Platform and AI glossary
Definitions for executives and decision-makers in regulated Swiss industries, focused on pharma, medtech and OTC.
Algorithm
A sequence of unambiguous logical instructions for solving a problem. In enterprise AI, algorithms detect patterns in unstructured data, automate review paths or forecast metrics.
Answer accuracy
The share of an AI system’s answers to a defined question that matches the company’s approved reference, measured across several runs rather than a single answer. Because generative systems are probabilistic, the spread across repeats is itself a finding.
Answer Engine Optimization (AEO)
Aligning your own content so that answer-giving systems use it as a source. Unlike GEO, AEO targets the answer to a specific question, while GEO targets the entity’s visibility in synthesised overviews. In practice the two overlap strongly.
Artificial intelligence (AI)
Umbrella term for systems that perform tasks that previously required human perception or judgement. Relevant in companies because of the shift from deterministic automation to probabilistic creation and synthesis of content.
Audit trail & traceability
Complete, tamper-proof documentation of data access, model outputs and approval steps. For high-risk AI systems the EU AI Act requires automatic record-keeping (Art. 12); Swiss data protection law can require logging in certain cases (Art. 4 DPO).
Bias (algorithmic)
A systematic skew in training data, model architecture or scoring logic that leads to disadvantageous or factually wrong results for certain groups or cases. The EU AI Act explicitly requires measures on data quality and bias for high-risk systems.
Big data
Large, heterogeneous and fast-changing data sets that overwhelm traditional relational databases. The basis for training and fine-tuning domain-specific models.
Computer vision
Automated capture and analysis of image and video data by neural networks, for example for quality control, sensor monitoring or document capture.
Conformity assessment
The procedure that demonstrates that a product meets the legal requirements. For AI systems, whether a notified body must be involved also decides whether they count as high-risk under Article 6(1) of the AI Act.
Deep learning
A branch of machine learning based on multi-layer neural networks. It detects hierarchical patterns in complex raw data without manual feature extraction.
Digital sovereignty
A company’s ability to control platforms, interfaces and data independently of individual vendors and proprietary ecosystems.
Drift (model and data drift)
The decline in model quality in operation because input data or reality has changed since training. Drift is why a system validated once silently gets worse without ongoing monitoring.
Ethical guardrails
Binding, technically enforced limits on the use of AI, as opposed to policy papers without a checkpoint. Guardrails only work when they act at a point in the process: an approval that is not given, an output that is filtered, a record that is not transmitted.
EU AI Act
Regulation (EU) 2024/1689, in force since 1 August 2024. It classifies AI systems by risk and attaches tiered obligations. The deadlines were moved by Regulation (EU) 2026/1744 (“Digital Omnibus on AI”, in force since 27 July 2026): Annex III systems from 2 December 2027, product-related systems under Annex I from 2 August 2028. Material that names 2026 or 2027 as the date for all high-risk systems has been out of date since July 2026.
Explainability
The ability to show on what basis a system produced an output. Without it, no root-cause analysis is possible after an incident, and without root-cause analysis no sound corrective action.
Generative AI
Systems that create new content (text, code, summaries, molecular structures) from large training data rather than merely classifying data.
Generative Engine Optimization (GEO)
The methodical maintenance of company information and approved sources so that generative search systems (ChatGPT, Perplexity, Gemini) demonstrably draw on them in their answers.
Grounding
Binding a model output to a named, verifiable source. A grounded system gives not only an answer but also the document it comes from. That shifts the check from “sounds plausible” to “is actually written there”.
Hallucination
A factually wrong or invented statement that a generative system outputs in the same tone and format as a correct one. The danger lies not in the error itself but in the fact that it reads like a correct answer. Hallucinations follow from how generative models work and cannot be removed by more training data alone.
High-risk AI system
A classification under the EU AI Act with extensive obligations on risk management, data quality, technical documentation, transparency and human oversight. There are two routes: Article 6(1) with Annex I covers systems that are a safety component of a product or are themselves such a product, including medical devices under the MDR and in-vitro diagnostics under the IVDR where a notified body is involved. Annex III covers stand-alone uses, for example in employment, education or law enforcement.
Human in the loop
A binding human review and approval step before a decision or publication. The term only holds if the reviewer can actually judge the output and has time to do so; otherwise it becomes a formal signature without real control.
Machine learning (ML)
Methods in which algorithms build statistical models from data to make predictions or classifications without rigid programming.
Natural language processing (NLP)
Methods for machine analysis, translation and structuring of natural language. The basis for document analysis and semantic search.
Predictive analytics
Using statistical methods, historical data and machine learning to calculate the probability of future events.
Prompt injection
An attack that smuggles instructions into content a system processes, such as a document, a web page or an email, to change its behaviour. Because language models do not reliably separate instruction from content, injected text acts like a command.
Red teaming
The systematic search for a system’s misbehaviour through targeted attacks and edge cases, before go-live and repeatedly afterwards. Unlike a functional test, it checks not whether the system does what is expected but what else it does under pressure.
Regtech
Technology solutions for meeting regulatory obligations in a standardised, efficient and traceable way.
Retrieval-augmented generation (RAG)
An architecture in which relevant passages are retrieved from a curated collection and passed to the model before it answers. This reduces invention and makes sources citable, but does not remove hallucinations: the synthesis step is still a language model. A search with wrong permissions can also expose confidential content to unauthorised users.
Shadow AI
Employees using unapproved AI services, usually with business data. It often arises where a ban was issued without a usable internal alternative; use then moves to private devices and out of control.
Source governance & ownership
The binding assignment of responsibility for primary sources: their currency, approval status and machine-readable access rights.
Supervised learning
A training method for machine-learning models using labelled data sets with known target values.
Unsupervised learning
A training method in which algorithms find groups, outliers and hidden patterns in raw data without given target values.
Zero data retention
A provider’s contractual and technical commitment not to store inputs and outputs beyond processing and not to use them to train its own models. The two points must be checked separately: no training does not automatically mean no retention.