AI Glossary — Practical Keywords for Sales
Definitions for terms you'll use in customer conversations, organized into 6 categories: 1. Fundamentals -> 2. Model makers -> 3. Cloud platforms -> 4. Latest engineering terms -> 5. Risk -> 6. Latest products. Especially in sensitive projects (e.g. My Number), the katakana terms in category 3 (Bedrock, region, etc.) always come up. Use the buttons above to jump between categories. Even just the bold one-liners will keep you in the conversation.
1. Fundamentals
The foundation for talking about AI at all. Start here.
Computer technology that can learn and make decisions like humans
Covers image recognition, voice recognition, text generation, prediction, and more. In today's business context, "AI" mostly means Generative AI and AI Agents.
AI that creates text, images, and code — ChatGPT is the familiar example
Traditional AI focused on classification and prediction. Generative AI is revolutionary because it creates new content: drafting, summarization, translation, code generation.
The engine inside Generative AI. GPT, Claude, Gemini are examples
Models trained on vast internet text. The model you choose changes cost, accuracy, and speed. Providers (2) build them; you call them via a cloud (3).
The smallest unit an AI processes. LLM cost is billed per token
Roughly 1 English word ≈ 1–2 tokens. LLMs bill on input + output tokens. Processing large volumes drives up cost, so token-reduction design is the core of cost control. Adsur has shown 75% reduction on real workloads.
The instruction text sent to an AI — quality directly drives output quality
How you ask matters enormously. Prompt engineering improves both accuracy and token efficiency — a key skill for enterprise AI.
The max amount of info an AI can read at once — its "short-term memory"
Each model has a limit (e.g. hundreds of thousands of tokens = hundreds of book pages). Exceed it and older content is "forgotten." For long documents or chat histories, balancing this limit against token cost matters.
A way to let AI "read" your internal documents — cheaper than fine-tuning
Relevant internal docs are searched, then handed to the AI. Turn manuals, policies, and past cases into a knowledge base for accurate answers. Greatly reduces hallucination (5).
Additional training on an existing LLM to specialize it for a task
Bakes business data into the model itself. Higher accuracy, but heavier cost, time, and retraining burden. Most business cases are well served by low-cost RAG first; fine-tuning is the option when RAG can't reach.
2. Model Makers (Providers)
The companies building the LLM "brains" and their flagship models.
The company behind ChatGPT. Model names like "GPT-4o", "GPT-5"
The spark of the GenAI boom. GPT models are highly general and the best-known brand. Usable via API, but enterprises mostly consume them through Microsoft's Azure OpenAI (3). "We want ChatGPT" usually means this family.
Builds Claude — strong on safety, long context, and coding. Adsur's primary model
Tiers: Opus (top performance) / Sonnet (balanced) / Haiku (fast, low-cost). Excellent at long-document work and code, and available on AWS Bedrock (3), making it a great fit for sensitive projects. Adsur's most-used model.
Google's flagship — handles text, image, audio, and video together
Tiers like Pro (performance) / Flash (fast, low-cost). Strong at multimodal and very long context. Used via Google Cloud's Vertex AI (3). A candidate when Google Workspace / YouTube integration is involved.
Freely published "open-weight" models you can host yourself
Meta's Llama is the flagship. You can host the model on your own servers/cloud, an option when data must never leave for an external API. The trade-off: you bear the operations and GPU cost. Choosing vs. commercial models (GPT/Claude/Gemini) is the debate.
3. Cloud Platforms & AI Services
Where AI runs — this layer decides environment, cost, and how sensitive data is handled.
Renting compute in a provider's data center vs. running servers in your own building
Cloud = rent AWS/Azure/Google infrastructure over the internet (low upfront cost, easy to scale). On-prem = servers in your own facility. "Cloud or on-prem?" is the basic question that sets where AI can run and where data lives.
The three dominant providers — the ones you meet most in enterprise and government deals
AWS (Amazon), Azure (Microsoft), Google Cloud (GCP) hold most of the market. Which cloud a customer uses shapes available AI services and the path forward. Sensitive workloads often run on AWS/Azure; Microsoft 365 shops pair well with Azure.
AWS managed service to call LLMs like Claude securely on AWS
AWS's "gateway to GenAI." Call Claude, Llama, Amazon Nova via API. Key point: data stays within your AWS environment and isn't used for training — a frequent pick for My Number, healthcare, and finance. If a customer says "Amazon's… Bedrock?", this is it.
Microsoft's AI platform — use GPT securely inside Azure
Azure OpenAI runs GPT models inside Microsoft's environment. Azure AI Foundry (formerly AI Studio) bundles model selection, RAG, agent building, and evaluation into one dev platform. Great fit for Microsoft 365 / Copilot shops. Note: Copilot (SaaS) and API use are separate contracts (6).
GCP's AI platform — the gateway to using Gemini in an enterprise
Vertex AI bundles Gemini (2) and other models with RAG, agents, and MLOps on Google Cloud. A candidate for Google Workspace / BigQuery integration. It appeared as the pre-existing environment in a government agency PoC.
Which country/area your data physically lives in — top concern for sensitive projects
Clouds run in regions worldwide. Choosing the "Tokyo region" keeps data in Japan. My Number and similar data often legally cannot leave the country, making region choice a prerequisite for the deal. Always confirm this.
