AI Literacy for Executives: A 10-Minute Meeting Prep Guide

The owner of a 60-person furniture manufacturer walked out of a meeting last month and said this: "I nodded for forty-five minutes. The guy said RAG, he said agents, he said tokens, and I kept saying 'sure, sure.' Then he left a proposal on my desk for about ten thousand dollars. I don't know what I said yes to." We have heard that sentence, in different cities and different industries, in almost exactly those words, more times than we can count.

The problem was never a lack of technical knowledge. Someone who built a factory, refinanced it through three downturns and still signs every payroll does not go quiet in an AI meeting because they are slow. They go quiet because nobody has translated the language for them. AI literacy for executives is not a shrunken version of what an engineer knows. It is a different skill: telling which word in the vendor's sentence means money, which means risk, and which means nothing at all.

This article does that translation. One-line meanings for the terms you will hear, the questions worth asking in any AI pitch, the sentences that quietly mark you as unprepared, and a second version of that furniture meeting that ended differently. For the wider picture, our end-to-end guide to AI for small business covers strategy, cost and rollout; this piece stays inside the meeting room.

Why AI literacy for executives can't wait another quarter

Because the decision is yours, and the person across the table knows it. Adoption is rising from a low base almost everywhere, which means more first-time buyers are walking into these meetings every month, while vendors are giving the same pitch for the hundredth time. That gap in experience is more dangerous than any gap in knowledge.

Turkey offers a clean data point on how uneven this is. The national statistics office reported in 2025 that 7.5 percent of firms with ten or more employees use some form of AI, up from 2.7 percent in 2021. Split by size, it reads 6.6 percent for firms with 10 to 49 staff, 9.6 percent for 50 to 249, and 24.1 percent above 250. One in fifteen small firms; one in four large ones. As that spread closes, the person deciding is the executive who nodded through their first meeting.

And "thinking you understand" is a documented trap. In BCG's 2026 survey of CEOs and board members, three quarters of directors said they understood AI. In the same survey, 35 percent of CEOs said their boards overestimate how far AI can replace people, and 61 percent said boards were rushing the transformation. Translated: the people at the decision table feel informed, and the people running the projects disagree. The riskiest meeting we see in practice is the one where everybody is sure they have understood.

Twelve terms you'll hear, each in one sentence

The list below covers roughly nine out of ten terms that come up in an AI meeting. You do not need to memorise them. You need to know which ones are about money, which are about risk, and which are just names. Most of them turn out to be three or four ideas wearing different labels.

• Artificial intelligence: The umbrella term. Software that learns from examples to predict, decide or generate. A family, not a product.
• Machine learning: Learning rules from data instead of writing them by hand. "Predict this year's sales from last year's" belongs here. It has existed for decades.
• Generative AI: The kind that produces text, images and code. ChatGPT, Gemini and Claude live here, and so does most of the noise of the past three years.
• Large language model (LLM): The engine behind generative AI. A system trained on enormous amounts of text to predict the next word. It does not know everything, but it writes very convincingly; keep those two facts apart.
• Prompt: The instruction you give the model. Think of the brief you hand a new hire. A bad brief gets bad work.
• Token: The unit the model counts text in, roughly a fragment of a word. Billing runs on tokens, and languages other than English often use more of them per sentence. When a vendor says "token," hear "cost."
• Hallucination: The model stating something false with full confidence. It is not a bug, it is the nature of the tool, which is why "human review" belongs in the contract.
• RAG: Retrieval-augmented generation, meaning the model reads your documents before answering. If you want "a chatbot that knows our contracts," this is the word you are looking for. If your documents are a mess, RAG gives messy answers.
• Fine-tuning: Retraining a ready-made model on your own data. Expensive and unnecessary in most small-business scenarios. If a vendor leads with it, ask why RAG would not do.
• Agent: AI that carries out multi-step work rather than answering one question: reading an email, updating the CRM, drafting a quote. You are granting authority, so the risk grows with it. Our decision tree for chatbots, agents and plain automation separates the three.
• Model, API, application: The model is the engine, the API is the tap you rent it through, the application is the product wrapped around it. Most "AI products" rent the same three or four engines. Knowing the vendor did not build the engine helps in price negotiations.
• Human in the loop: AI proposes, a person approves, on anything that matters. Regulators expect it, and it is the first thing to ask about when you hear "fully automated."

