Agentic AI glossary for business leaders

An activity is a type of step in a process, such as “check invoice” or “approve payment”. Each event in an event log records one activity, carried out for one case at one point in time.

Agentic AI is AI that can plan and carry out multi-step work toward a goal, using tools and acting with limited supervision. It goes beyond answering a question: it chooses steps, takes them and checks the result, within limits that people set.

See also: How to build AI agents for business

An agentic workforce is one in which AI agents carry out defined tasks alongside people, while people set goals and oversee the work. Via (transformation planning) is built to plan the move to an agentic workforce.

See also: What is an agentic workforce, Via agentic workforce transformation

An AI agent is software that uses a large language model to reason, plan and carry out tasks through tools, with limited supervision and within limits that people set. Without tools, an agent is only a chatbot.

See also: How to build AI agents for business

AI agent governance is the set of rules, roles and reviews that decide what each AI agent may do, which data it may use, which of its actions need a person's approval, and who answers for its results.

See also: Safe AI agents: a governance guide

AI value realization is the practice of stating the expected benefit of each AI initiative before it starts, then comparing it with the measured result afterwards, so that only the benefits that actually arrive are counted.

See also: AI value realization guide

An audit trail is a lasting record of what an AI agent did: the inputs it received, the actions it took, the approvals given, and every change to its permissions or instructions. It lets people trace and explain each action afterwards.

See also: Safe AI agents: a governance guide

A baseline is a measure of how work performs before a change, such as its volume, the time per task, its cost and its error rate. Without an agreed baseline, nobody can show later what an AI initiative changed.

See also: AI value realization guide

A bottleneck is a step in a process where work waits or piles up because the step cannot keep pace with the work arriving. A bottleneck sets the pace of the whole process, so it is often the first place to improve.

See also: Process bottleneck analysis

A business case is the comparison of what a change will cost with the savings and other benefits it is expected to bring over time, often summarized as a return on investment and a payback period.

See also: AI value realization guide

A case is one run of a process from start to finish, such as one order, one invoice or one insurance claim. Each case has its own ID, so that the events belonging to it can be grouped and followed.

Conformance checking is the comparison of real cases with a reference or approved process model, to find where and how cases deviate from it. It is not the same as measuring how many cases follow the most common path.

Current and future way of working means how a task is done today and how it is planned to be done after a change, including who performs it: a person or an AI agent. Comparing the two is how a change is planned and priced.

See also: Agentic workforce planning

Cycle time is the time a piece of work takes between two points in a process, such as from one step to the next. Long or uneven cycle times point to waiting, handoffs or rework that are worth investigating.

A device privacy policy is the policy an administrator publishes for Opus (task mining); without it, Opus collects no activity data. The policy also lists the applications and websites to exclude, and the text patterns to mask on the PC before upload.

See also: Task mining privacy

Digital interaction intelligence (DII) is software that records how people work in desktop applications and rebuilds their tasks and processes from that record. In Optimus Hive, Opus (task mining) is the product for it.

See also: Digital interaction intelligence

An event is a single recorded step in a business system, such as “invoice approved”, together with the case it belongs to and the time it happened. Events are the raw material of process mining.

An event log is a list of system events, each with a case ID, an activity and a timestamp. Favus (process mining) turns an event log into a process map that shows how cases flow from step to step.

See also: Favus process mining software

Forecast savings are the savings an initiative is expected to bring, estimated before the change from the baseline and the planned way of working. They become realized savings only when measured results confirm them.

Guardrails are the limits built around an AI agent to keep it within its role: the actions it may not take, the data it may not use, the checks its outputs must pass and the points where it must hand over to a person.

See also: Safe AI agents: a governance guide

Human in the loop means that a person reviews or approves an AI agent's work at set points, before it takes effect. The higher the risk of an action, the earlier and the more often a person should be involved.

See also: Safe AI agents: a governance guide

Idle time is time in a process or task when no work is being done on it, for example while a case waits for an approval or for missing information. In process analysis, idle time points to delays, not to how hard people work.

An initiative is a defined piece of change work with its own goal, owner, budget and timeline, such as moving one group of tasks to AI agents. A transformation program is usually split into several initiatives.

See also: Transformation program governance

The most common path is the sequence of steps that the largest number of cases follow through a process; it is also called the happy path. Comparing other variants with it shows where cases take longer or costlier routes.

Personal data masking is the replacement of names, account numbers and other personal details with neutral placeholders before data is stored, shared or sent to an AI model. It is also called PII redaction.

See also: Safe AI agents: a governance guide

Process mining is a method that rebuilds end-to-end business processes from the event data your business systems record. Favus (process mining) builds the process map from three fields per event: a case ID, an activity and a timestamp.

See also: Favus process mining software, Task mining vs process mining

Realized savings are savings that have actually arrived after a change, measured against the baseline and recorded against the plan. Until savings are realized, they are only a forecast.

See also: AI value realization guide

Rework is work done again because the first attempt was wrong or incomplete, such as a case sent back to an earlier step for correction. Rework adds time and cost without adding value.

Task mining is the analysis of desktop activity to show how people actually carry out individual tasks. Opus (task mining) does this with a Windows agent and with task recordings that employees start themselves.

See also: Opus task mining software, Task mining vs process mining

Throughput time is the total time a case takes to pass through a process, from its first event to its last. It includes both the time spent working on the case and the time the case spends waiting.

Tool access is the set of systems and functions an AI agent may use to act, usually through secure APIs, each with its own permissions. An agent's tool access should be no wider than its tasks need.

See also: How to build AI agents for business

A variant is one of the different paths a case or task can take through a process, such as an order that skips an approval or one that loops back for correction. Many variants often point to unclear rules or frequent exceptions.