"Use a finished product monthly" vs. "rent parts and build it yourself"
SaaS = a finished product used monthly over the internet (e.g. Microsoft 365). Managed = renting provider-run building blocks (e.g. Bedrock, Document Intelligence) to assemble your own system. "License Copilot and you get the GPT API too" is a myth — separate contracts. Matters for explaining cost.
4. GenAI Engineering Terms
New ways of building AI that emerged in 2024–2026. Adsur's core battlefield.
AI that autonomously executes multi-step tasks from an instruction
Not just "question → answer" but an autonomous loop of research → decide → run tools → verify. Used in automation, ops, and document review — Adsur's core solution area. "Agentic" is the buzzword for this autonomous quality.
How far AI proceeds to decide and act without step-by-step human input
Ranges from semi-autonomous (human approves key steps) to fully autonomous (no human in the loop). For business cases, start semi-autonomous — "AI proposes → human approves → execute" — it's safer and standard. Full autonomy spikes effort on exception handling.
The scaffolding that controls and constrains AI — structurally prevents cost blow-ups, misfires, and leaks
Not "build a better model" but "design an environment where AI runs safely, accurately, and cheaply." Proposed by OpenAI/Anthropic in 2025–2026. Adsur used it to hit 97.4% accuracy and 75% token reduction on Vex. Core of our in-house edge.
A standard "plug" for connecting AI agents to internal systems and tools
A standard proposed by Anthropic in 2024 for how AI connects to tools and data sources. Like a USB port — "support it once, swap connections easily." A major trend in agent development; pays off in automation with many integration points.
Splitting work across specialized AI agents, coordinated by a conductor
Divide labor across "researcher", "executor", "verifier" agents to raise accuracy and coverage. Orchestration is the coordination layer. Used for complex workflow automation; more robust than a single AI.
5. Risk & Quality Terms
How AI misbehaves — and how we contain it. The vocabulary to address customer fears.
When AI confidently generates false or fabricated information
AI can output plausible-sounding misinformation even when unsure. Mitigations: RAG (1) for grounded answers, mandatory citations, and rule-based validation. Adsur systems always attach source citations. The thing customers fear most.
AI repeating unintended actions until cost or processing spirals out of control
An autonomous agent enters a bad loop, hammering an API and inflating cost, or takes unintended actions. Mitigations: harness (4) action limits, run caps, and human approval gates. This design is your answer to "won't it run wild if we let AI do everything?"
Safety "rails" on AI input/output — stop forbidden or sensitive content
Enforce rules: reject inappropriate input, detect/mask sensitive data (PII, My Number), block dangerous output. Adsur's in-house Aegis AI is a gateway built on this idea, detecting 19 PII types to prevent leaks.
Crafted input that tries to bypass or hijack AI restrictions
Tricks like "ignore your previous instructions and…" try to strip guardrails. Essential to defend in AI that takes external input (chatbots). Defenses: input validation, privilege separation, output monitoring. A security question customers raise.
Design mistakes in volume or loops that balloon token use — and cost
Main causes: dumping long text, needless re-sends, agents spinning. Prevent with token-reduction design (1) and harness caps. Adsur quotes split "development" vs "production (usage-based)". In OCR projects, per-page OCR fees can bite before LLM cost.
6. Latest Products & Concepts
Adjacent tech that shows up in deals, and the keywords coming next.
Structuring tech that connects scattered data by "meaning" — core of data-AI platforms
Represents relationships between concepts (customer, account, deal) as a graph. Lets AI use data across multiple systems that RAG alone can't bridge. Central to Palantir's data platform. A keyword for next-gen data projects.
AI that operates in the real, physical world — robots, autonomous transport, factory automation
Beyond the digital, it works with sensors, robot arms, and AGVs to automate physical tasks. NVIDIA's Cosmos (world models) and Isaac (robot simulation) are leading platforms. One of Adsur's focus areas.
AI that sees a PC screen like a human and operates the mouse and keyboard
Recognizes screenshots to click and type, so it can drive systems that have no API. A next-gen take on RPA (6), drawing attention for office-task automation. Accuracy, speed, and stability are still maturing — verify before production use.
Turns text in paper/PDF/images into data — essential for automating forms
Scans forms to read "what is written where." Handwriting accuracy can make or break a project. Key cost point: OCR is usually billed per page and is often the main cost of one pipeline run (sometimes more than LLM tokens). "How many pages per month?" is the estimate's key.
The "connection port" through which systems exchange data
The entry point for calling AI or external services from a program. Whether a customer's system "has an API" is a fork for automation. API = clean, stable integration. No API = drive the human screen with a robot (RPA/Computer Use), adding constraints. "Does your system have an API?" is essential.
Software that mimics the clicks and typing a human does on a PC
For legacy systems without an API, RPA imitates human screen operations to auto-fill and transcribe. Caveat: most RPA assumes a human-logged-in screen, so fully unattended background runs can be hard. Simple transcription = RPA; if judgment is needed, combine with AI (4).
Small pre-production trials. PoC = "can it be done technically"; PoV = "does it deliver business value"
PoC confirms technical feasibility; PoV focuses on real effect and ROI once deployed. Both are the on-ramp for "start small, convert to a production order" — Adsur's standard way to run AI deals.