Glossary

The written rule set for staff on which AI uses are allowed, which are not, and which need approval. It is written for people, not for the model.

A variant of retrieval-augmented generation in which the model decides which sources to search, whether the results are good enough and whether to search again. It suits questions that span several sources, such as the tariff guide, CRM and outage notices, at the price of more calls, more latency, more cost and harder testing.

Not the same as Retrieval-augmented generation: Plain RAG runs one fixed retrieval step before the model answers; agentic RAG lets the model choose and repeat retrieval at run time.

A model in a loop that chooses its next action from a set of tools, observes the result and decides again. It fits cases where the path varies, such as gathering evidence for a complex billing dispute, provided each step can be checked; regulated, high-impact actions still need human approval.

Not the same as Deterministic workflow: A workflow fixes the path at design time; an agent picks it at run time, trading predictability for flexibility.

The single decision body that approves, rejects, defers or stops AI initiatives against shared budget, data, skills and risk. Members typically include the CIO or CTO, business-unit owners, enterprise architecture, finance, security, legal, data governance and the AI platform team; a smaller operator can give the remit to an existing architecture board or investment committee.

Under Article 4 of the EU AI Act, providers and deployers take measures to support AI literacy among the people who operate and use their systems. In practice it means staff understand what the system can and cannot do, and when to check or override its output; the Digital Omnibus amended the wording of the article.

Finding observations that depart from an established normal pattern, such as unusual traffic on a link, an odd usage spike on an account or a cabinet alarm sequence that has not been seen before. It is useful only where someone will investigate what it flags.

A defined interface through which one system exposes data or actions to another, for example a TMF Open API for product inventory or trouble tickets. AI applications reach systems of record through APIs like any other consumer.

Not the same as Model Context Protocol: An API exposes one system; MCP is a standard way for AI applications to discover and call such APIs and data sources.

A retained record of who asked, what was retrieved, which model and prompt version answered, which tools were called and who approved. It lets an operator show what a contact-centre agent was told, and why, if a customer later disputes it.

The measured value of the outcome before the change, agreed with finance before the pilot, such as current dispute handling time. Without it, an improvement cannot be shown afterwards.

Turning a claimed benefit into a measured one, tracked against the baseline through at least a full budget cycle after go-live. The business owner, not the AI team, owns it; saved minutes only become reduced cost if someone owns the operational change that banks them.

The choice between building an AI capability on models and a platform the operator controls, and buying it as a feature inside a product already licensed, such as a CRM or ITSM suite. Buying is faster but inherits that product’s data model and gives least control over behaviour and data path; building gives control and moves serving, evaluation and run cost onto the operator.

Splitting documents into passages before they are embedded and indexed. Chunk size and boundaries quietly decide answer quality: a tariff exception split from its rule can be retrieved without the context that qualifies it.

Tying each statement in an answer to the passage it came from, so a user can open the source clause before relying on it. The application should check that citations point to passages actually retrieved, and that the passage supports the claim; a citation proves where text came from, not that it is right.

Regression, decision trees, gradient boosting and similar models trained on structured data. They suit churn, fraud, capacity and fault prediction on tabular data, and need clean, well-defined features, which are often the hardest part in a brownfield BSS estate.

Not the same as Large language model: Classical models predict a value from structured features; an LLM generates text from unstructured input.

Interpreting images and video, for example checking a photo of a completed fibre installation or reading equipment labels in field service. Performance drops on images unlike the training set, such as poor light or unfamiliar kit.

The maximum amount of text, measured in tokens, that a model can take in and produce in one request, including instructions, retrieved passages, conversation history and the answer. A larger context raises cost and latency per request and does not by itself make the model use the content well.

A label on every source the assistant can retrieve from, which decides where it may be hosted, how long it is retained and who may ask about it. Network and subscriber data may also fall under national telecoms regulation.

The pipelines and stores that collect, clean, govern and serve data to models; a lakehouse or warehouse holds curated customer, billing, service and network data. A full platform pays off when many use cases share structured data, such as churn or fault correlation; a RAG assistant over a few hundred documents does not need one.

