| Repetitiveness |
How often work repeats and how similar its form is. A higher repetitiveness score favors automation. |
| Risk |
The harm caused by an error, broken into reach of impact, difficulty of reversal and time to detection. |
| Review cost |
Human time and effort needed to check results, compared with the original work time. |
| Exclusion criteria |
Conditions excluding automatic processing regardless of scores, such as consequential decisions without human review. |
| Trigger |
The event or time that starts work; determines scheduled versus event-triggered execution. |
| Completion condition |
A verifiable criterion establishing that a step has finished; the starting point for review criteria. |
| Processing time |
Time spent actually doing the work; the time AI mainly reduces. |
| Waiting time |
Time waiting for the next owner to start; reduced by changing procedures and assignment methods. |
| Normal and exception paths |
The route most cases take and routes for missing information, conflicts or overdue cases. |
| Delegation ladder |
Four levels of delegation to AI: organize and recommend, draft for review, execute after approval, limited automatic execution. |
| Human review point |
A location in the automation flow where a person checks or approves; placed before high-impact actions. |
| Approval |
Permission to perform an action; distinct from verifying content. |
| Approval fatigue |
Approving without reading because there are too many approval requests. |
| Rubber-stamp review |
Review points exist, but reviewers approve formally without catching errors. |
| Workflow |
A system where AI and tools follow predefined paths. |
| Agent |
An AI system that receives a goal, determines its own process and tool use, and handles multiple steps. |
| Automation design sheet |
A one-page document with purpose, step table, delegation levels, review points, excluded work, stop conditions and success criteria. |
| Knowledge space (project) |
A workspace grouping related conversations, files and instructions. |
| Shared project |
A knowledge space where multiple members accumulate conversations and materials together. |
| Custom assistant |
AI combining instructions and reference material for many people to use consistently, such as a GPT or Gem. |
| Instructions |
Behavior rules shared across conversations, defining role, procedure, format and exception handling. |
| Personal custom instructions |
Personal settings applying generally to one user's conversations. |
| Project instructions |
Instructions applying only to conversations in a particular project. |
| Knowledge |
Reference material for answers; relevant parts are retrieved as the collection grows. |
| Skills / skills |
Reusable instruction bundles loaded when needed. Anthropic Skills are folders of instructions, scripts and materials; Google is moving Gems to skills. |
| Plugins |
OpenAI packages bundling apps, skills or both into a workflow; recommended as the migration destination for custom GPTs. |
| RAG (retrieval-augmented generation) |
Retrieving material chunks relevant to a question and using them in the answer. |
| Context limit |
The maximum amount a model can read and process at once. |
| Connected source / uploaded copy |
Material linked to an original that may reflect changes, versus a copy fixed at upload time. Update behavior differs by product. |
| Materials register |
A management table recording material name, authority, effective date, owner, connected status and review date. |
| Authority level |
How official a source is; approved policies, departmental notices and personal notes carry different weight. |
| Test question set |
A fixed validation collection pairing normal, exception, refusal and boundary questions with expected behavior. |
| Boundary case |
A question prone to failure, such as conflicting sources, recent revisions or situations outside policy. |
| Project-only memory |
Memory that does not reference personal memories or conversations outside the project; applied automatically to ChatGPT shared projects. |
| Connector |
A feature linking AI to external services to find information and act; called an app in some products. |
| Permission inheritance |
A connector does not exceed the connected person's source-system permissions. Granted scopes and organizational settings may narrow access further. |
| Shared credentials |
Fixed authentication information such as an API key used by multiple users; its scope is available to all. |
| Authentication / access permissions / execution approval |
Three control layers determining whose account, what it can reach and whether this action may proceed. |
| MCP (Model Context Protocol) |
An open-source standard connecting AI apps to data, tools and workflows. |
| MCP host / client / server |
The user's AI app, the component maintaining one server connection, and the program providing context and capabilities. |
| Tools / resources / prompts |
Executable functions, contextual material and instruction templates provided by MCP servers. |
| Least privilege |
Granting only the actions and scope needed for the task. |
| Excessive Agency |
An OWASP risk where excessive functionality, permissions or autonomy enable harmful actions. |
| Tool permissions |
Per-tool connector settings such as allow, require approval and block. |
| Agent task specification |
A task instruction defining deliverable, completion criteria, scope, prohibited actions, stop conditions and reporting. |
| Scheduled task |
A task that repeats at specified times. |
| Event task |
A task triggered by a connected-app event, such as new email. |
| Prompt injection |
An attack where instructions in inputs or external materials unintentionally change AI behavior. |
| Indirect prompt injection |
Injection through hidden instructions in external material such as webpages, emails and files. |
| Idempotency |
Repeated execution of the same operation has the same result as one execution. |
| Permission matrix |
A table of who may do what to which materials, across people, materials and actions. |
| Sharing unit |
The scope of what is shared, such as a conversation link, project, notebook, task link or exported file. |
| Role-based access control (RBAC) |
Managing feature permissions according to roles. |
| Data retention |
How long conversations and files are stored in a service; separate from training use. |
| Training use |
Whether user data is used to improve models. |
| Temporary chat |
A conversation that does not enter history, create memory or train models while temporary. ChatGPT may retain it for up to 30 days for safety; saving it as a normal chat applies account settings. |
| Pseudonymization |
Deleting or replacing personal information so an individual cannot be identified without additional information (PIPA, Article 2). |
| Derived information |
AI-generated summaries, comparisons and quotations based on originals and containing their content. |
| Audit log |
A record of who changed settings or data and when. |
| Review checklist |
Items to inspect in results, with categories chosen according to work risk. |
| Full review / sample review |
Checking every case versus checking a randomly selected subset. |
| Error severity |
Classification of errors as minor, significant or critical. |
| Metric definition sheet |
A document specifying formula, denominator, inclusions/exclusions, collection and owner. |
| Rubric |
A table specifying grading criteria item by item. |
| Regression |
Previously correct behavior becoming incorrect after a change. |
| Staged rollout |
Applying a change to selected users before expanding to everyone. |
| Baseline |
A comparison value measured the same way before adoption or change. |
| Net savings |
Actual savings: reduction in total time per case multiplied by volume, minus maintenance time. |
| Manual fallback |
Stopping automation and having people take over through a defined procedure. |