AI Questions, Answered

A token is the basic unit of text that an AI model processes. Think of it as how AI "reads" — instead of processing whole words, it breaks text into smaller pieces called tokens.

One token is roughly 4 characters of English text, or about ¾ of a word. For example:

• "The quick brown fox" = approximately 4–5 tokens
• "Artificial intelligence" = approximately 3 tokens

Tokens matter because every AI model has a token limit — the maximum number of tokens it can process in a single interaction. This limit covers both your input (the prompt) and the AI's output (the response). When you hear about a model having a "128,000 token context window," that's the total combined limit for input and output.

For business leaders, understanding tokens helps you:

• Estimate costs (many AI tools charge per token)
• Gauge how much text you can feed an AI at once
• Understand why longer documents might get summarized or truncated

Token efficiency means getting the maximum value out of every token you use when interacting with AI. It's about being clear, specific, and structured so you don't waste tokens — or money — on vague prompts that produce low-quality responses.

Why it matters:

• AI tools often charge per token, so wasted tokens = wasted money
• Token limits mean inefficient prompts leave less room for the AI's response
• Well-structured prompts produce better, more relevant answers

How to improve token efficiency:

• Be specific: "Write a 200-word LinkedIn post about AI adoption for HVAC company owners" beats "Write something about AI"
• Use structure: Bullet points and numbered lists are more token-efficient than paragraphs
• Front-load key instructions: Put the most important instruction first
• Avoid repetition: Don't restate the same instruction three different ways
• Set constraints: Tell the AI what NOT to do ("No hashtags, no emoji, under 150 words")

A token-efficient prompt gets you a better answer, faster, at lower cost — which matters when you're scaling AI across a team or organization.

A prompt is the instruction, question, or input you give to an AI model to get a response. It's how you communicate with AI — and the quality of your prompt directly determines the quality of what you get back.

Think of a prompt like a brief you'd give to a new employee. The more context, clarity, and direction you provide, the better the result.

What to include in a good prompt:

1. Role or context: Tell the AI who it's acting as — "Act as a business strategist who specializes in small business operations..."
2. Task: Clearly state what you want — "Write a one-page AI adoption plan for a 50-person HVAC company..."
3. Constraints: Set boundaries — "Keep it under 300 words. No technical jargon. Focus on actionable steps, not theory."
4. Format: Specify how you want the output — "Return as a numbered list with a brief summary at the top."
5. Examples (optional but powerful): Show the tone or style you want — "Here's an example of the tone I'm looking for: [paste example]"

Example of a good prompt:

"Act as a business operations consultant. I run a plumbing company with 30 technicians. Write a simple 5-step plan for how my team can start using AI tools to reduce administrative work. Keep each step to 2-3 sentences. No jargon. Return as a numbered list."

Example of a bad prompt:

"How can AI help my business?"

The difference is night and day. Good prompts produce specific, useful, actionable answers. Bad prompts produce generic, unhelpful encyclopedic responses.

An AI hallucination is when an AI model generates information that sounds confident, detailed, and correct — but is actually wrong, fabricated, or completely made up.

Hallucinations are one of the most important risks for business leaders to understand, because they can be surprisingly hard to detect. The AI doesn't know it's wrong — it's simply predicting the most likely next word based on patterns, not verifying facts.

Examples of AI hallucinations:

• Citing a research study that doesn't exist
• Providing a statistic that sounds plausible but is fabricated ("73% of small businesses use AI for hiring" — made up)
• Inventing book titles, author names, or URLs
• Describing a product feature that doesn't exist
• Quoting a person on something they never said

How to spot AI hallucinations:

• Verify facts: If the AI cites a statistic, study, or source, look it up independently
• Check specificity: The more specific and "perfect" a detail sounds, the more you should verify it
• Watch for confident citations: AI can generate realistic-looking academic citations that are entirely fabricated
• Cross-reference: Ask the same question to a different AI tool and compare answers
• Ask for sources: Prompt the AI to include source links — then check if those links actually work

How to interrupt hallucinations:

• Ask the AI to explain its reasoning step by step
• Instruct the AI: "If you're not sure, say 'I don't know' rather than guessing"
• Provide your own source material and tell the AI to only use what you've provided
• Break complex questions into smaller, more specific questions
• Ask the AI to flag which parts of its answer it's most and least confident about

AI hallucinations cannot be eliminated 100%, but they can be significantly reduced with the right approach. The key is understanding that AI is a drafting tool, not a truth engine — and building verification into your workflow.

