for information about a chatbot that does not identify itself as one must answer completely and correctly, because inaccurate information given to the authorities is penalised separately; if the infringement is minor, the ACN may simply order it to add the notice by a set deadline.

Source: Legislative Decree No. 179 of 7 October 2026, Articles 4, 5, 6, 15, 18, 22 and 23 (Gazzetta Ufficiale No. 234 of 8 October 2026, in Italian); Law No. 132 of 23 September 2025, Article 20 (Normattiva, in Italian); AI Act, Article 70: national competent authorities and single point of contact. Read more: AI Act: what a manufacturing company must do · AI literacy: what the AI Act requires of companies · Company AI policy: usage rules and an example · ChatGPT, Copilot and Claude: using AI at work.

Net exposure to the metal price metal price exposure · net exposure

Net exposure to the metal price is the balance, by period and in tonnes per metal, between a company's flows that gain and those that lose when the price rises. Exposure arises from stocks, from purchases at a floating price and from sales at a fixed price, and it changes every time an order, a delivery or an invoice comes in: it is the part of the result that moves with the market. Net exposure is the number from which every hedging decision starts, and it is calculated from the data in the ERP (the company's management software), not from memory; how much of that exposure to keep or to hedge is for the company to decide.

In practice. A cable manufacturer with fixed-price orders for the quarter and copper purchases at the LME monthly average has a net exposure to price rises equal to the copper not yet bought: software recalculates it at every order and every delivery and shows how much is already hedged back-to-back and how much is not.

Source: London Metal Exchange, LME Digest: Physical and financial hedging, beginners guide (exposure, stocks, fixed-price sales); London Metal Exchange, Hedging (the company decides how much risk to accept). Read more: Metals hedging: what AI can do.

OCR optical character recognition

OCR (optical character recognition) is the conversion of an image of text, such as a scanned PDF, a photo of a certificate or a completed form, into text that a program can search and an artificial intelligence assistant can read. The current services of the large providers state that they read print, handwriting, tables and document structure, but how well they cope with stamps, forms filled in by hand and photocopies of photocopies has to be checked on the company's real documents. Without OCR, a scanned archive does not exist as far as the assistant is concerned.

In practice. A company that wants its material certificates of the last ten years, all scanned, to be findable tests OCR on a sample of a hundred documents and checks by hand whether the heat number, alloy and test values have been read correctly; only then does it index the archive, keeping a flag on the documents where OCR got it wrong.

Source: Microsoft Learn, What is Azure AI Document Intelligence? (printed and handwritten text, tables, searchable PDFs); Google Cloud, Document AI (OCR from scanned documents). Read more: AI assistant on company documents: how it works · AI inside the ERP: how to connect an agent.

On-premise on-premises · self-hosted · local AI

On-premise, that is, "on site" or local, describes software and artificial intelligence models installed and run on the company's own computers or servers, rather than in a provider's data centres reached over the internet (the cloud). With on-premise AI the data does not leave the company and the system works even without a network connection, but it takes suitable hardware, maintenance and open-weight models, which are generally smaller than those available in the cloud. Many companies use both: on-premise for what is confidential, the cloud for the tasks that require the most capable models.

In practice. The engineering department, whose drawings are covered by trade secrets, uses a local model to search and summarise; sales uses a cloud model, through the company API, for draft quotes.

Source: NIST, Special Publication 800-145: definition of cloud computing (on premises / off premises). Read more: AI in the cloud or on-premise: how to choose · ChatGPT, Copilot and Claude: using AI at work · AI assistant on company documents: how it works.

Open-weight and open-source models open weights · open-source model

An open-weight model is an artificial intelligence model whose developer publishes the file containing its parameters, the "weights", so that anyone can download it and run it on their own computers, even without a connection. It is not necessarily open source: strictly speaking, it is so only if the licence allows it to be used, studied, modified and redistributed for any purpose and if, according to the Open Source Initiative's definition, the code and the information on the training data are also available. The AI Act reserves lighter obligations for general-purpose AI models released under a free and open-source licence, with weights, architecture and information on usage made public, except those with systemic risk.

In practice. A company with data that must not leave its premises installs an open-weight model on an internal server and uses it to search and summarise documents; the most capable models of the large providers are not open-weight and can be used only in the cloud.

Source: AI Act, Article 53, paragraph 2 and Recital 102; Open Source Initiative, The Open Source AI Definition 1.0. Read more: AI in the cloud or on-premise: how to choose · AI assistant on company documents: how it works.

