AI is moving quickly, and so is the language around it. This enterprise AI glossary explains the key terms shaping conversations at the Accelerate Tomorrow AI Summit (ATS), from AI agents and frontier models to governance, implementation and responsible adoption.
It is written for business leaders, transformation teams and AI practitioners who want to understand what these concepts mean in practice and why they matter for organisations turning AI strategy into measurable business impact.
Accelerate Tomorrow AI Summit (ATS)
The Accelerate Tomorrow AI Summit, or ATS, is an Germany's leading enterprise AI conference for business leaders, AI practitioners and technology innovators.
ATS brings together decision-makers to explore how organisations can move from AI strategy and experimentation to responsible implementation and measurable business impact. ATS 2027 will take place on 15–16 June 2027 in Berlin, Germany.
Agentic AI
Agentic AI often describes AI systems that can plan, make decisions and carry out multi-step tasks toward a defined goal with limited human input. Unlike a chatbot that only responds to a prompt, an AI agent may retrieve information, use approved tools, take actions in connected systems and adapt its next step based on the result.
For businesses, the opportunity is to automate complete workflows rather than isolated tasks. The challenge is ensuring the agent has clear boundaries, appropriate permissions and human oversight where needed.
AI Agent
An AI agent is a software system that uses AI to perform tasks in pursuit of a goal. It can interpret instructions, access selected information or tools, and select or propose its next action.
For example, an agent may prepare a sales brief from CRM data, triage customer-service requests or support employees with internal knowledge.
AI Alignment
AI alignment refers to ensuring that an AI system behaves in ways that reflect intended human goals, values and constraints.
In an enterprise setting, this means making sure AI systems follow organisational policies, respect permissions, avoid harmful actions and support the outcome they were designed to achieve.
AI Bias
AI bias occurs when an AI system produces systematically systematically unfair, unjustified or disparate outcomes. It can result from biased data, flawed assumptions, poor system design or the way a model is used.
Managing bias requires deliberate testing, diverse perspectives and clear documentation. It also requires ongoing monitoring, not only a one-time technical check.
AI Governance
AI governance is the set of policies, processes, roles and controls that guide how an organisation develops, buys, deploys and manages AI.
Effective AI governance helps organisations use AI securely, responsibly and in compliance with relevant requirements. It also helps ensure that AI supports business objectives. This includes decisions about approved tools and models, data access and risk assessment. It also includes accountability, monitoring and human oversight.
AI Maturity
AI maturity describes how prepared and capable an organisation is to adopt and scale AI. It is shaped by factors such as strategy, leadership support, data quality, technical infrastructure, skills, governance and the ability to move successful pilots into everyday operations.
AI maturity is not only about technical capability. It is also about whether an organisation can consistently turn AI investment into business value.
AI Operating Model
An AI operating model defines how an organisation makes decisions about AI and puts those decisions into practice. It covers areas such as leadership, responsibilities, governance, technology, data, talent, risk management and ways of working.
The right model depends on the organisation. However, it should clearly identify who can develop, approve, deploy and monitor AI systems.
AI Use Case
An AI use case is a specific business problem or opportunity where AI can create value. Examples include automating document processing, improving demand forecasting, supporting customer-service teams or helping employees find internal knowledge.
Strong AI programmes begin with clearly prioritised use cases. Each should have a defined business owner, measurable outcome and realistic path to implementation.
B
Benchmarking
Benchmarking is the process of evaluating an AI model or system against defined tests, metrics or alternative solutions. It can measure factors such as accuracy, reliability, speed, cost and safety.
Organisations use benchmarking to compare models and assess whether an AI system performs well enough for a specific business use case. However, a benchmark result should not replace testing in the organisation’s real operating environment.
Black Box AI
Black box AI describes an AI system whose internal process for producing an output, recommendation or decision is difficult to understand or explain. Users may see the input and output without being able to determine how the system reached its result.
This can create challenges in high-impact or regulated use cases. Organisations may need testing, documentation, monitoring and human oversight to manage these risks and build trust in the system.
C
Computer Vision
Computer vision is a field of AI that helps systems interpret and analyse visual information. This can include images, video and scanned documents.
Businesses use computer vision for quality inspection, fraud detection, inventory management and document processing. It also supports medical imaging.
