Glossary

A structured, machine-readable description of a government service that an AI agent can understand and act on.

The dimension of a service description that declares what the service is, which department owns it, and what it does.

The dimension of a service description that defines who qualifies, expressed as deterministic rules that code can evaluate.

The dimension of a service description that defines the valid sequence of steps from start to completion.

The dimension of a service description that specifies what data fields are needed, from whom, and what consent is required before sharing.

Making a government service machine-readable so an AI agent can understand it, assess eligibility, and act on it for a citizen.

The strict separation between what the AI language model decides and what code decides.

A significant moment in a citizen's life that triggers interaction with multiple government services across departments.

A single government service that a citizen can access through the agent — such as registering a death, applying for bereavement support, or notifying DVLA.

A structured data-collection interface embedded within the conversation, replacing traditional government web forms.

A discrete unit of work within a journey — something the agent will do or something the citizen must confirm — with a tracked status.

A proactive message from the system alerting a citizen to a service they may be eligible for, a deadline approaching, or an action they need to take.

The citizen's portable store of verified credentials — digital equivalents of documents like driving licences, passports, and benefit entitlements.

A government-issued verifiable credential held in the citizen's wallet — the digital equivalent of a physical document.

The ability for one citizen to act on behalf of another — a parent for a child, an executor for a deceased person, a carer for a vulnerable adult.

A specific, informed permission from a citizen for the agent to share named data fields with a named department for a stated purpose.

A single piece of information a service requires from a citizen — such as date of birth, National Insurance number, or address.

Metadata on every data field recording which department originally holds or issued it.

Data sourced directly from a government department — the authoritative version of a fact about a citizen.

Data entered by the citizen themselves, not yet confirmed against a government source.

Data derived by the agent from context — such as inferring marital status from a bereavement report. Always flagged as inferred and confirmed before use.

Automatically populating data fields with information already held — from the wallet, a previous service, or a verified government source — so the citizen is not asked twice.

A mapping that recognises when different services ask for the same data under different names — "NI number", "national insurance number", and "NINO" all resolve to one field.

Merging data requirements across multiple services so the citizen is asked once for each piece of information, even when several services need it.

A single, timestamped record of something the system did — invoked a capability, evaluated eligibility, transitioned state, granted consent, or called the language model.

A separate, append-only audit layer that records every action the system takes — independent of the control logic that makes decisions.

A citizen-facing record of an action taken by the agent, showing what was done, the outcome, and what data was shared.

Stepping through a past interaction event by event, reconstructing the exact state of the conversation, data, and decisions at each point.

The complete, immutable record of every decision, data share, and consent grant across a citizen's interaction — available to the citizen, the department, and oversight bodies.

The central coordinator that manages the flow of a citizen's interaction — routing between services, injecting context, and delegating language-only work to the AI.

A formal model of the valid states a service interaction can be in and the valid transitions between them — preventing the agent from skipping steps.

The component that runs a citizen's data against published eligibility rules, returning a definitive pass or fail with reasons.

The single gateway through which every service action passes — ensuring that every invocation is policy-checked, consent-verified, and trace-recorded.

The component that checks whether consent exists before any data is shared, enforces standing preferences, and records all consent decisions.

The component that tracks which data fields a service requires versus which have already been collected or verified — so the agent only asks for what is genuinely missing.

When a citizen's situation falls outside the standard path defined by a service's policy — and the system must decide whether to adapt, flag, or escalate.

The structured escalation of a citizen's case from the AI agent to a human, with full context about what has already been done.

Moving a citizen's interaction to a higher level of support — from agent to human, or from front-line to specialist — when the situation requires it.

Automatic detection of distress, self-harm indicators, or vulnerability in a citizen's messages, triggering immediate escalation to human support.

The department-facing platform where service owners author, publish, and monitor the service descriptions that make their services accessible to agents.

A department-level view showing how many services have been made legible versus how many remain to be done.

A per-service assessment of which of the four dimensions (identity, eligibility, journey, data sharing) have been published, and which are missing.

The complete inventory of a department's services, whether or not they have been made legible — the starting point for prioritisation.

A classification system for government services based on their interaction pattern — application, registration, notification, lookup, payment — which determines the shape of their service description.