Whether the content and data a use case needs exist, are owned and are good enough, checked at a gate before solution architecture. Where it fails, fix the data at source with its owner or rescope; proceeding and promising to clean up later is the wrong response.

Machine learning with large neural networks, which handles unstructured input such as text, images and audio. It is harder to explain why a given output was produced, and for small tabular problems simpler models are cheaper.

A defined sequence of steps where the same input always follows the same path, with known branches and outcomes. Number porting, provisioning and refunds belong here; an AI step can sit inside the workflow without changing that.

Not the same as AI agent: A workflow decides the path at design time; an agent decides it at run time.

A list of numbers that represents the meaning of a piece of text, so that passages with similar meaning sit close together. The embedding pipeline splits, embeds and keeps documents current; a stale index answers from last quarter’s tariffs.

A model that turns text into vectors so that similar meanings sit close together, used to power semantic search and retrieval. It produces no text, and similarity is not relevance: the nearest passage can still be the wrong one.

Not the same as Large language model: An embedding model scores closeness of meaning; an LLM generates the answer.

Regulation (EU) 2024/1689, amended by Regulation (EU) 2026/1744 (the Digital Omnibus on AI): binding law on providers and deployers of AI systems placed on the market or used in the EU, and on providers of general-purpose AI models. It regulates uses, not technologies, with obligations scaled to risk: prohibited practices, high-risk requirements, transparency duties and AI literacy.

A curated set of realistic questions with agreed good answers, reviewed by people who know the domain, scored before go-live and after every change to model, prompt, index or content. It replaces pass/fail testing: a build is above or below an agreed quality bar, not "passed".

Further training an existing model on your own examples to change its behaviour, tone or format, such as classifying fault notes into your own categories. It is the wrong tool for facts that change, such as tariffs, and modifying a model can change an operator’s role under the AI Act, so take advice before fine-tuning at scale.

Not the same as Retrieval-augmented generation: RAG supplies current facts at request time; fine-tuning changes behaviour by training. Facts that change belong in retrieval.

A large model trained on broad data that can perform many tasks without task-specific training, and on which applications or further fine-tuning are built. Large language models are the most common kind in enterprise use.

Not the same as General-purpose AI model: "Foundation model" is a technical description; "general-purpose AI model" is the legal category the AI Act attaches provider obligations to.

The model is given descriptions of a few functions and returns a structured request to call one; the application, not the model, makes the actual call, for example a TMF API query on product inventory. It suits a small, fixed set of tools in one application.

Not the same as Model Context Protocol: Function calling wires tools into one application; MCP standardises tools so many AI applications can discover and reuse them.

Binding EU law on controllers and processors of personal data, applying from 25 May 2018: lawful basis and purpose limitation, minimisation, data subject rights, impact assessments where processing is likely to be high-risk, limits on solely automated decisions with significant effects, and rules on international transfers. It applies in full wherever an AI system processes personal data, whether or not the system is high-risk under the AI Act.

Not the same as EU AI Act: GDPR governs the processing of personal data; the AI Act governs AI systems by their intended use. Both can apply to the same assistant.

The AI Act category for models that can serve many purposes, with obligations falling mainly on the providers of those models. An operator that uses such a model through an API is normally a deployer of an AI system, not a model provider, although modifying the model can change that.

AI that produces new content, such as text, code, images or speech, rather than a score or a label, usually from a foundation model such as a large language model. The output is fluent but not guaranteed to be correct, and the same input can give a different answer each time, so it needs grounding, guardrails and evaluation that a classical model does not.

Not the same as Classical machine learning: Classical ML predicts a value or a class from structured data and is cheaper and easier to explain; generative AI writes new content from unstructured input.

Controls around a model that constrain what goes in and comes out: input checks, output filtering, refusal rules, relevance thresholds and limits on which tools can be called. They reduce harm; an instruction in the prompt is not access control, so permissions must still be enforced outside the model.

Fluent, plausible output that is not supported by the source data or by fact, produced because a language model generates the most plausible continuation rather than a verified answer. RAG and citations reduce it; they do not remove it.

Not the same as Retrieval miss: A hallucination is the model inventing content; a retrieval miss is the model faithfully answering from the wrong passage.

