Plain-English definitions of the AI adoption frameworks and core AI vocabulary every leader and manager should know.
The first half is the operating language for leading AI adoption (Skill Inversion, Five Stages of Expertise Disruption, AI activation vs adoption). The second half is the minimum AI literacy for a leader (LLM, GPT, RAG, AI agent, prompt engineering, hallucination). Built for executives and managers, not engineers. Companion to Teach Them to Drive by Toni Dos Santos.
29 terms, two sections.
AI ADOPTION FRAMEWORKS
The operating language for leading AI adoption inside a company. Most of these terms are defined and unpacked in Teach Them to Drive. Use them with your team to make conversations precise.
What happens when AI compresses execution speed and inverts the value stack: judgment, context, evaluation and taste become the rare and valuable skills, while producing a first draft drops to near zero cost.
Senior people who used to win on output now have to win on review. Junior people who used to lose on output can produce a credible first pass in minutes. Skill Inversion is the central force behind every adoption failure on a senior team — the part the senior was good at has collapsed in cost, so they no longer recognise the work.
From the book: defined in chapter 2. Get Teach Them to Drive on Amazon →
The shift in where economic value is created on a team after AI: away from execution speed and toward judgment, framing, evaluation, and taste.
Value Migration is what Skill Inversion produces at the org level. Roles, comp bands, and promotion criteria all need to follow the migration. If a senior IC is still being paid for first-draft speed two years into the AI era, the org is mispricing the work.
The five predictable stages senior experts move through when AI lands on their team: Denial, Quiet Trial, Crisis, Repositioning, and Advocacy.
Skip a stage and you lose your best people. Lead them through it and they become your most powerful adoption advocates. Each stage has a specific leadership move: reduce threat, make learning safe, reframe value, give ownership of evaluation, make the new champions visible.
From the book: full chapter and table on the book page.
The recurring meeting, report, or task that everyone on a team complains about every week. Painful Tuesdays are the highest-leverage starting points for an AI workflow pilot.
If you ask a team “what would you most like to never do again,” the answer is usually a Painful Tuesday: the Monday status report, the QBR prep, the customer escalation triage. Pilot AI on the Painful Tuesday and the team gives you their attention because you are removing pain, not adding initiative.
AI training that starts from a real recurring workflow and rebuilds it with AI in the loop, instead of starting from the tool and hoping people find use cases.
The opposite of tool-first training (a one-day “here’s how ChatGPT works” session). Workflow-first training picks one Painful Tuesday per team, redesigns the steps with prompts and human checkpoints, then runs the new workflow for two weeks with measured before-and-after. Workflow-first training is the only kind that produces adoption rather than activation.
The three-step human review ritual for any AI output going to a customer, regulator or executive: Check the facts, Challenge the reasoning, Confirm the tone and policy fit.
Check-Challenge-Confirm is the answer to hallucinations, brand drift, and the “the AI said it” defence. It scales: every AI-generated draft passes through the same three filters, owned by the same human. Codify it once per workflow and the team stops sending bad AI output by accident.
Activation is a license metric (a user logged into the AI tool at least once). Adoption is a workflow metric (the work itself changed because of AI). Most companies report 80–90% activation but under 20% real adoption.
This is the single most important distinction in the field today. If your dashboard shows people using AI but you cannot point to a workflow that runs differently this quarter than last quarter, you have activation, not adoption. Read the deeper breakdown in our blog post Why AI Adoption Fails in Companies.
A workflow designed from scratch around an AI model doing most of the execution, with humans in defined oversight and judgment roles.
An AI-native workflow assumes the model is the default executor and asks “where does a human need to step in?” Examples: an AI sales-research agent that drafts every account brief overnight; an AI triage agent that classifies every inbound support ticket. AI-native is harder to build but produces 5–10x leverage when it works.
A traditional human workflow with AI bolted in at one or two steps to accelerate drafting, summarising or analysis. Lower transformation than AI-native, easier to start with.
