Glossary of AI
in customer service.

AI support during a customer service employee’s work, for example with answers, summaries or work instructions. The employee conducts the customer conversation and assesses what to use.

The average time needed to handle a contact. Depending on the measurement setup, conversation time, time on hold and after-call work may be included. Specify which components your report uses.

An application that uses an AI model to choose next steps and use tools to achieve a goal. The permitted actions and boundaries are configured in advance.

The allocation of responsibilities, decision-making and controls around AI. This includes, among other things, who may modify an application, how performance is assessed and who follows up on incidents.

An agreed way for software to communicate with other software. An API can make data available or enable an action, within the permissions granted to the connection.

Automatic assessment of customer interactions against established quality criteria. AI can help assess conversations on a larger scale; people remain necessary to check the criteria and outcomes.

The share of an agreed set of tasks or contacts that is handled automatically. The outcome depends on what you count as a task, completion and human involvement.

An application that allows someone to have a conversation, usually via text. The answers may come from fixed routes, AI or a combination of both. A chatbot can also start connected workflows.

The share of bot contacts that, within the chosen measurement definition, remain without being handed over to an employee. This is not automatically the same as a resolved problem: someone may also drop out.

The information an application takes into account when generating an answer, such as previous messages, retrieved knowledge and relevant customer data. A language model can process only a limited amount at a time.

A collective term for AI techniques that enable systems to have a conversation via text or speech. The term itself says nothing about system access, autonomy or the quality of the answers.

Managing customer relationships and, in software form, recording customer data and interactions. A CRM can provide context for customer service, provided the right information and access rights are available.

A measure of customer satisfaction with an experience, product or service. It is often the ratio of satisfied respondents to all respondents. The wording, scale and timing of the measurement determine how you interpret the score.

Passing an issue on because additional expertise, authority or priority is needed. An escalation can originate from a bot, but can also take place between employees or teams.

The proportion of questions resolved during the first contact. Define how you determine that something has been resolved and within what period a follow-up contact still counts as relating to the same question.

A recurring process in which you collect and assess feedback, turn it into an adjustment and check the result again. Receiving feedback does not mean that an AI model automatically retrains itself.

AI that generates new content, such as text, based on learned patterns and provided input. A fluently worded answer does not necessarily have to be factually correct.

Supporting or constraining model output with information relevant to the task. This can be done, for example, through retrieved documentation or verified system data. The result still needs to be assessed.

AI output that is incorrect, fabricated or insufficiently supported, even though it may sound convincing. Better sources and checks can reduce this risk, but do not guarantee error-free answers.

Transferring a conversation or task from an automated application to a human. A useful handover includes the relevant context and a clear reason for taking over.

What someone is trying to achieve with a message or statement. The same intention can be expressed using different words; a message can also contain multiple intentions.

An organized collection of information, such as articles, terms and conditions, and work instructions. Its usefulness depends on being up to date, accessible, and clear to the target audience.

Organizing the entire knowledge lifecycle: creating, reviewing, publishing, updating, and removing information. This includes owners, versions, and agreements about who may use which information.

A large language model that has learned patterns in language and can process and generate text. It is one component of an application, not automatically a complete customer service solution.

The ability to process questions and answers in multiple languages. Language support does not yet indicate whether local terms and conditions, product names, and nuances are represented correctly in every language.

Techniques for deriving meaning, context, and intent from human language. NLU can, for example, help identify a reason for contacting an organization; it is different from merely recognizing words literally.

A score based on the willingness to recommend an organization. The percentage of respondents giving a 9 or 10 minus the percentage giving a 0 through 6 makes up the NPS.

An approach in which channels and relevant customer context are connected. Being reachable only by phone, chat, and email is not in itself proof that those channels also work cohesively.

The task and information used to direct a model. A prompt can describe style, purpose, and limitations, but it does not replace access rights or technical controls for a system action.

An approach in which relevant information is retrieved first and then provided to a language model to formulate the answer. The quality of the information found affects the usefulness of the result.

An automated assessment of the emotional tone in text or speech, such as positive or negative. Context and irony can make interpretation difficult; treat the result as a signal.

An agreement on the level of service to be provided, with definitions and responsibilities. Think availability, response time or resolution. A goal is only measurable if the measurement method is clear.

A piece of text or other model input that a model processes. A token can be part of a word, a word or a punctuation mark. The number of tokens therefore does not equal the number of words.

A model can select an available function and suggest arguments. The application validates and executes that request. Only the actual execution determines whether information is retrieved or a change is made.

An application that enables someone to communicate by voice. Speech recognition, conversation processing and speech synthesis are often combined. Call transfers and handling unclear speech must be deliberately configured.

An agreed sequence of steps, conditions and actions. A workflow can combine fixed rules, AI components and human approval. Not every automated workflow is therefore an autonomous AI agent.