The automation glossary, in plain English.

An AI system that can take actions toward a goal, not just answer a single question. An AI agent uses a model to decide and carry out steps - calling tools, reading data, escalating when unsure - in pursuit of an objective. In business operations, a well-built agent has clear boundaries, grounding in your data, and a path to hand off to a human on low-confidence cases.

See: Respond to GoHighLevel leads instantly with an AI first-touch.

A defined way for one piece of software to talk to another programmatically. An API lets systems exchange data and trigger actions through a stable contract, rather than through a screen a human reads. Integrating at the API layer is more reliable than screen-scraping because the contract changes less often than the interface. Most modern SaaS tools expose an API for exactly this purpose.

Handling work in scheduled groups rather than one item at a time. Batch processing collects work and runs it together on a schedule - nightly reconciliation, a weekly report. It's efficient for work that doesn't need to be instant, and contrasts with real-time processing.

Automating whole end-to-end business processes rather than isolated tasks. BPA targets a complete process - order-to-cash, employee onboarding, month-end close - rather than a single step. It usually spans multiple systems and departments and includes the exception handling and approvals a real process needs. It is the broader discipline that individual workflow automations serve.

The rate at which customers stop doing business with you. Churn measures lost customers over a period. Automation touches it indirectly - faster response, consistent follow-up, and reliable service all reduce the friction that drives customers away.

Data trapped in one system, disconnected from the rest of the business. A data silo is information locked inside a single tool or team, invisible to others who need it. Silos force manual copying between systems and prevent a unified view; breaking them is a core aim of integration work.

See: Systems that don't talk to each other.

Moving data from source systems into a destination, cleaning it along the way. ETL pulls data out of one or more systems (extract), reshapes and cleans it (transform), and writes it somewhere useful like a warehouse or report (load). It's the backbone of automated reporting and analytics pipelines.

Deciding what an automation does when it hits a case it can't handle confidently. Exception handling routes the messy cases - an unfamiliar document, a low-confidence result, a mismatch - to a human or a defined fallback instead of forcing a guess through. It's the feature that separates automations that survive production from ones that quietly corrupt data.

See: Re-keying the same data between tools.

Keeping a person reviewing an automation's output where judgment is needed. Human-in-the-loop design has a person check or approve results at the points that need judgment, while the automation handles the repetitive work. Over time, as trust grows, that review can shift from gatekeeping to spot-checking.

A property that means running the same operation twice has the same effect as running it once. Idempotency protects automations from duplicate processing - if a webhook fires twice or a job retries, the outcome stays correct. It's a small technical detail with big consequences for data integrity in unattended systems.

The type of AI model that understands and generates human language. An LLM is trained on large amounts of text to predict and produce language, which lets it summarize, classify, extract, and draft. In operations it's the engine behind document understanding, drafting replies, and routing - but it needs guardrails and grounding to be reliable on real data.

Tools that are mostly visual but let developers drop into code for the hard parts. Low-code sits between no-code and full custom development: a visual builder for speed, with escape hatches to write real code where the logic demands it. It suits teams that want to move quickly but still need to handle non-standard cases properly.

Tools that let people build automations or apps through a visual interface, without writing code. No-code platforms use drag-and-drop builders and pre-made blocks so non-developers can assemble workflows. They're fast for standard, well-behaved processes. They can get brittle or expensive when a workflow grows complex, high-volume, or needs logic the platform wasn't designed for.

Technology that converts an image of text into machine-readable text. OCR turns a scanned page or photo into characters a computer can process. It's often the first step in a document pipeline, with Document AI or an LLM interpreting the extracted text afterward.

Coordinating multiple automated steps or services into one reliable process. Orchestration manages the order, timing, and dependencies of the steps in a workflow - including what happens when one fails. It's what turns a collection of individual automations into a dependable end-to-end process.

A technique that grounds an AI model's answers in your own documents instead of only its training data. RAG retrieves relevant passages from your knowledge base and feeds them to the model as context, so answers are based on your actual policies, products, or records. It's how an AI support or knowledge assistant stays accurate to your business rather than making things up.

Handling each event the moment it happens. Real-time processing acts on data as it arrives rather than waiting for a scheduled batch. It's the right choice when latency matters - routing a hot lead, alerting on a failed payment - and typically relies on webhooks or streaming.

Automatically re-attempting a step that failed, instead of giving up. Retry logic handles the transient failures every integration hits - a rate limit, a brief outage - by trying again sensibly rather than dropping the work. Combined with alerting, it's the difference between a resilient automation and one that fails silently.

Software 'bots' that mimic a human clicking through an application's user interface. RPA automates by driving the screen - moving the mouse, typing, reading fields - rather than calling an API. It's useful for legacy systems with no API, but it's fragile: a form redesign or layout change can break a bot. API-level integration is more durable where it's available.

See: Automations that break every time something changes.

One agreed place where a piece of data lives, so every team sees the same number. A single source of truth means a metric or record is defined and stored once, and everything else reads from it. Without one, teams reconcile conflicting numbers by hand and 'the same KPI' means different things on different dashboards.

See: Month-end close stuck in spreadsheets.

A committed standard for how a service will perform, such as response time. An SLA sets measurable expectations - uptime, response time, resolution time - that a provider commits to. For automations, monitoring against an SLA is how you know a system is actually meeting its promise rather than silently degrading.

Data organised in a defined format, like rows and columns. Structured data lives in a consistent shape - fields, types, tables - that software can process directly. Turning unstructured inputs (emails, PDFs) into structured data is often the first step that makes automation possible.

The single authoritative source for a given type of data. A system of record is the tool your business treats as the truth for something - the CRM for customers, the accounting system for finances. Clear systems of record prevent the 'which spreadsheet is real?' problem and make integrations far easier to reason about.

The future cost of quick or brittle solutions that will need reworking. Technical debt accrues when systems are built fast or patched rather than built durably - brittle scripts, undocumented workarounds, tools bent out of shape. It shows up later as breakages and maintenance that crowd out new work.

See: Automations that break every time something changes.

The event that starts an automation. A trigger is the condition that kicks off a workflow - a new email, a form submission, a scheduled time, a status change. Every automation begins with one, and choosing the right trigger is half of designing a reliable workflow.

Information without a predefined format - free text, images, documents. Unstructured data is the messy majority of business information: emails, scanned invoices, chat logs. Extracting structure from it (via Document AI, OCR, or an LLM) is what lets it flow into automated systems.

An automated message a system sends to another the moment an event happens. A webhook is a 'push' notification between systems: when something occurs (a form is submitted, a payment clears), the source system immediately calls a URL you control with the event data. It's the real-time alternative to repeatedly polling an API to ask 'has anything changed yet?'.

Software that runs a multi-step business process automatically instead of a person doing each step by hand. Workflow automation chains together the steps of a recurring process - triggers, actions, and conditions - so it runs without manual effort. It covers everything from a single app-to-app trigger to a branching pipeline across many systems. The goal is to remove repetitive human handling from work that follows consistent rules.

See: Re-keying the same data between tools.