Most enterprise AI today is AI-assisted: a marketer using ChatGPT to draft a post they will heavily rewrite, a lawyer using an AI summariser on a contract before reading it themselves. AI-assisted is the right starting point for most teams. AI-native is the destination.
Skill Inversion, the Five Stages of Expertise Disruption, the 90-Day AI Adoption Playbook, and the four core skills — all unpacked with examples, prompts and templates in Teach Them to Drive.
The minimum AI literacy for a leader. Knowing these terms is the difference between asking a vendor good questions and getting talked into the wrong product. Plain English, no engineering background required.
A statistical model trained on huge amounts of text to predict the next word. Modern LLMs (GPT, Claude, Gemini, Llama) can write, summarise, translate and reason in natural language. They power most chat-based AI tools used in business.
LLMs are not databases and they are not search engines. They generate text one token at a time based on patterns learned during training. That is why they sometimes hallucinate: confident-sounding wrong answers are a feature of how the technology works, not a bug to be fixed by the next model release.
Generative Pre-trained Transformer. The family of large language models built by OpenAI. GPT-4 and GPT-5 power ChatGPT and many enterprise AI products.
The acronym GPT is often used loosely in conversation to mean any modern chat-style LLM, but technically only OpenAI’s models are GPTs. Claude is built by Anthropic, Gemini by Google, Llama by Meta. They are all LLMs, only OpenAI’s are GPTs.
AI systems that produce new content (text, images, audio, video, code) rather than just classifying or predicting. ChatGPT, Midjourney, GitHub Copilot and ElevenLabs are all examples of generative AI.
The category that exploded in 2023 and changed the conversation in every boardroom. Predictive AI (fraud detection, demand forecasting, recommendation engines) has existed for years; generative AI is what made AI feel like a tool every employee could use directly.
A very large, general-purpose AI model trained on broad data that can then be adapted to many downstream tasks. GPT, Claude, Gemini and Llama are foundation models.
Most enterprise AI features today are built on top of a foundation model rather than from scratch. The strategic question for a leader is rarely “which model should we train” — it is “which foundation model should we build on, and how do we feed it our data through prompts and RAG.”
The instruction you give to an AI model. A prompt can be a single sentence, a long brief, an example to mimic, or a structured set of rules. The quality of the prompt is the single biggest lever on the quality of the output.
Most managers’ first AI experience is “the model gave me a generic answer.” That is almost always a prompt problem, not a model problem. A two-line prompt produces two-line output. A prompt with role, goal, constraints, examples and a request for step-by-step reasoning produces work you can ship.
The discipline of designing prompts that get reliable, useful, on-brand output from an AI model. Includes role-framing, constraints, examples, step-by-step reasoning and structured output.
The non-technical, manager-level version is usually called “prompt literacy”: a small set of repeatable habits non-technical employees can use every day. Most enterprises do not need a prompt engineering team; they need their managers to be prompt literate. See our blog post on prompt literacy for non-technical managers.
A hidden instruction set the AI tool uses for every conversation in a given context: who the model should pretend to be, what rules to follow, what tone to use, what it must never say.
Custom GPTs, Claude Projects, Copilot Agents and Gemini Gems are all wrappers around a system prompt. As an enterprise leader, the system prompt is where you encode your brand voice, your compliance rules and your task-specific instructions once, so every employee using that assistant gets consistent output without having to remember to ask.
The unit an AI model reads and writes in. A token is roughly three quarters of an English word. Pricing, speed and context-window limits are all measured in tokens, not characters or words.
Practical implication: 1,000 tokens is roughly 750 English words or 1.5 pages. When a vendor quotes “$3 per million input tokens,” that is about $3 per 1,500 pages of text fed in. Tokens are also why model output sometimes cuts off mid-sentence: it hit a token limit.
The maximum amount of text an AI model can hold in memory at one time, measured in tokens. Larger context windows let you paste longer documents.