D
Data Sovereignty
Data sovereignty is the principle that data is subject to the laws, regulatory requirements and government-access rules of the jurisdictions connected to it. These often include the places where data is stored, processed or controlled.
For organisations using AI, data sovereignty can affect decisions about cloud providers and model vendors. It can also affect where data is processed, what information can be shared with external systems and how organisations meet compliance requirements.
E
EU AI Act
The EU AI Act is the European Union’s legal framework for artificial intelligence. It takes a risk-based approach, with obligations that vary according to an organisation’s role, such as provider or deployer, and how an AI system is used.
For organisations operating in Europe, the Act makes AI governance, documentation, transparency, risk management and AI literacy practical business priorities. Its requirements are being introduced in stages. Some rules already apply. Key requirements for high-risk AI systems will apply from late 2027 and 2028.
F
Fine-Tuning
Fine-tuning is the process of further training an existing AI model on a specific dataset to improve its performance for a particular task, industry or style of response.
It can be useful when prompting alone is not sufficient. However, it requires high-quality data, clear evaluation and careful governance.
Foundation Model
A foundation model is a large AI model trained on broad datasets that can be adapted for many different tasks. Large language models are one type of foundation model.
Businesses commonly adapt foundation models through prompting, fine-tuning or retrieval-augmented generation (RAG), rather than building a model from scratch.
Frontier Model
A frontier model is among the most capable general-purpose AI models at this time. These models are generally trained on very large datasets and can perform a broad range of advanced tasks, such as reasoning, writing, coding, analysing documents and working across text, images and audio.
Frontier models can unlock significant new capabilities. However, organisations need to consider cost, security, reliability, data handling and where a smaller or more specialised model may be more suitable.
G
Generative AI (GenAI)
Generative AI, or GenAI, creates content in response to an instruction or other input. It can generate text, images, audio, video and code.
Businesses use generative AI to draft content, summarise information, analyse documents and create software. It can also assist employees. However, organisations still need controls to check accuracy, protect sensitive data and oversee high-impact uses.
H
Human-in-the-Loop
Human-in-the-loop means that a person remains involved in reviewing, validating or approving an AI system’s output or action.
This is especially important when AI affects customers, employees, financial decisions, safety, compliance or other high-impact areas. Therefore, organisations should design human oversight into the workflow, rather than add it only after an issue occurs.
I
Inference
Inference is the stage when a trained AI model processes new input and generates an output. That output may be a prediction, recommendation or response.
For example, when an employee asks an AI assistant a question, the model performs inference to produce its answer. Inference affects the speed, cost and user experience of an AI system.
J
Jailbreak
A jailbreak is an attempt to bypass an AI system’s safeguards, instructions or restrictions. For example, someone may use prompts to make a model reveal restricted information or perform actions it should refuse.
Organisations should test AI systems for jailbreak attempts. This is especially important when a system can access sensitive information, use tools or take actions in connected business systems.
K
Knowledge Graph
A knowledge graph organises information as connected entities and relationships. For example, it can link customers, products, suppliers, locations and documents.
Businesses use knowledge graphs to show how information relates to other information. They can improve search, data integration and decision-making. They can also give AI systems more reliable organisational context.
L
Large Language Model (LLM)
A large language model, or LLM, is an AI model trained to process and generate human language. It powers many generative AI tools used for writing, summarising, searching, analysing and conversational assistance.
However, an LLM does not automatically know an organisation’s internal policies, data or context. Businesses, therefore, need to decide what information it can access and how its outputs are checked before use.
M
Model Monitoring
Model monitoring is the ongoing process of checking whether an AI system continues to perform as intended after deployment. It can include measuring accuracy, reliability, cost, bias, security and changes in user behaviour or input data.
AI systems can change in performance over time. Monitoring, therefore, helps organisations identify problems early and maintain trust in the systems they use.
Multimodal AI
Multimodal AI can work with more than one type of information in the same system. These inputs can include text, images, audio, video and structured data.
For example, a multimodal system could analyse an invoice image, extract the relevant data and answer questions about it. As as a result, AI can support many business processes that involve more than text alone.
N
Natural Language Processing (NLP)
Natural language processing, or NLP, is a field of AI that helps systems understand, interpret and generate human language.
