PSM-AI key terms

A model trained on very large text collections to predict and generate language. It produces responses by estimating plausible word sequences rather than by looking facts up in a verified database.

The small chunk of text, often a word fragment, that a language model reads and generates. Token counts set context limits and drive usage cost, and they do not match word counts.

The input a person gives a generative AI tool: instructions, background context, examples, or constraints. Its clarity and specificity strongly shape the quality and consistency of the response.

A confident but incorrect or invented output, such as a fabricated citation or a non-existent function. It occurs because the model generates plausible text rather than verifying facts against a source.

The maximum amount of text, measured in tokens, that a model can consider in one interaction. Material beyond that limit is ignored unless it is summarized, retrieved, or supplied in a new request.

The examples used to build a model's behaviour. Its gaps, biases, and knowledge cutoff date shape what the tool knows, how fairly it performs, and where it will fail.

Additional training of an existing model on a smaller, targeted dataset so it better fits a task, tone, or domain. It changes the model itself, unlike prompting, which only changes the input.

A pattern where a tool first searches an approved knowledge source, then has the model answer using those retrieved passages. It reduces invention and keeps answers tied to current, organization-owned material.

The tendency of the same prompt to produce different outputs on different runs. Teams must plan for variability, so a single good result is not evidence of reliable behaviour.

Systems that can plan several steps and call tools or take actions with limited human involvement, instead of only answering one prompt. Greater autonomy increases the need for explicit limits and oversight.

The practice of designing and repeatedly refining prompts, with clear goals, context, constraints, and examples, to obtain more useful and consistent output from a generative tool.

Pre-set instructions that define a tool's role, tone, and boundaries before a conversation begins. They apply to every turn and shape responses differently from a one-off user request.

Including a handful of worked examples in a prompt so the model imitates the desired format, tone, or reasoning pattern. It is often the quickest way to stabilize output.

Asking a model to work through intermediate steps before presenting a final answer. It can improve multi-step accuracy but lengthens output, raises cost, and exposes reasoning that may be wrong.

Technical or procedural limits on an AI tool, such as content filters, approved data sources, or required approvals. They constrain what the system will do and route exceptions to people.

A deliberate checkpoint where a person reviews, corrects, or approves AI output before it is used or acted on. It preserves accountability and catches errors the tool cannot detect.

A structured way to judge whether an AI tool suits a defined purpose, using test cases, rubrics, or comparison against a baseline. It replaces impressions of quality with evidence.

A team's shared working understanding of what AI tools can and cannot do, including their typical failure modes. It lets members judge output critically instead of accepting it.

Explicit team norms covering where AI may be used, what must be disclosed, which data is off-limits, and who reviews output. The team revisits it as it learns.

A climate in which team members can try unfamiliar tools, admit mistakes, and share weak results without embarrassment. It is a precondition for honest learning about what actually helps.

Openly telling colleagues, customers, or stakeholders when content or decisions involved AI, and to what extent. Disclosure lets others apply the scrutiny they consider appropriate.

Accepting AI output without adequate checking, so the team gradually loses its own judgement. It shows up as unchallenged summaries, thin reviews, and unexamined assumptions.

A shared, versioned collection of prompts the team has found useful, with notes on context and results. It spreads good practice and cuts repeated trial and error.

The team's checklist for when work is complete. Where AI contributes, teams often add checks for factual accuracy, sourcing, human review, and appropriate disclosure.

The policies, roles, and decision rights an organization uses to choose, use, and monitor AI. It clarifies who may approve a tool and who answers when something goes wrong.

An organizational document stating which AI tools are approved, what data may be entered, and which uses are prohibited. Teams need it in plain, practical form to make daily choices.

Sorting information into levels such as public, internal, confidential, or personal, so people know what may be shared with an external AI service. It converts vague caution into clear rules.

Reviewing a supplier's data handling, security, model update practices, and exit terms before adoption. It addresses the chance that a tool changes behaviour or retains information beyond the organization's control.

A time-boxed, narrowly scoped trial of an AI tool with agreed success measures and a decision point. It generates real evidence before the organization commits to wider adoption.

Moving an AI practice from one team's trial to broader organizational use, which requires updated policy, training, support, and evidence that the benefit still holds at larger scale.

The ongoing expense of AI tools, driven by subscriptions and by tokens processed, plus the human time spent reviewing and reworking output. It belongs in the business case.

Systematic skew in AI output that disadvantages certain groups or viewpoints, often inherited from training data and design choices. It persists unless teams deliberately test for it.

An attack in which hidden instructions inside content the model reads, such as a document or web page, redirect its behaviour. It is a security concern whenever tools process untrusted input.

Unintended exposure of sensitive information through an AI tool, for example pasting confidential content into an external service that stores or learns from it. Classification and clear rules reduce the risk.

Uncertainty about the ownership and provenance of AI-generated content, including whether it reproduces protected material. Teams should check licensing, attribution, and policy before publishing.

The human tendency to trust a machine's suggestions more than the evidence warrants, especially under time pressure. It is the mechanism that turns convenient tools into unchecked decisions.

How far a person can understand why an AI system produced a particular output. Low explainability makes it hard to challenge results or defend decisions to stakeholders and regulators.

The principle that a named person or role stays answerable for decisions and deliverables, whatever share of the work AI contributed. Systems cannot hold accountability or consent to being governed.

Decline in a tool's accuracy or usefulness over time as the world, data, or underlying model changes. It argues for periodic re-evaluation rather than one-time approval.

Legally binding rules that classify AI uses by risk and impose duties such as disclosure, documentation, testing, or outright prohibition. Scrum Masters should know which uses in their context fall into higher-risk categories.