Modern context windows range from 8,000 tokens (about 12 pages) on small models to 1–2 million tokens (multiple books) on the largest. But beware: even with a 1M context window, models still pay more attention to the start and end of the prompt than the middle. Long does not always mean reliable.
When an AI model produces a confident, plausible-sounding answer that is factually wrong. Hallucinations are the single biggest reason every AI output going to a customer or executive needs a human Check-Challenge-Confirm step.
Hallucinations are not a bug being patched in the next release — they are a structural property of how LLMs generate text. Newer models hallucinate less but never zero. The leader’s job is not to eliminate hallucinations but to design workflows where humans catch them before they reach customers.
Continuing to train a foundation model on your own data so it learns your domain, voice, or task. Fine-tuning is rarely the right first move for enterprises today.
Prompt engineering and RAG usually solve the same problem at one tenth of the cost and one hundredth of the maintenance burden. Fine-tuning makes sense when you need consistent output in a very specific format or voice and prompt engineering has hit a clear ceiling. For most companies, that day has not arrived.
A pattern where the AI model first retrieves relevant snippets from your own documents (a knowledge base, a Notion, a SharePoint) and then writes its answer using those snippets.
RAG is how most enterprises ground AI in their own data without fine-tuning. The flow: a user asks a question, the system searches your document store for relevant chunks (usually using embeddings and a vector database), passes those chunks to the LLM with the question, and the LLM writes an answer that cites the snippets. Almost every “chat with your docs” product is a RAG system under the hood.
An AI system that can take multiple steps on its own to accomplish a goal: search the web, call an API, write a file, send an email, retry on failure. Agents are how AI moves from chat to actually doing work.
The difference between a chatbot and an agent is autonomy. A chatbot answers your message and stops. An agent takes a goal (“research these 20 prospects and draft a tailored outreach email for each”) and executes the multi-step plan on its own. Agents need tight scope, clear guardrails and human checkpoints — they fail loudly when they are pointed at fuzzy goals with no boundaries.
A numeric representation of a piece of text (or image, audio) that captures its meaning. Embeddings are how AI systems find semantically similar content — the engine behind search, recommendations and the retrieval step in RAG.
Two pieces of text that mean the same thing have similar embeddings, even if they share no exact words. That is what makes embedding-based search smarter than keyword search: a query about “parental leave” can surface a policy document titled “maternity and paternity benefits” without the exact phrase appearing.
A database that stores embeddings and lets you search by semantic similarity instead of exact keywords. Vector databases (Pinecone, Weaviate, pgvector, Qdrant) are the storage layer behind most enterprise RAG systems.
For a leader, the practical takeaway is this: when a vendor pitches a “chat with your data” product, ask “what vector database do you use, where does it live, and who has access to the embeddings.” The answer tells you where your sensitive data is actually being stored.
An AI model that can read and produce more than one type of content — text, images, audio, video — in the same conversation.
GPT-4o, Claude and Gemini are multimodal: you can paste a screenshot of a chart, a photo of a whiteboard, or an audio recording, and the model can reason about it alongside text. Multimodal AI removes the “copy the data into a spreadsheet first” step from a lot of analyst workflows.
The act of running a trained model to get an answer. Every time you send a prompt, the provider runs an inference. Inference is what you actually pay for in the API price per token.
Training a foundation model costs hundreds of millions of dollars and happens once. Inference costs a fraction of a cent per request and happens billions of times a day. The economics of any AI feature you ship depend on inference cost, not training cost.
A setting that controls how random or creative an AI model’s output is. Low temperature (close to 0) produces consistent, conservative answers. High temperature produces more varied, creative answers.
Use low temperature for extraction, classification, structured output and anything where you want the same answer every time. Use higher temperature for brainstorming, creative writing, ideation. Most chat tools default to a middle setting; in the API you set it explicitly per request.
The questions that come up most often when leaders are getting up to speed on AI vocabulary.