Automation or AI? The distinction vendors blur most often

Automation is a pre-written rule running on its own; AI is the ability to handle a situation the rule never described. "Forward the invoice to accounting when it arrives" is automation. "This invoice doesn't match the supplier's usual pricing, please check" is AI. A meaningful share of the products you will be shown belong to the first kind and carry the name of the second.

The industry calls this AI washing: sticking an AI label on rule-based software. Two questions cut through it. First: "What data does the system learn from, and how does it improve over time?" No answer means nothing is learning. Second: "What does it do that a rule couldn't?" The honest answer sounds like "recognise patterns or free-form language nobody defined in advance." A product that fails both is not a bad product. It is simply not what it was sold as.

One more thing: sometimes automation is exactly what you need, and it is cheaper. The furniture manufacturer's order-confirmation process did not need AI; three rules and an email template solved it. An executive who knows the difference can turn a ten-thousand-dollar proposal into a one-thousand-dollar automation.

Seven questions to ask in any AI pitch

These seven questions measure the seriousness of a proposal without any technical background. You do not have to ask all of them; if the first three get no clear answer, you can politely end the meeting. They get harder as they go, and each one closes a door a vendor might otherwise slip through.

• "Which kind of AI? Prediction, language model, or rule-based?" A vendor who cannot say either does not know or does not want to.
• "Show it live, on our data, now." Not a recorded demo. Take ten rows from your own order list and test it on the spot. Real products work on real data.
• "Where does it get things wrong?" A vendor who names no limits knows the product less well than you will. Trust the one who says "on that type of question the error rate is around 15 percent."
• "Where is our data processed, and is it used to train the model?" This is a legal question, not a technical one; the data-protection section below explains why.
• "What's the total cost: licence, usage, integration, data preparation, human review?" The number in the proposal usually covers the first item. The other four often exceed it in the first year.
• "Which specific number will have changed in 90 days?" Quote turnaround, return rate, call length, anything. A benefit that cannot be measured is a slide.
• "When it makes a mistake, who is responsible, is there a log, and can it be undone?" For agent products this becomes the most important clause in the contract.

If you want the pricing side of these questions in more depth, our piece on twelve questions to ask before hiring an AI consultant covers retainer and project ranges as well.

Five sentences that quietly mark you as unprepared

Some sentences, the moment they are spoken, tell the technical people in the room that the executive has not done the reading. These are the ones we hear most. Under each is what is wrong with it and what to say instead. Knowing these earns more respect than knowing the vocabulary.

• "Our data is a bit messy, but the system will sort it out." It will not. AI produces what it was fed. Projects that start from scattered spreadsheets spend their first three months on data cleaning, and that rarely appears in the proposal. Instead: "How much time and budget should we set aside for data preparation?"
• "We asked ChatGPT and it said the same." A general model does not know your company; it knows general information up to its training date. Instead: "How will the model access our documents?"
• "Let's roll it out company-wide right away." This is precisely what 61 percent of CEOs in the BCG survey complained about. Instead: "Which single process do we start with, and how do we measure it in 90 days?"
• "Hallucinations will probably be fixed in the next version." They are shrinking, not disappearing. Human review on critical output is a permanent cost line. Instead: "Which outputs will go through human approval?"
• "Let's train our own model, it'll be smarter." Expensive and unnecessary for most small-business needs; good prompting and RAG come first. Instead: "Why would we need fine-tuning?"

