Glossary

Plain-language AI glossary.

50 terms every European mid-market leader runs into. Plain definitions, why each matters, and how each relates to the arqmetrica AI Maturity Index.

53 / 53

AI Act

EU regulation (Regulation (EU) 2024/1689) that establishes risk-based rules for AI systems placed on the EU market. It is the first comprehensive horizontal AI law in any major jurisdiction.

Why it matters.High-risk obligations become enforceable from August 2026 — every European mid-market company offering AI-touched products or services to EU customers is in scope.

Related dimension:Governance & ethics

AI Act conformity assessment

The pre-market evaluation that proves a high-risk AI system meets the AI Act requirements on data quality, documentation, transparency, human oversight, robustness and cybersecurity.

Why it matters.No high-risk AI system can be sold or used in the EU without a passed assessment and a CE marking — analogous to the safety regime for medical devices.

Related dimension:Governance & ethics

AI agent

An AI system that can plan and execute multi-step tasks autonomously, using tools (APIs, code, web browsing) to take actions in the world rather than just produce text.

Why it matters.Agents move AI from advice to action. They also raise the stakes on oversight, audit trails and reversibility, because their mistakes have real consequences.

Related dimension:Tooling & infrastructure

AI centre of excellence

A small, central team that sets AI standards, owns shared infrastructure, builds reusable patterns and supports business units delivering AI projects — without owning every project itself.

Why it matters.A working centre of excellence is the difference between every business unit reinventing the wheel and the company compounding learning across projects.

Related dimension:People & capability

AI change management

The structured set of activities — communications, training, role redesign, incentive change — that move an organisation from announcing AI to actually using it.

Why it matters.The technology is rarely the bottleneck. The bottleneck is whether middle managers understand how their team’s work changes, and whether they are rewarded for changing it.

Related dimension:People & capability

AI coach

A specialist who works one-to-one or in small groups to upskill employees on practical AI use within their actual workflows — not generic training.

Why it matters.Coaching closes the gap between AI tools your company has paid for and AI tools your employees actually use. The licences are usually the cheap part.

Related dimension:People & capability

AI incident reporting

The duty under AI Act Article 73 to notify market surveillance authorities of serious incidents and malfunctions involving high-risk AI systems.

Why it matters.You need an internal incident-detection process and a 15-day reporting clock before deployment — not after the first incident lands.

Related dimension:Governance & ethics

AI investment prioritisation

A documented framework that ranks candidate AI initiatives against explicit ROI thresholds, risk gates and capability fit — not whoever shouts loudest.

Why it matters.Without a prioritisation framework, AI budget tends to be allocated by vendor sales cycles. The result is a portfolio of pilots that nobody asked for.

Related dimension:Strategy & vision

AI literacy

The practical understanding employees need to use AI tools well and recognise their limits — covering how the tools work, what they get wrong, and when not to use them.

Why it matters.Mandated by AI Act Article 4 since February 2025. Also the cheapest way to lift productivity: a workforce that knows how to prompt is a workforce that ships faster.

Related dimension:People & capability

AI literacy mandate

AI Act Article 4 obligation that providers and deployers ensure a sufficient level of AI literacy among their staff and any other person involved in the operation of AI systems on their behalf.

Why it matters.In force since February 2025. The duty is on the company, not the staff — you need a documented AI literacy programme that is auditable.

Related dimension:Governance & ethics

The structural choice of how AI capability is organised: a central team, embedded specialists in business units, a federated hub-and-spoke, or a centre of excellence supporting decentralised delivery.

Why it matters.The wrong operating model bottlenecks delivery for years. Most mid-market companies should start federated and centralise specific capabilities, not the other way round.

Related dimension:Strategy & vision

AI pilot purgatory

The state in which AI pilots produce demos but never reach production. Symptoms: every quarter starts with new pilots and ends with no measurable business outcomes.

Why it matters.MIT Sloan/BCG 2024 research finds the median European mid-market company gets only 22% of pilots into production — the rest stall.

Related dimension:ROI & measurement

AI ROI scorecard

A structured tracker that records, for each AI initiative, the business outcome it targets, the baseline, the measured result, and the net financial impact.

Why it matters.No scorecard, no reckoning. Companies without one almost always overestimate AI value because they remember the wins and forget the pilots that died.