Techniques to reduce hallucinations:

1. Provide source material: Instead of asking the AI to generate facts from memory, give it documents and tell it to only use what you've provided. This is called "grounding" the AI.
2. Use retrieval-augmented generation (RAG): Some AI tools connect to your own documents or databases, so the AI pulls answers from your verified data rather than its training data.
3. Set explicit boundaries: Include instructions like "Only use information from the documents I've provided" and "If the answer isn't in the provided material, say 'I don't have enough information to answer that.'"
4. Verify before you trust: Treat every AI output as a first draft. Check facts, statistics, and citations independently before using them in any business context.
5. Use AI for drafting, not deciding: AI is excellent at generating first drafts, organizing thoughts, and suggesting options. It should not be the final decision-maker on factual claims.
6. Keep a human in the loop: The most effective AI workflows have a human reviewing and approving outputs before they reach customers, clients, or public-facing channels.

The bottom line: you can't eliminate hallucinations, but you can build habits and guardrails that catch them before they cause damage. The leaders who win with AI aren't the ones who trust it blindly — they're the ones who verify systematically.

A context window is the maximum amount of text an AI model can process in a single interaction — measured in tokens. It includes both your input (your prompt and any documents you attach) and the AI's output (its response).

For example, if a model has a 128,000-token context window, that's roughly 96,000 words — or about a 300-page book. Everything you put in plus everything the AI generates has to fit within that limit.

Why it matters for business:

• It determines how much information you can feed the AI at once
• It affects whether you can paste in entire documents, reports, or contracts
• Larger context windows allow longer conversations without the AI "forgetting" earlier context

If you've ever had an AI seem to "forget" something you mentioned earlier in a conversation, it's likely because that content fell outside the context window.

A Large Language Model (LLM) is the type of AI behind tools like ChatGPT, Claude, and Google Gemini. It's trained on massive amounts of text data to predict what word should come next in a sentence — which, at scale, allows it to write, summarize, answer questions, and reason through problems.

Think of an LLM as a super-powered autocomplete. It doesn't "understand" language the way humans do — it recognizes patterns in how words are used together and predicts the most likely continuation.

What this means for business leaders:

• LLMs are excellent at language tasks: writing, summarizing, editing, translating, brainstorming
• They can reason through problems when prompted well
• They don't have real-time information unless connected to a search tool or given current data
• Different models (GPT-4, Claude, Gemini) have different strengths — some are better at writing, others at analysis

AI (Artificial Intelligence) is the broadest term — any system that performs tasks that normally require human intelligence. It includes everything from rule-based systems to modern chatbots.

Machine Learning (ML) is a subset of AI. Instead of programming explicit rules, you give the system data and let it learn patterns on its own. A spam filter that improves as you mark emails is machine learning.

Deep Learning is a subset of ML that uses neural networks with many layers (hence "deep"). This is the technology behind modern breakthroughs — ChatGPT, Claude, image generators, and voice assistants all use deep learning.

The nesting is: AI ⊃ Machine Learning ⊃ Deep Learning. All deep learning is machine learning. All machine learning is AI. But not all AI uses machine learning, and not all machine learning uses deep learning.

Data safety depends on the tool, the configuration, and what you put in. Most major AI tools (ChatGPT, Claude, Gemini) offer business and enterprise tiers that prevent your data from being used to train their models. Consumer/free tiers may use your conversations for training.

Best practices for business use:

• Never paste sensitive data into public AI tools: Social Security numbers, financial statements, customer PII, trade secrets
• Use business-tier versions: ChatGPT Team, Claude for Work, and Gemini for Google Workspace all have stronger data protections
• Create a team AI usage policy: Define what's safe to share and what isn't
• Train your team: Most data leaks come from well-meaning employees pasting something they shouldn't have

If you're using AI for anything involving customer data, financial information, or proprietary business data, use an enterprise tier and consult your IT or legal team.

