— GLOSSARY

The AI terms to know before you decide.

Terminology from a board's perspective. Primary sources, Polish context, no textbook generalities.

— SEARCH THE GLOSSARY

29 entries

Agent washing (rebranding chatbots as AI agents)

The practice of rebranding classic chatbots, workflow automation, or RPA as agentic AI. The vendor adds a GPT layer to an existing solution and sells it for three times the price. Out of the thousands of vendors claiming agentic capabilities in 2025, Gartner estimates that only around 130 actually have genuine agentic functions.

AI Act (the EU regulation governing artificial intelligence)

Regulation (EU) 2024/1689 of the European Parliament of 13 June 2024, governing artificial intelligence systems in the European Union. It applies in stages: a ban on prohibited practices from February 2025, requirements for general-purpose models from August 2025, full requirements for high-risk systems from August 2026, and full application from August 2027.

AI governance (oversight and decision-making for AI in a company)

A set of rules, roles, and procedures that define who in the company decides on AI deployments, who monitors risks, and who responds to incidents. Without AI governance, projects escalate on their own into problems no one knows how to solve. Firms with documented AI governance see three times higher ROI from their deployments.

AI Operating Model (a company's operating model for AI)

The way an organisation decides, builds, deploys, and maintains AI solutions at scale. Four main models: a central AI unit (Center of Excellence), a federated distributed model, a hybrid model, and AI embedded in business units. The choice of model determines 60 to 80 percent of the long-term ROI from AI in a company.

AI Readiness Score (a measure of a company's readiness to deploy AI)

A measurable indicator of an organisation's readiness to deploy AI, expressed on a 0 to 100 scale, calculated in the proprietary yesfor.ai methodology from eight component areas. It shows where the company is ready and where it has blockers, without interpretation. A score below 50 means that deploying AI in the organisation's current state has a high probability of failure.

AI-ready data (data ready for an AI deployment)

Data that meets Gartner's five technical conditions, making it possible to train models and keep them running in production. Alignment with a specific use case, clear ownership of the asset, automated pipelines with quality gates, maintained metadata, and continuous quality control. Without all five conditions, AI learns chaos and replicates it at scale.

Bias in AI (algorithmic bias)

Systematic deviations in an AI model's predictions that harm specific groups of people or cases. Bias can come from training data (historical inequalities), from model design, from the data labelling process, or from prompt engineering. AI models without a bias audit categorically fall under the documentation obligations of the AI Act from August 2026.

Capability transfer (handing competence from the consultant to the company)

A deliberate process in which the consultant or vendor builds the knowledge and the ability to maintain the solution into the client's organisation, instead of creating dependency. Without capability transfer, an AI deployment becomes an operational time bomb: the cost of maintenance reaches thirty to fifty percent of the deployment cost per year, and every change requires re-engaging the consultant.

Context window (a model's context window)

The maximum number of tokens (fragments of vocabulary) that an LLM can take in and process at once in a single query. The context window covers both the user's prompt and the model's response. As of 2026: GPT-4o 128,000 tokens, Claude 3.7 Sonnet 200,000, Gemini 2.5 Pro 2,000,000.

Embeddings (vector embeddings)

A representation of text, an image, or another artefact as a numerical vector in a high-dimensional space, where geometric closeness corresponds to semantic closeness. Embeddings are the foundation of modern semantic search, recommendation systems, and RAG architecture. The dimension of a typical embedding: 768 to 3,072 numbers.

Explainable AI / XAI (the explainability of AI model decisions)

A set of techniques that make it possible to understand why an AI model made a specific decision. Required by the AI Act for high-risk systems, by the financial sector through the KNF Recommendation, and by the medical sector through the MDR. Without XAI, a model is a black box that is not acceptable in regulated contexts.

Fine-tuning (tuning an AI model to a specific task)

The process of adapting a pre-trained LLM to a specific task by further training it on dedicated data. Fine-tuning costs 5 to 50 times more than prompt engineering and 2 to 10 times more than RAG, but it delivers better results in three narrow classes of use case: replicating a style, narrowing scope, and reducing inference costs at high volume.

Foundation model (a base AI model)

A large machine learning model trained on a vast, diverse dataset, designed as a universal starting point for adaptation to specific tasks. The term was introduced by Stanford CRFM in 2021 as a counter to narrow specialist models. Current foundation models include GPT, Claude, Gemini, Llama, and Mistral.

GenAI Divide (the gap in generative AI deployments)

A concept introduced by MIT Project NANDA in a July 2025 report describing the split between the five percent of firms that achieve a measurable return from generative AI deployments and the ninety-five percent whose pilots end with no impact on the P&L. After spending thirty to forty billion dollars.