Prompt (request to a model) system prompt

A prompt is the request, in natural language, given to a generative artificial intelligence model: the instructions, the context, the examples and the documents it has to start from. The quality of the answer depends largely on the quality of the prompt (role, objective, expected format, limits, what to do if a piece of data is missing). In business applications the system prompt, written once by whoever builds the software, sets rules and tone, and the user adds only the request of the moment; it should not be confused with the Windows "Command Prompt". "Prompt injection" is the attempt to manipulate the instructions through the data the model reads.

In practice. "You are the assistant of the quotes office. From the attached PDF, extract items, quantities and delivery terms into a table; if a piece of data is missing write 'missing', do not make it up."

Source: OpenAI, prompt engineering (developer documentation); Anthropic, prompt engineering overview; NIST AI 600-1 (2024): prompt injection. Read more: AI agents: what they are and what they can do · Company AI policy: usage rules and an example.

Prompt date prompt · LME prompt date

The prompt date is the delivery and settlement date of a London Metal Exchange contract. Unlike other exchanges, the LME quotes daily prompt dates out to three months, weekly ones out to six and monthly ones beyond, normally the third Wednesday of the month: in this way a company can match its hedge to the day of physical delivery, instead of adapting to a few standard dates. For holidays, the LME publishes a notice giving the replacement date; the last moment to trade a contract is the trading day before the prompt date, by 12.30 London time.

In practice. The calendar of prompt dates and of the contracts' quotation periods is one of the sources of the automatic alerts that an ERP (the company's management software) can give to purchasing and administration: a hedge falling due, the last day to settle it, the end of a monthly averaging window.

Source: London Metal Exchange, Prompt date structure; London Metal Exchange, Contract types. Read more: Metals hedging: what AI can do.

Punto Impresa Digitale (PID)

The Punto Impresa Digitale (PID) is an Italian service: the digital support desk of Italy's Chambers of Commerce, launched in 2017 to support the 4.0 digitalisation of businesses, which helps a business understand its level of digitalisation and choose its next steps. It offers an online self-assessment of digital maturity (SELFI4.0), with an automatic report at the end of the questionnaire, and an assessment guided by a Digital Promoter from the Chamber (ZOOM4.0), with interviews with management and staff and a more detailed report. Since July 2026 the PIDs have also run a national funding call with local application windows, with vouchers for technology, training and consultancy and a specific focus on the adoption of artificial intelligence, and local laboratories called AI-Lab.

In practice. A company in Italy that wants to understand where to start completes SELFI4.0 and reads the report; for a more in-depth reading it asks its Chamber of Commerce for ZOOM4.0 with the Digital Promoter. The tools cover digitalisation in general, not artificial intelligence in particular, but they give a free snapshot before paying a consultant. Dates, supplier requirements and the size of the grant are decided by each Chamber in its own local call.

Source: Punto Impresa Digitale, assessment of businesses' digital maturity (SELFI4.0 and ZOOM4.0, in Italian); Punto Impresa Digitale, Bando Voucher Doppia Transizione 2026 (the twin transition voucher call, in Italian); Unioncamere (the union of Italy's Chambers of Commerce), Punto Impresa Digitale: the Chambers of Commerce project (17 October 2017, in Italian). Read more: AI consultant: what they do and how to choose one.

RAG (retrieval-augmented generation)

RAG (retrieval-augmented generation) is the technique by which an artificial intelligence system, before answering, searches the authorised documents and data for the passages relevant to the question and passes them to the language model together with the request; the model answers on the basis of those passages and can cite them. It makes it possible to use up-to-date and confidential information without training the model, reduces made-up answers and makes it possible to check where every statement comes from. It is the basis of AI assistants on company know-how.

In practice. Asked "what is the thickness tolerance in the customer's specification?", the system retrieves the page of the specification currently in force and answers by citing it, with the link to the document.

Source: Anthropic, glossary: RAG (retrieval augmented generation). Read more: AI agents: what they are and what they can do · AI assistant on company documents: how it works.

Risk levels (AI Act) risk-based approach · risk categories

The AI Act regulates artificial intelligence systems according to risk, on four levels. Prohibited AI practices (Article 5), such as inferring people's emotions in the workplace, have been banned since 2 February 2025. High-risk AI systems (Article 6: safety components of the regulated products in Annex I and the uses in Annex III) carry heavy requirements for the provider and obligations for the deployer, which apply from 2 December 2027 for Annex III and from 2 August 2028 for products. Some systems have only transparency obligations (Article 50, since 2 August 2026), such as chatbots, which must identify themselves as such. Everything else, that is, the great majority of office uses, has no specific obligations beyond the general ones, such as AI literacy among staff; the rules for general-purpose AI models stand apart, and fall on those who produce them.