Data protection in two minutes, with a Turkey spotlight

Whatever your jurisdiction, the core principle is the same: the company using the AI tool is the data controller, the platform is a processor, and responsibility stays with you. "The tool made the mistake" is not a defence. Cloud-hosted models usually process data abroad, so cross-border transfer rules apply, and data minimisation means pasting a full customer list into a chat box is a risk in itself.

Turkey is a useful example of how quickly this is being formalised. Its data protection authority published a 64-page generative AI guide in November 2025 that spells out exactly this controller-processor split, plus a short "15 questions" summary written for managers rather than lawyers. If you operate there, our breakdown of the ten rules in that guidance is the practical version. Elsewhere, the fourth question above, where is our data processed and is it used for training, is the one that keeps you on the right side of whichever rulebook applies.

A concrete scenario: the furniture meeting, second attempt

Three weeks later the same manufacturer sat down with the same vendor. This time there was a single sheet on the table: twelve terms, seven questions. The meeting again ran forty-five minutes, and ended differently. In the first ten minutes the vendor presented "AI-powered order management." The owner asked: "What data does it learn from?" The answer wandered. Second question: "What does it do that a rule couldn't?" The vendor conceded that the product was, at heart, a rules engine sending emails based on order status.

The second item in the proposal was more interesting: a language model that read free-text requests from dealers and extracted dimensions, finish and delivery date into the ERP. That was genuine AI work. Dealers wrote things like "same as last time but walnut veneer, by month end" over messaging apps, and someone spent two hours a day keying it in. The owner asked for a live test. Out of that day's ten messages the system got eight right and confused the dimensions on two. The answer to "where does it get things wrong" appeared on the table: a 20 percent error rate, acceptable with a human-approval step.

The outcome: a package worth around ten thousand dollars became a single-process pilot at roughly fifteen hundred. The rules-engine part was solved for free with a setting in the existing ERP. The language-model part got a 90-day metric: two hours of manual entry a day to under 30 minutes. The contract gained a clause on where data is processed and a ban on using it for training. The owner still cannot explain how RAG works. But they know what they are paying for, and that was the whole point.

Frequently asked questions

How much AI do I need to understand as an executive?

The decision, not the technology. Recognising the twelve terms and being able to ask the seven questions is almost all the knowledge an executive needs. How the code is written is your team's or your vendor's problem. Your problem is which measurable result the money is tied to.

Nobody on my team is technical. Who should join the meeting?

Whoever knows the process best. If order entry is being discussed, bring the person who does order entry; they will ask the sharpest question during the live demo. An independent adviser is valuable in the second meeting, but the process owner matters more in the first.

When a vendor says "our model," is it really theirs?

Rarely. Most products rent an engine from a handful of providers and add an interface and a workflow. That is fine, but "which engine, in which country, under which terms" determines your data-protection obligations.

How much should I budget for an AI project?

For a single-process pilot we typically see figures in the low thousands of dollars; broader integrations run several times that. Before the number, settle "which single metric will change." The budget takes its shape from that metric.

My team already uses ChatGPT. Is that a problem?

Without rules, yes. Data minimisation principles target exactly this scenario. A one-page usage policy manages both the risk and the productivity gain better than a ban.

So what should you actually do?

• Prepare a one-page sheet: the twelve terms and seven questions from this article. Bring it to the meeting and put it on the table in the first ten minutes.
• Announce the live-demo condition in advance: write "we will test live on ten examples from our own data" into the meeting invitation. A vendor who arrives unprepared has already answered.
• Start with one process, measure one number: the wider the scope, the more invisible the failure. Set a single 90-day metric.
• Move the data question into the contract: where is it processed, is it used for training, when is it deleted. A verbal answer is not enough; our eight-clause vendor contract checklist gives these questions their written form.
• Seat the process owner at the table: the person who does the work evaluates the demo best.

What the owner said after the second meeting is the best summary of this article: "I learned to say no to what I don't understand, and yes to what I do." If there is a proposal heading for your desk, that one-page sheet is something we can draft with you; half an hour before the meeting is usually enough.