Related dimension:ROI & measurement

AI strategy

A board-endorsed document that says where AI will create value for the business in the next 18 to 36 months, what will not be pursued, who owns delivery, and how much will be invested.

Why it matters.In MIT Sloan/BCG longitudinal research, the presence of a clear AI strategy is the single strongest predictor of measurable AI value capture.

Related dimension:Strategy & vision

Attribution method

The technique used to isolate the share of a business outcome that an AI initiative actually caused — separating it from market effects, seasonality and other simultaneous changes.

Why it matters.Without a defensible attribution method, every claimed AI ROI is just a story. With one, you can defend numbers in front of a CFO.

Related dimension:ROI & measurement

Bias audit

A structured test of an AI system to detect whether outputs differ systematically across protected groups — for example, by gender, age, ethnicity or nationality.

Why it matters.Required for high-risk AI systems under the AI Act, and increasingly demanded by enterprise procurement. Most off-the-shelf vendor models have not had one.

Related dimension:Data foundations

Chief AI Officer (CAIO)

The named executive accountable for AI outcomes across an organisation — strategy, governance, capability building, vendor selection, and ROI reporting to the board.

Why it matters.Without a single accountable owner, AI work scatters across IT, operations and individual business units, and nobody is on the hook for the result.

Related dimension:Strategy & vision

Context window

The maximum amount of text — measured in tokens — that a language model can consider at once when generating a response. Frontier models in 2026 typically support 200,000 to 2,000,000 tokens.

Why it matters.Context window limits decide which use cases are feasible. Long-document analysis, long-running agents and multi-file code review all live or die on the available window.

Data governance

The organisational framework — roles, policies and controls — that defines who owns each data domain, how data quality is enforced, and how access is granted and revoked.

Why it matters.AI Act Article 10 makes data governance a legal requirement for high-risk systems. It is also the single biggest predictor of whether AI projects reach production.

Related dimension:Data foundations

Data lineage

The end-to-end record of where a piece of data came from, how it was transformed, and where it is used. Sometimes called provenance.

Why it matters.Without lineage you cannot prove training data quality, you cannot answer regulator questions about bias, and you cannot delete personal data on request.

Related dimension:Data foundations

Data residency

The geographic location where data is stored and processed. EU data residency means all storage and compute happens inside the European Union.

Why it matters.Essential for GDPR confidence and post-Schrems II compliance. Many enterprise procurement teams now treat non-EU residency as a hard disqualifier.

Related dimension:Data foundations

DPIA (Data Protection Impact Assessment)

A structured risk assessment required by GDPR Article 35 before processing personal data in ways likely to result in a high risk to individuals — including most AI use cases.

Why it matters.A DPIA is the document a data protection authority will ask for first. No DPIA, and you have no defence when a complaint lands.

Related dimension:Data foundations

Embeddings

Numerical representations of text, images or other content as vectors of numbers, so that semantically similar items end up close together in the vector space.

Why it matters.Embeddings are the foundation of semantic search, recommendation, classification and RAG. Almost every useful enterprise AI pipeline produces them at some stage.

EU AI Act risk classification

The four-tier system the AI Act uses to sort AI systems by societal risk: prohibited, high-risk, limited-risk and minimal-risk. Each tier triggers a different set of legal obligations.

Why it matters.Your obligations — and your compliance budget — are decided entirely by which tier your AI use cases fall into. Misclassifying is the most expensive mistake you can make.

Related dimension:Governance & ethics

FAIR data principles

A widely adopted set of guidelines saying data should be Findable, Accessible, Interoperable and Reusable. Originated in 2016 in the open-science community.

Why it matters.FAIR is the closest thing to a universal benchmark for data foundations. If you cannot tick all four letters for the data your AI uses, that is your bottleneck.

Related dimension:Data foundations

Fine-tuning

The process of taking an already-trained AI model and continuing training on a smaller, specific dataset so it performs better on a particular task or in a particular style.

Why it matters.Fine-tuning is often pitched as the answer when retrieval (RAG) or better prompting would be cheaper, faster and easier to govern. Pick the right tool, not the impressive one.

Fractional CAIO

A senior AI executive engaged part-time — typically two to four days per month — instead of as a full-time hire. Common for mid-market companies that need executive-grade AI leadership without a 250k salary.

