AI and automation glossary

Agent assist is a copilot layer inside the agent desktop that drafts replies, retrieves relevant policy, and suggests the next best action while a human stays in control of the conversation. It shortens handle time and reduces variation between agents without removing the person from the interaction.

An AI agent is software that resolves a customer request end to end, taking actions in connected systems rather than only answering questions. Unlike a scripted chatbot, an AI agent can look up an order, issue a refund, or update a record, then confirm the completed outcome back to the customer.

Automation rate is the percentage of incoming contacts handled entirely without a human agent. It measures reach rather than quality, so support teams read it alongside containment rate and satisfaction scores to confirm that automated volume is genuinely resolved rather than simply absorbed and later reopened by the customer.

Bot handoff is the transfer of a live conversation from an automated agent to a human, ideally with the full transcript and any collected context attached. A clean handoff prevents the customer from repeating information, which is the most common complaint recorded about automated support experiences.

Containment rate is the share of automated conversations that end without escalation to a human. It is the primary efficiency measure for AI support, but a high rate only indicates success when paired with satisfaction and repeat-contact data confirming customers actually got their issue resolved.

Conversational AI describes systems that understand and respond in natural language across chat, email, and voice. The category spans intent-based bots and modern language-model agents, and the practical difference between them is whether the system can only reply or can also take action in business systems.

A copilot is an assistive AI surface that works alongside a human agent instead of replacing them, suggesting replies, summarising history, and surfacing policy at the moment of need. Copilots are typically measured on handle time and answer consistency rather than on deflection or containment.

Entity extraction is the process of pulling structured values — an order number, an IBAN, a product name, a date — out of a customer’s free-text message. Reliable extraction is what allows an automated agent to act inside a business system rather than only classify the request.

An evaluation set is a fixed sample of real conversations used to score an AI agent’s quality before and after changes. Holding the sample constant is what makes results comparable, so evaluation sets are versioned and refreshed only deliberately when the underlying contact mix genuinely shifts.

A fallback is the safe response an AI agent gives when its confidence is too low to act, typically an offer to hand the conversation to a human. Well-designed fallbacks fail early rather than guessing, because a confidently wrong answer costs far more than an admitted gap.

Grounding is the practice of constraining an AI agent’s answers to approved knowledge sources and live system data rather than to the model’s own training. Grounding is the main defence against hallucination, and it is what makes an AI answer auditable back to a specific document or record.

Guardrails are the policy, tone, and action limits an AI agent is not permitted to cross, such as refusing to issue refunds above a threshold or to discuss topics outside its remit. Guardrails are enforced in the system rather than merely requested in the prompt.

A hallucination is a fluent but unsupported answer produced when a model is not grounded in real data. In customer service the risk is specific: an invented policy, price, or delivery date reads as authoritative to the customer and creates a commitment the business never made.

Human in the loop describes a review step where a person approves or corrects AI output before it reaches the customer. Teams typically apply it to high-risk intents such as refunds, cancellations, and complaints, then relax it selectively as measured accuracy on those intents proves out.

An intent is what a customer is actually trying to do, inferred from what they wrote or said rather than from the specific words used. Two very different messages can share an intent, which is why routing and automation are built on intent rather than on keyword matching.

An intent taxonomy is the structured list of intents a support operation recognises and reports on. A good taxonomy is granular enough to drive different actions and stable enough to trend over time; taxonomies that grow uncontrolled become unusable for reporting well before routing.

A large language model is the model class behind modern conversational AI, summarisation, and reply drafting. Trained on broad text corpora, it predicts language fluently but holds no inherent knowledge of a specific business, which is why grounding and retrieval are required before answering customer questions.

Latency is the delay between a customer’s message and the first useful response. In live chat and voice it directly shapes perceived quality, and it is one of the few AI quality measures a customer notices immediately, well before judging whether the answer was correct.

Model drift is the quiet degradation in AI performance that occurs as products, policies, and customer language change while the model and its knowledge stay fixed. Drift rarely announces itself; it appears as a slow rise in escalations and fallbacks on intents that previously performed well.

Natural language understanding is the layer that maps free text to intents and entities so a system can act on what a customer wrote. It answers what the customer wants and which values are involved; generation is a separate step deciding what to say back.

A prompt is the instruction set that defines how an AI agent behaves in a given task, covering role, tone, permitted actions, and output format. Prompts express intent but do not enforce it, which is why hard limits belong in guardrails rather than in prompt wording.

Retrieval-augmented generation grounds answers by retrieving an organisation’s own documents at the moment of answering and passing them to the model as context. It keeps responses current without retraining, and it makes each answer traceable to the specific source the system actually consulted.

An utterance is a single thing a customer said, exactly as the model receives it, before any cleaning or interpretation. Utterances are the raw material for intent classification and evaluation sets, and preserving them verbatim is what allows AI errors to be reproduced and diagnosed later.

Voice AI is automation that handles spoken conversations using speech recognition, language understanding, and speech synthesis. It carries constraints text automation does not: recognition errors on names and numbers, sensitivity to accents and dialects, and a much lower customer tolerance for response latency.

Workflow automation chains system actions — issue a refund, reissue a document, update a record — so a resolution completes itself without an agent performing each step. It separates an AI agent that resolves requests from a chatbot that can only describe how a request would be resolved.

Zero-touch resolution is an issue resolved end to end with no human involvement and no follow-up contact from the customer. It is a stricter measure than containment, because it requires the customer not to return, and it is the cleanest available proof that automation genuinely worked.