Under Article 6 of the AI Act, either a safety component of a product covered by the EU legislation in Annex I or a use listed in Annex III, carrying requirements on risk management, data governance, documentation, logging, human oversight, accuracy and robustness. Telco uses that can qualify include evaluating staff performance, assessing individual creditworthiness and safety components in critical digital infrastructure; classification follows the use, so scope creep can change it.

A defined point where a person reviews, approves, corrects or rejects AI output before it takes effect. It belongs in front of any action that changes a system of record or reaches a customer, such as a goodwill credit, and the approver must see enough to judge it.

Combining semantic (vector) search with keyword search and metadata filters. Exact identifiers such as fault codes or tariff codes are better found by keyword than by similarity, so hybrid search catches cases a vector index ranks badly.

Using a trained model: one request in, one output back, every time someone asks a question. Almost all enterprise AI work at an operator is inference, so the recurring cost and the operational load sit here.

Not the same as Training: Training produces a model, rarely and mostly by providers; inference uses it, on every request.

ISO/IEC 42001:2023, a voluntary, certifiable standard setting requirements for an AI management system: policy, roles, risk and impact assessment, controls, monitoring and continual improvement. It shares the management-system structure of ISO/IEC 27001, and is not by itself evidence of AI Act compliance.

Not the same as NIST AI Risk Management Framework: ISO/IEC 42001 is a certifiable management-system standard; the NIST AI RMF is voluntary guidance and not certifiable.

The owned, versioned content an assistant retrieves from: tariff guides, terms, handling procedures, fault procedures. It is business content with named owners, not an IT asset, and when nobody owns its freshness the assistant becomes confidently out of date.

A foundation model trained on large volumes of text that generates language: drafting, summarising and answering from supplied documents. It is the wrong tool for arithmetic, exact lookups or decisions that must be identical every time.

Not the same as AI agent: The LLM generates text; the application around it (agent or assistant) handles identity, retrieval, logging, refusal and integration. Swapping the model rarely fixes an answer broken by bad retrieval.

The time a request takes, usually measured as time to first token and end to end. Agents that make many model and tool calls add latency step by step, which matters on a live customer call.

The operational discipline for models in production: versioning, deployment, monitoring of quality, drift and cost, and rollback; MLOps is the older term from classical machine learning. It is needed from the first production release and needs an owning team.

Algorithms that learn patterns from historical examples rather than from hand-written rules, such as which tickets will breach SLA. It learns the past, including its biases and gaps; where a rule is known and stable, write the rule.

A service that exposes tools or data to AI applications through the Model Context Protocol, for example one wrapping a trouble-ticket API. When it is down the agent cannot act, so it is registered in the CMDB as a dependency of the AI service and monitored with synthetic calls.

An open protocol that standardises how an AI application discovers and calls tools and reads data sources, so a connector written once can serve several applications. It carries calls; it does not decide who may make them, so it is not a security model.

Not the same as API: An API exposes a system; MCP is a standard way for AI applications to discover and call such APIs. TMF637 still serves product inventory; an MCP server can wrap it.

A decline in output quality over time without a deliberate change: the world moves away from the training data, content changes, or a provider updates a model version. It is caught by pinning versions and re-running the evaluation set, not by availability monitoring.

A single front door between applications and models that handles authentication, routing, rate limits, cost tracking and logging. It is worth having once more than one application or model is in play; without one there is no central control of cost or model version.

Several agents with separate roles that hand work to each other, such as one diagnosing a fault and another drafting the customer update. It is emerging, not established: errors compound across hand-offs, so it belongs in bounded experiments rather than processes that need predictable outcomes.

Understanding and producing text: summarising, classifying, extracting and answering, for call notes, dispute letters, contract clauses and knowledge search. Where the data is already structured in a system of record, query the system instead.

Voluntary guidance of US origin (AI RMF 1.0, January 2023) organised in four functions: Govern, Map, Measure and Manage, with a Generative AI Profile (NIST AI 600-1). It helps structure thinking about AI risk; it is not certifiable.

Not the same as ISO/IEC 42001: The NIST AI RMF is guidance; ISO/IEC 42001 is a certifiable management-system standard.