Human-in-the-loop (a person who approves AI decisions)

An architectural pattern in which an AI system's decision requires human approval before its consequences take effect. It is not a safety net, it is part of the architecture, and without it a solution is not production-ready in high-stakes applications. The share of deployments using HITL in the financial and medical sectors exceeds 95 percent.

Last mile problem (the final stretch of AI deployment)

A phenomenon in which a demo and a pilot show promising results, but integrating the model with the organisation's actual workflow turns out to be ten to twenty times more expensive than the model itself. The final stage, where AI plugs into the real processes of people and systems, absorbs 70 to 80 percent of the deployment budget. Most AI projects never reach that stage.

LLM hallucinations

A language model producing answers that are formally correct in syntax but factually untrue. A hallucination is not an implementation bug or a defect, it is an architectural trait of probabilistic models. The hallucination rate in production enterprise applications ranges from 3 to 27 percent depending on the type of query and the benchmark.

MLOps (DevOps for machine learning)

A set of practices that automate the lifecycle of a machine learning model from research experiment to production and maintenance. It covers versioning of models and data, automated training pipelines, deployment, monitoring and retraining. An organisation's MLOps maturity is the single strongest predictor of sustained AI ROI in the second and third year after deployment.

Model drift (degradation of an AI model over time)

The decline in an AI model's prediction quality in production, caused by a change in the input data (data drift) or a change in the relationship between the data and the outcome (concept drift). Without monitoring and retraining a model typically degrades by 5 to 20 percent in quality per year. Model drift is one of the three most common causes of AI deployment failure six months after go-live.

Multi-tenant SaaS (one SaaS platform serving many customers)

An architecture in which a single application instance serves many customers (tenants), with logical separation of data. It is the counterpart to single-tenant, where each customer gets a separate instance. In the AI context, multi-tenant lets you share the cost of infrastructure and the model across many customers, while single-tenant gives stronger compliance isolation.

Process mining (reconstructing processes from system logs)

A technique for reconstructing how business processes actually run, based on system logs (ERP, CRM, BPMS) rather than on documentation or interviews. It shows what really happens, not what is supposed to happen. The gap between the documented picture and the real one is typically 40 to 60 percent.

Prompt injection (injecting malicious instructions into an LLM)

A class of attacks on LLM-based applications in which external data (from a user, a document or a database) contains instructions that hijack control of the model. OWASP ranks prompt injection number one on its LLM Top 10 list of security threats for 2025. The vulnerability rate in production enterprise applications hovers around 76 percent.

RAG (Retrieval-Augmented Generation)

An architecture in which, before generating an answer, an LLM receives dynamically retrieved fragments from the client's knowledge base instead of relying solely on its training knowledge. RAG lowers the hallucination rate from 10 to 27 percent down to 0.5 to 3 percent for information tasks, and makes AI deployment possible without fine-tuning the model.

Shadow AI (hidden use of AI by employees)

The use of AI tools by employees outside the company's knowledge and policy. Pasting customer data into ChatGPT, source code into Claude, contract fragments into Copilot. The consequence: data leakage, GDPR risk, AI Act risk and breach of professional confidentiality.

TCO AI (total cost of ownership of an AI solution)

The sum of all costs of implementing, maintaining and developing an AI solution over a 3 to 5 year horizon. It covers model licences, infrastructure, team work, integrations, training, compliance and monitoring. AI TCO is typically 3 to 5 times higher than the initial deployment cost, which is the only figure presented in a vendor's commercial proposal.

Time to Value AI (time to the first measurable value)

The number of months from the decision to deploy AI to the first measurable impact on P&L or an operational KPI. In enterprise-class projects the median TTV is 14 months. In projects with workflow-first design and a clear scope it drops to 4 to 7 months. TTV is today the one metric that boards use in front of the CFO.

Tokenisation (splitting text into pieces for an LLM)

The process of turning input text into a sequence of tokens, the smallest units a model operates on. A token is typically part of a word, a whole word or a punctuation mark. In English one token corresponds on average to 0.75 of a word, in Polish to 0.5 of a word. The number of tokens determines the cost of a query, the context window limit and the response time.

Vendor lock-in (dependence on an AI vendor)

A state in which migrating from one AI model vendor to another becomes economically irrational. In the AI context it stems from three sources: an incompatible API, incompatible prompt semantics (model A and model B interpret the same instructions differently), and fine-tuned customisation that cannot be transferred. The cost of migration typically equals the cost of the original deployment.

Workflow-first design (processes before choosing technology)

A principle for designing AI deployments under which you first design the target business process and only then select the technology. The reverse of the dominant practice, in which the choice of AI model precedes the redesign of the process. Companies that apply workflow-first achieve twice the ROI from AI.