In practice. Reading orders, preparing quotes or answering questions about procedures carries no specific obligations; a customer-facing chatbot must say that it is an AI; ranking CVs with AI is high-risk; measuring employees' stress from their chats is prohibited.

Source: AI Act, Article 5: prohibited AI practices; AI Act, Article 6: classification of high-risk AI systems; AI Act, Article 50: transparency obligations; Legislative Decree No. 179 of 7 October 2026, Article 23: the fines applied in Italy for each level (in Italian). Read more: AI Act: what a manufacturing company must do · AI literacy: what the AI Act requires of companies · ChatGPT, Copilot and Claude: using AI at work.

RPA (robotic process automation) software robots

RPA (robotic process automation) is automation carried out by "software robots" that repeat on screen the actions of a person: opening a program, copying a field, pressing a button. It is used for repetitive, rule-based tasks, even on old programs without an API; it breaks when the screen changes and it does not understand new cases. It is not artificial intelligence unless it contains a model: ISTAT, Italy's national statistics institute, excludes general automation systems without artificial intelligence components from AI. UiPath, one of the main RPA providers, now presents it as the layer that carries out what an AI agent has decided.

In practice. The robot enters into the old ERP (the company's management software), line by line, the data that the agent has extracted from PDF orders and that a person has confirmed: the agent understands the document, the robot presses the keys. If the ERP provider updates the screens, the robot has to be rebuilt.

Source: UiPath, What is Robotic Process Automation; Microsoft Learn, Introduction to desktop flows (Power Automate); ISTAT, Imprese e ICT 2025 report (in Italian): definition of artificial intelligence in the report's glossary. Read more: AI inside the ERP: how to connect an agent.

Shadow AI ungoverned AI use · unsanctioned AI

Shadow AI is the use of artificial intelligence tools by employees and departments without the company knowing about it or governing it: personal accounts for ChatGPT or similar services, used to translate, write, summarise or analyse work data. It is spontaneous bottom-up adoption, a sign of real demand, but it exposes the company to risks concerning confidential data, unchecked results and unmet obligations, because the AI Act also applies to those who use only general-purpose tools. The answer is not a ban, which makes it invisible, but company accounts, written usage rules and training.

In practice. The salesperson who pastes a competitor's quote and a customer's records into a personal chat to "tidy up the text" is doing shadow AI, even if the aim is to work better.

Source: European Commission, questions and answers on Article 4: employees' use of ChatGPT falls within the obligation. Read more: AI adoption in companies: top-down or bottom-up · Company AI policy: usage rules and an example · ChatGPT, Copilot and Claude: using AI at work.

Single sign-on (SSO) SAML

Single sign-on (SSO) is access with the company identity: the employee signs in to a service, for example an artificial intelligence assistant, with the same work account they use for email and documents (Microsoft Entra, Google or another identity provider), without a separate password. In this way the company creates, suspends and removes access from a single place, and with the SCIM protocol it does so automatically when a person joins or leaves. It is the first setting that distinguishes business use of AI tools from a collection of personal accounts: ChatGPT Business includes it, Claude Enterprise adds it together with automatic user management, and Microsoft Copilot inherits it from the work account.

In practice. When a person leaves the company, with single sign-on their access to the AI assistant closes at the same moment as their email; with personal accounts, by contrast, the conversations and the uploaded documents stay in their hands. It is one of the reasons why the usage rules allow only company accounts.

Source: OpenAI Help Center, Setting up single sign-on (SSO) for ChatGPT Business; Anthropic Help Center, Team and Enterprise plans: identity management (SSO, JIT, SCIM); Microsoft Learn, Microsoft 365 Copilot overview (Microsoft Entra account, AI administrator role). Read more: ChatGPT, Copilot and Claude: using AI at work.

Source citation citation · source attribution

Source citation is the indication, next to each statement made by an artificial intelligence assistant, of the document and the place (page, paragraph) the information was taken from, so that the reader can open it and check. It is what distinguishes an assistant on company documents from a general chat, and it is the first check against invented answers: providers treat it as part of the technique, with the page number for PDFs and, in the more advanced systems, with a grounding score sentence by sentence. An answer without a source is not used to make decisions.