It supports applications such as chatbots, search, translation and sentiment analysis. It also supports speech recognition and document processing.
O
Open-Source Model
An open-source AI model is made available under terms that allow others to use, study, modify and share it. This normally requires more than releasing model weights alone.
Open-weight models make trained model parameters publicly available, but may not include the training code, data information or other materials needed to reproduce or fully modify the system.
Both can give organisations more flexibility and control. However, organisations still need to assess security, licensing, model quality, data handling and support requirements.
P
Prompt Engineering
Prompt engineering is the practice of designing and refining instructions for an AI system. It aims to improve the relevance, accuracy and consistency of its outputs.
It can help employees use generative AI more effectively. For repeatable business processes, organisations may standardise prompts, templates and guardrails. This is more reliable than relying on ad hoc instructions.
Q
Quantization
Quantization is a technical method that reduces the precision of the numbers used by an AI model. As a result, it can make a model smaller, faster and less expensive to run.
This is particularly relevant when organisations want to run models efficiently on their own infrastructure. It can also help when devices have limited computing capacity.
R
Responsible AI
Responsible AI is generally undestood as the practice of developing and using AI in a way that is fair, safe, transparent, accountable and respectful of privacy and human rights.
It is not a separate project that begins after an AI system is built. Instead, responsible AI needs to be part of how organisations select use cases, design systems, and manage data. It must also be part of how organisations test outcomes and oversee deployment.
Retrieval-Augmented Generation (RAG)
Retrieval-augmented generation, or RAG, is an approach that gives an AI model access to selected external information when responding to a question. This information can include company documents, product information or policy libraries.
Instead of relying only on what the model learned during training, RAG retrieves relevant information at the time of the request. Used well, it can make responses more accurate, current and grounded in approved organisational knowledge.
S
Synthetic Data
Synthetic data is artificially generated data that reflects the characteristics of real-world data. It is designed to reflect patterns or characteristics found in real-world data, without simply copying the original records. However, synthetic data is not automatically anonymous or free from privacy risk.
Organisations can use synthetic data to train, test or evaluate AI systems. It can help when real data is limited, sensitive or subject to privacy restrictions. However, organisations must still check its quality, as it can reproduce or introduce bias.
T
Training Data
Training data is the information used to teach an AI model to recognise patterns, make predictions or generate outputs.
The quality and relevance of training data strongly affect an AI system’s performance and reliability. Its governance can also affect the risk of bias.
U
Unstructured Data
Unstructured data is information that does not follow a fixed format or database structure. Examples include documents, emails, presentations, images, audio recordings and video.
Much of an organisation’s knowledge exists as unstructured data. AI can help businesses search, analyse and use this information. However, they must respect access permissions and data-governance requirements.
V
Vector Database
A vector database stores numerical representations of information, known as embeddings. It helps systems find content based on meaning and similarity, rather than only exact keyword matches.
Vector databases are common in retrieval-augmented generation (RAG) systems. They help AI models find relevant company knowledge before generating a response.
W
Workflow Automation
Workflow automation uses technology to complete repeatable tasks or move work between people and systems. It follows defined rules.
AI can make workflow automation more flexible. For example, it can interpret unstructured information, make recommendations or handle more complex steps. However, organisations should still set clear boundaries, approvals and accountability.
X
XAI (Explainable AI)
Explainable AI, often shortened to XAI, refers to methods and system designs that help people understand and communicate how an AI system produces an output, recommendation or decision.
The level of explanation needed depends on the use case. In high-impact or regulated settings, organisations may need to show what information influenced an outcome. They may also need to explain what limitations apply and when a human should review it.
Y
YAML
YAML is a human-readable file format used to structure configuration data. The name originally stood for “YAML Ain’t Markup Language.”
AI teams often use YAML files to define model settings, prompts, data pipelines, deployment environments and automated workflows. It helps teams manage complex configurations in a consistent, version-controlled format.
Z
Zero-Shot Learning
Zero-shot learning describes an AI model’s ability to complete a task without receiving task-specific training examples or examples in the prompt.
For example, a general-purpose language model may classify a new type of customer enquiry based only on an instruction. While this can be useful, organisations should test zero-shot outputs for accuracy in important business use cases.
