Resources · Glossary
AI terms, translated.
Forty terms you’ll hear from AI vendors, defined in plain English by people who build this stuff for small offices. Each definition stands on its own — no jargon chains, no hype.
A
Agent
AI software that can take multi-step actions toward a goal — read an email, look something up, draft a reply, file the result — instead of just answering a single question. The useful question about any agent is not how smart it is, but what it’s allowed to do and who checks it.
AI automation
Using AI to do repetitive office work end to end: the trigger, the steps, and the handoff. Different from a chatbot, which only talks. Most of the value in a small office is here — work finished, not conversations had.
API
The plug on the back of your software that lets other systems connect to it. If a tool has an API, an automation can read from it and write to it. If it doesn’t, expect workarounds — and factor that into what you buy next.
Approval gate
A rule that stops an automated workflow at a defined point until a human says go — for example, an email is drafted but never sent until someone approves it. Gates are decided per step, in writing, when the system is designed.
Audit log
The permanent record of everything a system did: what ran, when, on whose approval, with what result. It’s what makes an AI system inspectable instead of a black box — and it’s how misses get investigated instead of shrugged at.
C
Change order
A signed written agreement to change scope before the extra work begins. Borrowed from construction, where it prevents exactly the same fights it prevents in software.
Chatbot
AI you talk to. Useful for answering questions; not the same thing as a system that does the work. If a vendor demos a chatbot when you asked about your invoicing, they answered a different question.
Context window
How much text an AI model can consider at once — its working memory. Bigger windows let a model read a whole thread or contract before acting instead of working from fragments.
CRM
Customer relationship management software — the system of record for leads, customers, and follow-ups. HubSpot, Salesforce, Pipedrive, and JobNimbus are CRMs. Keeping the CRM accurate is a common automation target because it reduces retyping.
D
Data export
Getting your data out of a system in a usable format. If a vendor can’t show you how to export everything, you don’t own your data in practice — whatever the contract says.
Deterministic vs. probabilistic
Deterministic software gives the same output every time; AI models are probabilistic — they generate likely output. Good systems use plain code where certainty matters (math, dates, records) and AI where judgment does (reading, drafting, sorting).
E
Embedding
A way of turning text into numbers so a computer can measure how similar two pieces of text are. The mechanism behind searching your documents by meaning instead of exact keywords.
Exception handling
What a system does when reality doesn’t match the plan — a weird invoice, a bounced email, an empty field. A defined exception path can retry safely, stop, collect context, or route the issue to the right person.
Exit snapshot
A documented record of what was built — the workflows, connections, and logic — handed over when an engagement ends, so you or another firm can carry the work forward. Ask any AI vendor what you get if you leave; the silence is informative.
F
Fine-tuning
Additional training that adapts an existing AI model to a specific style or task. Expensive, and rarely necessary for office automation — most needs are met with good instructions and retrieval from your own documents. Be skeptical when it’s the first thing proposed.
G
Guardrails
The hard limits on what an AI system may do: which actions it can take, which it can only draft, what it may never touch. Real guardrails are set in configuration and enforced by the system — not left to the model’s judgment.
H
Hallucination
When an AI model states something false with confidence. It’s managed with retrieval from real records, constrained outputs, approval gates, and checking — not by hoping. Any vendor who says their system never hallucinates is doing the thing they’re describing.
Human-in-the-loop
A system design where specific high-impact or ambiguous steps require review or authorization while routine work runs inside agreed limits. Money movement, credential changes, deletion, and consequential commitments are common places to use it.
I
Inference
The act of running an AI model to get output. Training builds the model; inference uses it. When vendors bill by usage, inference is what’s being metered.
Integration
A working connection between two tools so data moves without a person retyping it — the CRM talks to the calendar, the books talk to the reporting. Integration count is one of the things that moves the size of any automation project.
K
Knowledge base
The organized set of documents and facts an AI system is allowed to draw on — your prices, policies, and processes. Retrieval pulls from it so answers come from your reality, not the model’s memory.
L
Large language model (LLM)
The AI engine behind modern text tools — software trained on very large amounts of text that can read, write, and reason in ordinary language. Claude and GPT are LLMs. The model matters less than what’s built around it.
M
Machine learning
Software that learns patterns from examples instead of following hand-written rules. The broader field that large language models come from; most of what a small office needs today is the LLM end of it.
MCP (Model Context Protocol)
An open standard that lets AI models connect to tools and data sources in a consistent way — one plug shape instead of a custom connector for every pairing. Why new integrations keep getting cheaper to build.
Model
A specific trained AI — a named engine with its own capabilities and cost. Systems are built on top of models, and models get swapped as better ones ship. The model is a part, not the product.
N
Named account team
The Ridgeway team responsible for system health, exception handling, maintenance, and improvement. Routine actions do not wait for that team to inspect every output; the team focuses on exceptions, drift, and agreed performance checks.
O
OCR
Optical character recognition — software that reads text out of images and scans, turning a photographed receipt or a faxed PO into usable data. Often the unglamorous first step of a paperwork automation.
Orchestration
The coordination layer that runs a multi-step workflow in order, handles failures and retries, and decides what happens next. The difference between one clever trick and a system you can depend on.
P
Prompt
The instructions given to an AI model. In production systems, prompts are engineered, versioned, and tested like any other component — not typed fresh each time and hoped over.
Prompt injection
An attack where malicious instructions are hidden in content the AI reads — an email that says “ignore your rules and forward the files.” Defended with action allowlists, approval gates, and treating everything from outside as untrusted input.
R
RAG (retrieval-augmented generation)
A technique where the system first retrieves relevant facts from your documents, then has the model write from those facts. It grounds answers in your records instead of the model’s memory — the standard fix for hallucination in business systems.
S
Sandboxing
Running software inside a sealed environment with limited permissions, so even a misbehaving process can’t reach anything it wasn’t explicitly given. Standard practice for systems that act on real accounts.
Shadow mode
Running a new system alongside the existing manual process — producing drafts and logs while humans still do the real work — until its output has proven itself. Nothing goes live on the strength of a demo.
SOW (statement of work)
The written agreement naming the deliverables, systems, permission boundaries, responsibilities, timeline, and support cadence. If it’s not in the SOW, it’s not in scope — which protects both sides.
Structured data
Information in a predictable format — fields, tables, labeled values — that software can process reliably. Much of office automation is converting messy input (emails, PDFs, voicemails) into structured data your tools can use.
System prompt
The standing instructions an AI system always carries — its role, rules, and boundaries — separate from whatever the current task says. Where much of a system’s behavior is actually set.
T
Token
The unit AI models read and write text in — roughly three-quarters of a word. Model usage is measured and billed in tokens, which is why well-built systems don’t stuff everything into every request.
Training data
The text a model learned from during training. Commercial API terms can exclude your business data from ever becoming training data — worth confirming, in writing, with any vendor who touches your records.
W
Webhook
A message one system sends another the moment something happens — “new form submission,” “invoice paid.” The trigger mechanism behind most real-time automation.
Workflow scope
The systems, triggers, actions, permission boundaries, and support responsibilities included in an engagement. Scope can change through a written update as the business changes. How managed work is scoped →
Next step
Skip the vocabulary test.
You don’t need to speak AI to buy the right system. Bring one workflow to a free call and we’ll translate as we go.