AI agent
A program that holds a conversation and can complete a task inside it, not merely answer. The difference from a chatbot is that an agent reaches for data and triggers operations, usually with human consent.
Resources
Glossary
Explanations without marketing. Forty terms that come up around AI agents in customer service, written so you can take them into a meeting with legal or with the board.
Terms
No term matches this query.
Language model
A model trained on text that predicts the next words. It has no access to your data and does not know what is true; it can only assemble a sentence fitting the context it received.
Hallucination
An answer that sounds credible but has no grounding in reality or in sources. It follows from how the model works rather than being a fault, which is why it must be countered outside the model.
Prompt
The instructions and context handed to the model before it generates an answer. It usually holds the agent's role, behaviour rules and knowledge passages retrieved for the question.
Persona
A description of how the agent should address the other party: tone, form, level of detail. It governs style, not substance, because substance comes from the knowledge base.
Intent
The recognised purpose of what the other party said, for example an order status question or a refund request. Used to route the conversation to the right place.
Context window
How much text the model can consider at once. This is why selected knowledge base passages reach the model rather than its entire contents.
Temperature
A parameter controlling how far the model strays from the most likely next word. A lower value gives repeatable answers, a higher one looser ones.
RAG
Retrieval augmented generation. Instead of relying on the model's knowledge, the system first finds the right passages in your documents and only then asks the model to phrase an answer.
Knowledge base
The body of content the agent grounds answers in: articles, procedures, policies. In AIvatary content has a state and a version, and the agent sees only published versions.
Embedding
Turning a passage of text into a sequence of numbers describing its meaning. That makes it possible to find a passage matching a question even when it uses different words.
Chunk
The piece a document is split into before retrieval. Too large blurs the topic, too small loses the context.
Hybrid retrieval
Combining meaning based and word based search. The first copes with questions asked in different words, the second with catalogue numbers and proper names.
Reranker
A step that reorders retrieved passages against the specific question. It improves precision at the cost of response time.
Citation gate
A mechanism checking whether an answer is grounded in sources before it reaches the other party. In AIvatary an ungrounded answer is rejected whole and the conversation goes to a human.
Source citation
Naming the document and place an answer came from. It lets the other party verify the content and the team work out what to fix.
Gap register
A list of questions with no coverage in the knowledge base. It is built from real conversations, so it orders editorial work by actual need.
Editorial cycle
The path of content from draft through review to publication and archive. It keeps a working note from becoming an answer to a customer.
Widget
A conversation window embedded on a page with one line of script. It runs in a separate style tree, so it does not mix with the host page's appearance.
Hero mode
Placing the avatar inside the page layout instead of a corner bubble. The conversation begins where the visitor is already looking.
Handoff
Passing a conversation to a human agent along with the full history and the reason. A normal part of support, not a failure.
Confirmation card
A request to perform a data changing operation that a human must approve. Instead of issuing a refund, the agent prepares it for approval.
Barge in
A voice agent's ability to fall silent immediately when the other party starts speaking. Without it a conversation feels like a recorded announcement.
Speech synthesis (TTS)
Turning the answer text into sound. It decides how the agent sounds, not what it says.
Speech recognition (STT)
Turning what the other party says into text. The quality of this step decides everything downstream in a voice conversation.
Turn latency
The time from the end of the other party's utterance to the start of the agent's reply. Far more noticeable in voice than in chat.
Audit log
An append only record of events that lets you reconstruct what happened in a conversation. A correction is a new event, not a swap of the old one.
Decision trace
A record of which content was retrieved and what was decided in a given turn. The answer to how the agent knows that.
Sensitive data redaction
Replacing identification numbers, account numbers and similar values with markers before they reach a record. Not the same as anonymisation, because a person can still be identifiable from context.
Kill switch
A switch that cuts off a channel or an operation immediately. It acts in the execution layer, so it cannot be bypassed from above.
Quality supervision
Automated answer scoring running beside the critical path so it never delays the conversation. The verdict is stored with the conversation and never reaches the customer, and the panel screen for reading it is still being built.
Closed tool catalogue
A finite list of operations the agent can trigger at all. What is not on the list does not exist from the agent's point of view.
Deflection rate
The share of conversations closed without a human. On its own it can mislead, because a conversation abandoned by an impatient customer also counts as closed.
First contact resolution (FCR)
The share of cases settled on the first attempt, without a repeat contact. Closer to customer experience than deflection alone.
Average handling time (AHT)
The average time one ticket takes. Useful for calculating team relief, less so for judging quality.
Customer satisfaction (CSAT)
A rating given by the customer after a conversation. A simple measure, burdened by the fact that only a minority respond.
Service level agreement (SLA)
Agreed availability and response parameters written into a contract. Without the contractual clause it is a statement, not an obligation.
GDPR
The European regulation on personal data protection. It sets out, among other things, the information duty, legal bases for processing and the rights of data subjects.
Controller
The entity deciding the purposes and means of processing. For conversations with our customer's customers, that customer is the controller, not us.
Processor
An entity processing data on the controller's instruction. AIvatary acts in that role towards the content of conversations with our customers' customers.
Data processing agreement (DPA)
The agreement required by Article 28 GDPR, describing the scope, duration and rules of processing entrusted to a processor.
Subprocessor
A further processor engaged by a processor. It requires the controller's consent and should be named individually.
Standard contractual clauses (SCC)
A contract template approved by the European Commission that allows transfers of data outside the European Economic Area.
AI Act
The European regulation governing AI systems. Article 50 requires, among other things, telling a person that they are speaking to a system rather than a human.
Multi-tenancy
An architecture where one instance serves many organisations. What matters is the level at which separation of their data is enforced.
Retention
How long data is kept and when it is deleted. It should follow from the purpose of processing, not from disk capacity.
Where these definitions come from
The definitions describe how we use these terms in AIvatary and how the industry understands them. Where something is understood differently by different people, we say so rather than picking the more convenient reading.
A term is missing
Tell us what you could not find. The glossary grows from questions that actually get asked.