Being able to see what an AI service did on each request: traces of what was retrieved, which tools were called, token use, latency and outcome, alongside availability. Conventional monitoring shows whether the service is up; AI observability is needed to see whether it is right and what it costs.

A model whose trained weights are published so an operator can run it on its own infrastructure. It gives full control of version and data, and full responsibility for serving, patching and evaluating it.

How AI capability is organised. Centralised: one team builds and runs everything, which becomes a bottleneck. Federated: each business unit has its own capability, which duplicates platforms and security. Hybrid (hub and spoke): a central team owns platform, guardrails and standards while business units own use cases and outcomes; most operators move towards it as the portfolio grows.

The gate at which run teams accept ownership of a live AI service and the project team steps back. It needs a trained run team, a funded run cost and an improvement backlog; a service still run by the project team has not been handed over.

The time it takes for cumulative net benefit to repay the up-front investment. It is sensitive to the measured benefit: halving the saving per contact can stretch payback far beyond the plan.

A live trial with real users doing real work in a limited scope, such as one care site and one product family, with measurement against the baseline in place. Its exit gate is evaluation scores, user feedback and a measured effect on the work, approved by the business owner.

Not the same as Proof of concept: A PoC shows technical feasibility on representative data; a pilot shows measured value with real users.

Estimating a future value or likelihood from history: churn risk, traffic growth on a link, likely failure of a cabinet battery. It is worth doing only where a decision changes because of the forecast, and it breaks when conditions change in ways the history did not contain.

The gate before deployment that confirms a support model, monitoring including quality, runbooks and rollback. It stops services with no run owner, no CMDB entry or no quality monitoring.

An AI use banned outright under Article 5 of the AI Act. Among them is inferring the emotions of people in the workplace except for medical or safety reasons, which is directly relevant to any proposal to score contact-centre agents’ emotional state.

The input a model receives: instructions, retrieved passages, conversation and the user’s question. Prompting is the cheapest way to shape behaviour, but prompts are brittle across model versions and must be versioned and tested like code.

Text written to redirect an assistant, placed in content it reads, such as a customer email attached to a trouble ticket or a field in a CRM record. No known control removes it, so design assumes it will sometimes succeed and limits the damage: fewest tools, read separated from write, user-scoped credentials and human approval for actions.

A test of technical feasibility on representative data, not curated examples, such as an assistant scored against historical billing enquiries. Its exit is a met quality threshold and an estimated cost per request; a working demo is not evidence.

Not the same as Pilot: A PoC answers "can it work?"; a pilot answers "does it help real users enough to fund?"

The agreed threshold on the evaluation set, plus safety and cost limits, that an AI release must meet. It is set by the business owner before the pilot; lowering it after a miss is the wrong response.

A second scoring pass over retrieved passages that reorders them by relevance to the question before they go to the model. It catches cases where the nearest passage by similarity is not the right one, such as a superseded tariff ranked above the current one.

A review of the gap between what an AI initiative needs and what the organisation can supply today, across ten dimensions from business strategy and data to change management and operational support. Programmes more often stall on these gaps than on model limits.

Deliberately attacking an AI system before release, including with prompt injection and misuse cases, to find failures the evaluation set does not cover. Findings feed back into the evaluation set and security controls.

A layered map of where AI components sit in an operator’s estate and who owns each in production: channels, AI applications, models and gateway, retrieval, MCP and API integration, data platforms, security and identity, hosting, observability and IT service management. It is used to check that a solution has an owner at every layer.

An umbrella term for the principles and controls that keep AI use lawful, fair, transparent, secure and accountable. It has no single legal definition; what makes it real is named risk owners, controls and evidence, not a statement of principles.

Finding the passages most relevant to a question from indexed content, filtered by the asking user’s entitlements, and returning them with document references. Entitlement filtering happens before retrieval, not after.

A failure in which retrieval returns the wrong passage, or none, and the model answers from it. The answer can be cited, fluent and wrong; the fix lies in content, chunking, ranking or threshold, not in the model.

Not the same as Hallucination: In a retrieval miss the model is faithful to the wrong source; in a hallucination it invents content.

A pattern that retrieves relevant passages from the organisation’s own sources, gives them to the model and has it answer from them with citations. It suits answers that must reflect current, organisation-specific documents; it reduces unsupported answers, it does not remove them.