Why it matters.For most European mid-market companies, the AI workload does not justify a full-time CAIO yet — but the absence of one is the bottleneck.

Related dimension:Strategy & vision

General-purpose AI (GPAI) model

An AI model trained on broad data that can perform a wide range of tasks and be integrated into many downstream systems — for example, the foundation models behind ChatGPT, Claude and Gemini.

Why it matters.GPAI providers carry their own AI Act obligations, including model documentation and copyright transparency. If you fine-tune or deploy one, parts of those obligations cascade onto you.

Related dimension:Governance & ethics

Hallucination

When a language model produces output that is fluent, confident and plausible-sounding — but factually wrong. Sometimes called confabulation.

Why it matters.Hallucination is the headline AI risk for any deployment that touches customers, regulators or accountants. Mitigation (grounding, citations, human review) is a design choice, not an afterthought.

High-risk AI system

An AI system listed in AI Act Annex III or used as a safety component of a regulated product — for example, AI used in hiring, credit scoring, education, critical infrastructure or law enforcement.

Why it matters.High-risk systems require conformity assessment, technical documentation, human oversight and post-market monitoring before they can be placed on the EU market.

Related dimension:Governance & ethics

Holdout test

An experimental design where part of the population is deliberately excluded from an AI intervention so the difference in outcomes between treatment and holdout can be measured.

Why it matters.A holdout is the gold standard for proving AI ROI. It is also the test most companies refuse to run because they are afraid of what it will show.

Related dimension:ROI & measurement

Human oversight

The requirement that high-risk AI systems are designed and used so that natural persons can effectively monitor them, intervene, and override outputs — defined in AI Act Article 14.

Why it matters.Oversight is a design requirement, not a slogan. The system must surface its own confidence, expose review controls, and allow the human reviewer to understand what the model is doing.

Related dimension:Governance & ethics

Inference

The act of running a trained AI model to produce an output — as opposed to training, which is the process of building the model in the first place.

Why it matters.Inference is where AI cost is incurred at scale. Inference cost per query, multiplied by query volume, is the unit economics question every CFO will eventually ask.

Kill criteria

Pre-agreed conditions under which an AI initiative will be stopped — defined before it starts, so the decision is automatic rather than political.

Why it matters.Without kill criteria, failing initiatives keep getting funded because nobody wants to admit they were wrong. With them, capital recycles into what works.

Related dimension:ROI & measurement

LLM (large language model)

An AI model trained on vast text corpora to predict the next token in a sequence — the technology underneath ChatGPT, Claude, Gemini and similar assistants.

Why it matters.LLMs are the general-purpose AI substrate of the next decade. Understanding what they do well — and badly — is now baseline executive literacy.

MCP server (Model Context Protocol)

A standard introduced by Anthropic in late 2024 for connecting AI models to external tools and data sources. An MCP server exposes tools and data that any compatible AI client can use.

Why it matters.MCP is becoming the universal connector between AI assistants and enterprise systems — the equivalent role that USB plays for hardware.

Related dimension:Tooling & infrastructure

MLOps

The engineering discipline that brings DevOps practices to machine learning — versioning models and data, automating deployment, monitoring drift, and managing retraining.

Why it matters.Without MLOps, every AI model in production is a one-off science project. With it, you can ship and operate dozens.

Related dimension:Tooling & infrastructure

Model card

A short, structured document that describes an AI model — its intended use, training data, performance metrics, known limitations, and ethical considerations.

Why it matters.AI Act Article 11 effectively makes model cards mandatory for high-risk systems. Even outside that, they are the cheapest tool for vendor due diligence.

Related dimension:Tooling & infrastructure

Model drift

The gradual decay in a model’s performance as the real-world data it sees diverges from the data it was trained on. Comes in two flavours: data drift (the inputs change) and concept drift (the underlying relationship changes).

Why it matters.Drift is silent. A model can degrade for months before anyone notices, particularly if its outputs feed downstream automated decisions.

Related dimension:Tooling & infrastructure

Multimodal AI

AI systems that can process and generate more than one type of content — for example, text plus images, or text plus audio plus video.

Why it matters.Multimodal models open use cases that text-only AI cannot touch: document understanding with charts and tables, visual quality control, voice-first interfaces, and accessibility.

