A rule of thumb before the definitions: vendor fog is a signal. Every term below has a plain-English meaning, and a salesperson who won't give it to you plainly is usually hiding a meter, a cloud, or a permission you wouldn't grant if you understood it.
A
Agent
An AI that doesn't just answer questions but takes multi-step actions with tools: searching a database, building a file, updating a record, running on a schedule. The word matters when buying because an agent acts, and anything that acts needs guardrails you can inspect. When a vendor says agent, your next two questions are what it can touch and what stops it.
AI chief of staff
An AI that holds a whole desk instead of answering one question at a time: permanent memory of your clients, fees, and promises, a daily cadence of briefs and follow-up nudges, hands on your real tools, and accountability that names the call you've been avoiding. It's distinct from an assistant, which waits to be asked, and from the generalist AI employee category, which automates broad business tasks without knowing your trade. RecruiterClaw is this, built specifically for recruiters.
Additive-only writes
A safety design where an AI connected to your ATS or CRM can create records, add notes, add candidates to a pipeline, and update fields, while delete, merge, and archive are impossible by design: the tool never even requests those permissions. It's the difference between a tool that promises to be careful with your database and one that structurally cannot wreck it. Twenty years of candidate history deserves the second kind.
Answer-first content
Web content built so the opening paragraph fully answers the question in the title, because that's the paragraph AI assistants and search snippets quote. It matters to recruiters as marketers: when a hiring manager asks an AI tool who recruits superintendents in his market, the firms with answer-first pages are the ones that get cited.
C
Contact enrichment
Turning a name and a profile into verified ways to reach the person: email plus direct and mobile phone numbers. Verified contact data costs real money, so the buying question is control. Enrichment you trigger on candidates you picked, like telling RecruiterClaw "enrich the top 5" after reviewing a shortlist, beats enrichment that runs on everyone and bills you for people you were never going to call.
Credit metering
A pricing model where every task draws down a purchased balance, so the meter is always running. On a desk it changes behavior: you hesitate to run one more search on a hunch, and a heavy week costs more than a slow one. The honest exception is contact data, which genuinely costs vendors money. Ask any vendor which parts are metered and which are flat, and get it in writing.
F to H
Frontier model
One of the handful of most capable AI models available at any given time, built by the major AI labs. Most serious AI products, recruiting-specific or not, run on the same class of frontier models under the hood. That means the real differences between tools live in what's built around the model: data, memory, tools, and guardrails, not the engine itself.
Hallucination
When an AI states something false with total confidence, in fluent, plausible prose. The fix in a buying context is grounding: good tools work from real data, like actual professional profiles, your own ATS records, or a document you dropped in, rather than generating facts from thin air. The working rule on a desk stays the same either way: nothing AI-drafted goes to a client or candidate unread.
Human-in-the-loop
A design where the AI prepares the work and a person approves the consequential step: the email doesn't send and the record doesn't change until you sign off. Look for mechanisms, not promises. In RecruiterClaw, for example, sending an email requires you to type the exact confirmation phrase, so nothing ever sends on its own.
L to M
LLM (large language model)
The engine underneath modern AI tools: a model trained on enormous amounts of text that can read, reason over, and produce language. When you chat with any current AI product, an LLM is doing the reading and writing. The product wrapped around it decides what the model can see, remember, and do, which is where your evaluation should focus.
MPC (Most Placeable Candidate)
A recruiting term far older than AI: a standout candidate you proactively market to companies instead of waiting for a matching job order. AI changed the economics of the play. RecruiterClaw's MPC Builder takes one dropped resume and returns 5-10 companies with live matching openings pulled from their own career pages, the right hiring manager at each with verified email and phone, and an anonymous 3-step outreach sequence, all in one CSV.
O to P
On-device OCR
OCR is the technology that reads text out of images and scans: a photographed resume, a scanned fee agreement, an iPhone photo of a business card. On-device means it runs on your own machine, so the document never leaves it. When the documents in question are resumes and signed agreements, where that processing happens is a privacy question, not a technical footnote. RecruiterClaw runs OCR on-device.
Plain-English search
Describing who you want the way a client would say it, "superintendents in Dallas, ground-up commercial, currently at a GC," and letting the AI build the structured search. It rewards knowing your market instead of knowing operator syntax, and refinement becomes conversation: "drop the estimators," "tighten to 30 miles."
Prompt injection
A trick where someone hides instructions inside content your AI reads, like a resume, an email, or a web page, hoping the AI treats them as commands from you. You won't stop strangers from trying it, so the defense is structural: choose tools where consequential actions either require your explicit sign-off or are impossible by design. A poisoned document can't send your email if nothing sends without your confirmation phrase, and it can't delete your ATS records if delete was never a permission the tool holds.
How to use this vocabulary in a demo
These fourteen terms collapse into five demo questions. Does it come with candidate data? What's metered and what's flat? Where does my data live? What can it never do to my ATS? Who built it? RecruiterClaw's answers: 400M+ professional profiles included, flat rate with an allowance only on sourcing and enrichment, everything stored on your own machine, additive-only writes with delete impossible by design, and builders with 30+ years of combined placement experience. Hold every vendor, us included, to that standard of plainness.
What's the difference between an AI agent and an AI assistant?
An assistant answers when asked and stops there. An agent takes multi-step actions with tools: searching a database, updating a record, running on a schedule. An AI chief of staff is an agent plus three things a desk needs: permanent memory, a daily cadence, and accountability aimed at one specific recruiter's business.
Which term matters most when comparing prices?
Credit metering. It decides whether a heavy week costs more than a slow one and whether you hesitate before running one more search. Ask every vendor which parts are metered and which are flat. A fair answer meters only what costs the vendor real money, like verified contact data, and keeps everything else flat.
Which glossary terms are about safety?
Human-in-the-loop, additive-only writes, prompt injection, and on-device OCR. Together they answer one buying question: what happens on the tool's worst day? The strongest answers are structural. Sending requires your explicit sign-off, destructive database writes are impossible by design, and sensitive documents never leave your machine.
Do I need to understand LLMs to buy recruiting AI?
No. Most serious AI products run on the same class of frontier models, so the model is rarely the differentiator. Judge the layer built around it instead: what data it can search, what it remembers, what tools it can touch, what it structurally cannot do, and whether the people who built it have run a desk.
What questions should a recruiter take into an AI demo?
Five, straight from this glossary: does it come with candidate data or just advice about candidates, what is credit-metered versus flat, where does my data live, what can it never do to my ATS, and who built it. If a vendor can't answer those plainly, the fog is the answer.