In practice. "The payment term granted to that customer is 60 days (quote 2024-117, page 3)": the reader opens the quote and checks. If the same question receives an answer with no reference, or with a reference that does not contain the figure, the answer is discarded and the case noted for the person responsible for the sources.

Source: AWS, What is Retrieval-Augmented Generation? (source attribution); Anthropic, Citations (citations with page number, text only); Google Cloud, Check grounding with RAG (sentence-by-sentence grounding score); Microsoft Learn, Data, Privacy, and Security for Microsoft Copilot (citations stored with the answers). Read more: AI assistant on company documents: how it works.

Tenant Microsoft 365 tenant · cloud organisation

A tenant is an organisation's reserved space inside a shared cloud service: users, permissions, documents and settings sit inside that boundary, logically separated from those of the other companies using the same service. In Microsoft 365 the tenant is also the perimeter within which Copilot and its logs operate: the assistant reads only the tenant content for which the person already has permissions, and the location where the tenant is registered decides whether the data stays within the EU Data Boundary, the European boundary within which Microsoft commits to storing and processing customers' data.

In practice. A company with its email on Microsoft 365 already has a tenant: switching on Copilot with the work account means having it work inside that perimeter, under the same contract as email and files. Before switching it on, it is worth checking who can read what in SharePoint, because Copilot shows everything the person can already open, including folders shared by mistake.

Source: Microsoft Learn, Data, Privacy, and Security for Microsoft Copilot (logical isolation of tenants, permissions); Microsoft Learn, What is the EU Data Boundary?. Read more: ChatGPT, Copilot and Claude: using AI at work · AI assistant on company documents: how it works.

Token

A token is the smallest unit of text a language model reads and writes: a fragment of a word, a short word, a punctuation mark. A page of text is worth roughly a thousand tokens, with differences depending on the model and the language (Italian takes more than English). Tokens serve two practical purposes: measuring how much text fits into the context window, and calculating the consumption of cloud models, which are paid for by tokens read and written, at different rates depending on the size of the model.

In practice. A thirty-page contract sent to the model "costs" about thirty thousand tokens to read, plus the tokens of the answer; the quotation for a project estimates these volumes and the pilot verifies them.

Source: Anthropic, glossary: tokens. Read more: What an AI project costs: what it depends on · AI in the cloud or on-premise: how to choose · AI assistant on company documents: how it works.

Training data

Training data are the texts, images and examples from which an artificial intelligence model learns its parameters; the AI Act defines them as data "used for training an AI system through fitting its learnable parameters". For a language model they amount to billions of pages and, in the personal versions of many assistants, they may also include users' conversations: in ChatGPT until the setting that improves the model for everyone is switched off, in Claude according to the choice made by the user, in Gemini until training is switched off in the History settings, in Copilot with a personal account unless the user opts out in the settings. In the plans for organisations, the customer's content does not become training data by default, and the contract says so.

In practice. There is only one question to ask every provider: do our requests and our documents go into the training data? In the plans for organisations the answer is no by default; in the personal ones it depends on a setting that should be checked on the first day. A price list pasted into an account that trains the model is information that can never be taken back.

Source: AI Act, Article 3, point 29: definition of training data; OpenAI Help Center, How your data is used to improve model performance; Anthropic, Updates to Consumer Terms and Privacy Policy (28 August 2025); Google, Gemini Apps Privacy Hub; Microsoft, Privacy FAQ for Microsoft Copilot. Read more: ChatGPT, Copilot and Claude: using AI at work.

Webhook web hook · HTTP callback

A webhook is the notification that one system automatically sends to another when something happens: an order confirmed, an invoice posted, a customer or supplier record changed. Unlike an API, where it is the external program that asks, here it is the ERP (the company's management software) that calls; the program that receives the notification has to subscribe, prove its identity with a shared secret and renew the subscription before it expires. For an artificial intelligence agent it is the way to react to an event at the moment it happens, instead of querying the ERP continuously.

In practice. When the customer confirms the order, the ERP notifies the agent with a webhook; the agent prepares the draft order confirmation with the dates from the production plan and leaves it with the salesperson. If the webhook subscription expires and nobody renews it, the notifications are lost silently: it is the first check to put in the log.

Source: Microsoft Learn, Working with webhooks in Dynamics 365 Business Central. Read more: AI inside the ERP: how to connect an agent.

A missing term, or a definition to discuss?

Write to us: the entries come from the questions companies ask us. And if you want to understand how these tools fit into your processes, let's talk.