PII (personally identifiable information)

Any data that can identify a natural person — directly (name, email, ID number) or indirectly in combination with other data (postcode plus date of birth plus job title).

Why it matters.The threshold for PII under GDPR is much lower than most teams assume. Internal AI tools that look anonymised often are not.

Related dimension:Data foundations

Prohibited AI practices

AI uses banned outright by AI Act Article 5, including social scoring, untargeted facial-image scraping, emotion recognition in workplaces and schools, and most real-time remote biometric identification in public spaces.

Why it matters.These prohibitions have been in force since February 2025 and carry the largest AI Act fines — up to 7% of global annual turnover.

Related dimension:Governance & ethics

Prompt engineering

The discipline of designing, testing and refining the instructions given to a language model to produce reliable, accurate and useful outputs.

Why it matters.A well-engineered prompt frequently outperforms a fine-tuned model at one tenth of the cost — which is why prompt engineering is the highest-leverage skill to build internally.

Related dimension:People & capability

RAG (retrieval-augmented generation)

A pattern where a language model answers a question by first retrieving relevant documents from a knowledge base and then generating an answer grounded in them.

Why it matters.RAG is how most enterprise AI assistants stay accurate and citeable. It is also the cheapest alternative to fine-tuning when you need the model to know your data.

Related dimension:Tooling & infrastructure

Responsible AI

The umbrella term for designing, building and operating AI systems that are fair, transparent, accountable, safe and aligned with human values — operationalised through frameworks such as NIST AI RMF and ISO/IEC 42001.

Why it matters.Responsible AI is increasingly a procurement requirement: large enterprise and public-sector buyers expect documented practices, not slogans, before they sign.

Synthetic data

Artificial data generated by a model to mimic the statistical structure of real data, without containing actual records of real people.

Why it matters.A practical way to train and test AI systems without exposing real customer data — but only if generated correctly. Poor synthetic data leaks the originals.

Related dimension:Data foundations

Temperature

A setting on a language model that controls how random its outputs are. Low temperature (near 0) produces deterministic, conservative outputs; high temperature produces more varied, creative outputs.

Why it matters.Most enterprise use cases want low temperature for predictability and audit. Default settings are often higher than they should be for compliance-sensitive work.

Tokenisation

The process of breaking text into the smaller units (tokens) that a language model actually processes. Roughly, one English word equals 1.3 tokens; one Portuguese word slightly more.

Why it matters.Tokens are the unit of pricing for almost every AI vendor. Understanding tokenisation lets you forecast costs and explain why Portuguese, French and German workloads are systematically more expensive than English ones.

Transformer

The neural network architecture, introduced by Google in 2017, that powers nearly all modern large language models. Its key innovation is the attention mechanism that lets the model weigh the relevance of every word to every other word.

Why it matters.Almost every commercially significant AI advance since 2018 — from GPT to AlphaFold — sits on top of the transformer.

Transparency obligations

AI Act Article 50 duties to disclose AI involvement: users must be told when they are interacting with an AI system, and AI-generated or manipulated content (deepfakes, synthetic media) must be labelled as such.

Why it matters.Even minimal-risk systems carry transparency duties. Customer-facing chatbots, AI voice agents and AI-generated marketing imagery all need clear disclosure.

Related dimension:Governance & ethics

Value realisation

The discipline of tracking an AI initiative beyond go-live, all the way to the financial outcome it was meant to deliver — typically over the following four to eight quarters.

Why it matters.Most AI value is captured (or lost) in the months after deployment, not at launch. Companies that stop tracking at go-live capture only a fraction of the available upside.

Related dimension:ROI & measurement

Vector database

A database optimised for storing and searching high-dimensional embeddings — the numerical representations of text, images or other content that AI models work with.

Why it matters.Vector databases are the storage layer behind almost every retrieval-augmented (RAG) application. Choice of vector database constrains scale, cost and EU residency.

Related dimension:Tooling & infrastructure

Vendor lock-in

The state of being tied to a single AI vendor because switching costs — re-training, re-integration, re-procurement — are prohibitive.

Why it matters.AI vendor lock-in compounds faster than traditional SaaS lock-in because data, prompts, fine-tuning and agent state all become vendor-specific.

Related dimension:Tooling & infrastructure