Not the same as Fine-tuning: RAG supplies facts at request time; fine-tuning changes behaviour by training.

Net benefit relative to the investment that produced it. For AI, benefits are avoided cost, revenue or risk reduction, each uncertain until measured; ROI is one input to a funding decision, alongside strategic fit, feasibility, risk and readiness.

Not the same as Total cost of ownership: TCO counts all costs over the life of the service; ROI sets benefits against those costs.

A model run by the operator on its own private cloud, on-premises hardware or tenancy, rather than called as a provider service. Compute cost replaces per-call fees, and model upgrades, patching and evaluation become the operator’s permanent work.

A number expressing how close a passage is in meaning to the question, used to rank retrieval results. A threshold below which nothing is passed to the model lets the assistant say it does not know instead of answering from nothing.

A smaller language model that runs on modest hardware, including on-premises, suited to narrow, high-volume tasks such as ticket classification or where data must not leave the operator’s estate. It is less capable outside the task it is tuned for; routing simple requests to one can change model cost substantially.

Not the same as Large language model: An SLM trades breadth for lower cost and easier local hosting; an LLM handles broad reasoning over long, varied input.

Cloud services operated within a jurisdiction, under local legal control and staffing, to meet national security or residency obligations. It is usually priced above standard cloud and offers a narrower range of services.

Converting speech to text, for example transcribing contact-centre calls for summarising and quality review. Accents, line quality and crosstalk cause errors that flow into everything downstream, and a transcript has value only if something consumes it.

A decision point between delivery stages where named approvers check defined evidence before more money is spent: business case, data readiness, architecture, security, pilot exit, production readiness and handover. Stopping at a gate is a valid outcome, not a failure.

Constraining a model’s output to a defined format, such as JSON with named fields, so that software can consume it reliably, for example a dispute category and reason code. The format can be enforced; the correctness of the values cannot.

The system that holds the authoritative version of a fact: CRM for the customer, billing for charges, inventory for what the customer has, the product catalog for commercial terms. In an AI design the model is rarely a system of record and the knowledge base is a system of reference; an AI that holds a fact no system of record holds will diverge.

A generation setting that controls how much randomness the model uses when choosing each token. Lower values make output more repeatable, not more correct; no setting makes a language model deterministic in the sense a workflow is.

Standard REST interfaces for telco systems, such as TMF621 trouble tickets, TMF622 product ordering and TMF637 product inventory. AI tools that read or change operator systems go through them like any other consumer, often wrapped by an MCP server.

The unit a language model reads and writes: a word or part of a word. Context limits, latency and usage charges are all counted in tokens.

The recurring cost of model use: tokens × price per token × volume of requests. It grows with adoption, longer prompts and agent step count, so a design that is cheap with ten agents can be expensive with two thousand.

All costs of an AI service over its life: licences, model usage, infrastructure, integration, data preparation, evaluation, run and support, and change management. Data preparation and change management are the lines most often underestimated or omitted.

Not the same as Return on investment: TCO is the cost side only; ROI weighs benefits against it.

Producing a model by running large compute over large datasets, done rarely and mostly by model providers. Operators mostly consume trained models and occasionally fine-tune them.

Not the same as Inference: Training creates the model; inference uses it on every request and carries the recurring cost.

Under Article 50 of the AI Act, duties such as designing a system that interacts directly with people so they know they are dealing with an AI, unless that is obvious from the context. A customer-facing chatbot will usually carry it.

The standard record every AI proposal arrives in: problem, business owner, outcome, candidate capability, data availability, feasibility, complexity, risk, investment, expected return and strategic fit. Each field forces a question the proposer may not have asked, including whether the answer is AI at all.

A store that finds the passages closest to a query vector, often combined with keyword search and metadata filters. It is one way to search, not the only one; for small document sets keyword search may be enough, and access control must be carried into the index.

Coordinating a defined sequence of steps with known branches, retries and compensation, the same discipline as order orchestration in SOM. AI can be one step inside it, such as classifying a dispute before routing; most production designs that work are workflows with AI steps, with agents confined to bounded sub-tasks.