Skip to content
QRKTech

Resources · AI glossary

AI, in plain English.

The words that come up when you start putting AI to work, without the jargon. If a term on our site is unclear, it’s defined here.

A

Agentic AI

AI systems that pursue a goal across multiple steps, choosing which tools to use and when. They’re most useful when work needs judgment, not just a fixed script.

See alsoAI agentRPA (Robotic Process Automation)

AI agent

Software that uses a language model to plan and take actions with tools, such as reading records, writing documents, or sending messages, rather than only generating text.

See alsoDigital coworkerMCP (Model Context Protocol)

Air-gapped AI

AI that runs entirely on isolated infrastructure with no connection to outside networks, so data never leaves the environment.

See alsoPrivate AIOpen-weight model

Approval gate

A required human decision before an action proceeds. QRK coworkers can request approval through a Teams card tied to the exact version of the plan.

See alsoHuman-in-the-loop

C

Connector

The integration that lets a coworker use a business system: an MCP server, a command-line tool, an API client, or a tool built from recorded browser workflows.

See alsoMCP (Model Context Protocol)

D

Deny-first networking

A firewall posture where every outbound connection is blocked unless a rule explicitly allows it. QRK rules can scope access by program, host, port, HTTP method, and path.

See alsoLeast privilegeSandbox

Digital coworker

An AI agent deployed into a defined role, with its own workstation, scoped access to business systems, memory, schedules, and a human manager. Also called a digital worker or AI employee.

See alsoAI agentHuman-in-the-loop

Dry run

Validating an action without performing it, to catch errors before anything changes.

See alsoRead-back verification

H

Human-in-the-loop

A design where people approve, correct, or take over at defined points, such as before customer messages, financial actions, or record changes.

See alsoApproval gate

L

Least privilege

The security principle of giving every user, service, or agent the minimum access needed to do its work, and nothing more.

See alsoRole-scoped accessDeny-first networking

Long-term memory

Facts and observations an AI retains across sessions. QRK coworkers use memory banks with separate read and write permissions, personal and departmental.

See alsoDigital coworker

M

MCP (Model Context Protocol)

An open standard for connecting AI agents to tools and data. A server publishes a list of tools the agent can call. In a role-scoped design, each role sees only its own tools.

See alsoRole-scoped accessConnector

O

Open-weight model

A language model whose trained weights are publicly available, so it can be run on your own hardware or by a provider of your choice, with no lock-in to one vendor.

See alsoZero data retention (ZDR)Air-gapped AI

P

Private AI

AI deployed so that your data, prompts, and outputs stay under your control: on your infrastructure or with contractual zero data retention.

See alsoAir-gapped AIZero data retention (ZDR)

Prompt injection

An attack where text in an email, ticket, or document tries to instruct an AI to do something else. The defense is to treat content as data and confine the agent technically.

See alsoDeny-first networkingRole-scoped access

PSA (Professional Services Automation)

Software that service businesses, especially MSPs, use for tickets, projects, time, contracts, and billing. Examples include HaloPSA, ConnectWise, and Autotask.

See alsoRMM (Remote Monitoring & Management)

R

Read-back verification

After writing to a system, independently reading the record back to confirm the change actually landed as intended. A success response alone isn’t proof.

See alsoDry run

RMM (Remote Monitoring & Management)

Software IT providers use to monitor and manage endpoints and servers, such as N-able N-central or NinjaOne.

See alsoPSA (Professional Services Automation)

Role-scoped access

Granting an AI only the tools and permissions its job requires, enforced by the server rather than by instructions. A triage role can’t call project-management tools, even if asked.

See alsoLeast privilegeMCP (Model Context Protocol)

RPA (Robotic Process Automation)

Software bots that replay fixed sequences of clicks and keystrokes. RPA is fast for stable, rule-bound tasks, but brittle when screens or inputs change.

See alsoAgentic AI

S

Sandbox

An isolated computing environment with restricted network, file, and process access. Each QRK coworker runs in its own sandboxed workstation.

See alsoDeny-first networking

Skill

A written, versioned procedure that tells a coworker how to do a job: inputs, steps, approvals, verification, and helper scripts. Think of it as an SOP an AI can follow.

See alsoDigital coworker

Subagent

A short-lived helper agent that a lead coworker launches to gather evidence or do part of a task in parallel.

See alsoWork board

W

Work board

A durable, kanban-style queue where large jobs are split into one card per item and worked by dedicated worker lanes, with history, dependencies, and safe retries.

See alsoSubagent

Z

Zero data retention (ZDR)

A provider commitment not to store prompts or outputs beyond processing the request, and never to use them for training.

See alsoOpen-weight model

Next step

Rather talk it through?

We’re happy to explain any of this in the context of your business, with no buzzwords and no pressure.