The biggest mistake business leaders make with AI is trying to do everything at once. The second biggest mistake is doing nothing. Here's a practical starting point:

Step 1: Assess where you are. Take the free AI Leadership Scorecard at /ai-scorecard to identify your organization's AI readiness and gaps.

Step 2: Start with one use case. Pick one repetitive, time-consuming task — email drafting, meeting summaries, proposal templates — and use an AI tool to speed it up. Don't try to transform your whole company on day one.

Step 3: Build AI literacy. Make sure your team understands the basics: what a prompt is, what tokens are, how to spot hallucinations, and what data is safe to share.

Step 4: Get leadership aligned. AI adoption fails when leadership isn't on the same page. Use the AI Opportunity Blueprint at /blueprint to create a shared strategy across your leadership team.

Step 5: Scale what works. Once you've proven value in one area, expand to others. Document what works, create prompt templates, and train your team.

The goal isn't to become an AI company. The goal is to use AI to become a better version of your company.

A loop in AI is a repeating cycle where the AI takes an action, evaluates the result, and uses that result to improve the next step — continuing until a goal is met or a human intervenes. It's the difference between AI answering a single question and AI actively working through a multi-step task.

Types of loops business leaders should know:

• Human-in-the-loop: The AI does the work, but a human reviews and approves before anything is finalized. This is the safest model for business use — AI drafts, humans decide.
• Agentic loop: The AI acts autonomously in a cycle — plan, execute, evaluate, repeat — until it completes the task or hits a boundary. Think of it like a delegate you've given a task to with clear guardrails.
• Feedback loop: The AI's output is evaluated (by a human or another system) and that evaluation is fed back into the next prompt or iteration, improving results over time.

Why it matters for leaders:

• Human-in-the-loop is the standard for anything customer-facing, financial, or high-stakes. AI should never send an email, publish content, or make a purchase decision without human approval.
• Agentic loops are powerful for research, data analysis, and repetitive tasks — but they need clear boundaries, permissions, and a shutdown path.
• Feedback loops are how your team gets better results from AI over time — by refining prompts based on what works.

The key leadership question isn't "Can AI do this task?" It's "Where in this task does human judgment need to enter the loop?"

In AI engineering, a graph is a structured plan for how multiple AI agents work together on a task — a network of jobs ("nodes") connected by dependencies ("edges") that show which job needs to happen and which job has to wait for another. It's the evolution beyond a simple loop (one agent repeating try → check → adjust → repeat, on its own) into a system where multiple agents check and correct each other.

The two building blocks of a graph:

• Node: One job. One agent doing one task, with one input and one output — researching a competitor, writing a draft, checking a claim, verifying quality.
• Edge: The connection between nodes. It carries the output of one node into the next, and defines the order — which job needs what another job already produced.

A practical example — "The Graph Blueprint" pattern:

1. Goal — define what you're solving for
2. Split — break the work into pieces
3. Workers (parallel) — separate agents handle research, comparison, fact-checking, gap-finding
4. Verifier — a dedicated agent checks quality and catches mistakes or hallucinations before anything moves forward
5. Merge — combine everything that passed verification
6. Final Answer — one clear, actionable output

The key idea: instead of trusting one agent's single pass at a task, you build in dedicated checking and correction steps as part of the system itself.

Why this matters for business leaders:

• If you're evaluating an AI vendor or building an internal AI workflow, "is this a single loop or a graph?" is a fair question to ask. A single agent looping on its own has no built-in check on its own mistakes. A graph with a verifier step does.
• This is the same principle behind human-in-the-loop — except here, the "check" step can be another AI agent instead of (or in addition to) a person. For anything customer-facing or high-stakes, a human should still be the final check.
• The underlying idea isn't new — it's workflow orchestration and dependency mapping, concepts that have run critical business systems for decades. AI is just applying it to agents instead of people or servers.

Note: "graph engineering" is a fast-moving term in AI engineering circles — you may also hear it called "agent graphs" or "multi-agent orchestration." Source: adapted from Anatoli Kopadze's "Graph Engineering Explained" thread (X/Twitter, July 2026).