Most of the confusion in AI conversations comes from one word covering very different technologies. A fraud-detection system, a chatbot and a software agent that books meetings are all called "AI", yet they fail in different ways, need different oversight and create value differently. Three categories are enough to classify almost anything a vendor will show you.

Predictive AI learns patterns from historical data to forecast or classify: which clients are likely to leave, whether a transaction is fraudulent, how long a contract review will take. Its output is a score, a probability or a label. It has long been used in credit scoring and demand forecasting. Its strength is consistency at scale. Its weakness is that it can only reflect the data it learned from, so if that data was biased or the world changes, its predictions drift, often quietly. The leadership question is: how accurate is it, on whose data, and what happens when it's wrong?

Generative AI creates new content — text, images, code, audio — in response to an instruction called a prompt. ChatGPT, Claude, Gemini and Microsoft 365 Copilot are the best-known examples. Its output is a draft: a summary, an email, a clause, a slide. Its errors look exactly like its successes: a wrong answer arrives in the same fluent, confident prose as a right one. The leadership question is: who checks the draft before it matters?

Agentic AI goes a step further. An agent is given a goal rather than a single instruction, and it works towards that goal in several steps, choosing and using tools — searching a database, reading documents, updating a record — and checking its progress as it goes. Most agents today use a generative model to do the reasoning. Their output is an action, not just a draft. That is where the value lies, because work gets done without someone directing every step, and it is also where the risk lies, because an action can have consequences before anyone has looked at it. The leadership question is: what is the agent allowed to do without asking?

Two cautions. Many products combine types: a CRM might score leads (predictive), draft the follow-up email (generative) and book the call (agentic), so classify the components, not the brand. And "agent" is now used loosely in marketing; Gartner has warned of "agent washing", where existing assistants and chatbots are relabelled as agents. The test is whether the system can pursue a multi-step goal and take actions with tools without step-by-step instruction.

The short version for a leadership team: predictive AI tells you what is likely, generative AI drafts what you asked for, and agentic AI does what you delegated.

A large language model (LLM) is trained on a very large collection of text — books, websites, code and other documents — to do one thing: predict what comes next. Text is broken into tokens, which are words or fragments of words. In training it repeatedly guesses the next token across billions of passages and is adjusted each time it is wrong. Over time it builds an immensely detailed statistical picture of how language fits together: which words follow which, how arguments are structured, what a contract clause or a board paper typically looks like. The underlying design, the transformer, was published by Google researchers in 2017.

When you type a prompt, the model writes its answer one token at a time, each time choosing a likely next token given everything so far: your prompt, any documents you attached and the words it has already written. The amount of text it can take into account at once is its context window. Developers then add further training, including feedback from human reviewers, so that the model follows instructions and answers helpfully. A large general-purpose model of this kind is called a foundation model: thousands of products are built on top of one, which is why the same few providers sit behind so many tools.

Three consequences matter for leaders. First, the model doesn't look facts up; it generates text that is statistically likely. Usually likely and true coincide, because the training data contained the true answer many times. When they don't — an obscure case, a recent event, a precise figure — the model still produces something fluent. That is a hallucination, a by-product of the same mechanism that makes the model fluent, which is why it is unlikely to be engineered away entirely. Second, there is a deliberate element of chance in which likely token gets chosen, so the same question can produce different answers. That helps with drafting and gets in the way of anything that must be repeatable. Third, the model knows nothing about your organisation unless you give it the information. Enterprise tools handle this by grounding: they retrieve relevant documents from your own systems and add them to the prompt, so the answer draws on your material. Grounding reduces hallucination but doesn't eliminate it, because the model can still misread or embellish what it retrieved.

A useful working picture is a very well-read, very fast junior colleague with no memory of your firm, no reliable sense of when it is wrong and a tendency always to produce an answer. You would give that colleague a lot of work. You wouldn't let them send anything to a client or a court unchecked.

The AI market looks chaotic, but it has four layers.

At the base are the foundation model providers, the few companies that train the largest models: OpenAI (the GPT models behind ChatGPT), Anthropic (Claude), Google (Gemini), Meta (Llama) and a growing group of others, including Mistral in Europe and several Chinese labs. Some models are closed, available only through the provider's own services; others are released as open models that organisations can run on their own infrastructure. Leaders rarely buy from this layer directly, but it sets what everything above it can do.

The second layer is general assistants: ChatGPT, Claude, Gemini and Copilot used as stand-alone chat tools, usually licensed per user. They are quick to adopt, and they are where most staff first meet generative AI, often on personal accounts before any policy exists. That unsanctioned use is called shadow AI, and it is a common way for company data to end up in tools nobody approved.

The third layer is AI embedded in the enterprise software you already run. Microsoft 365 Copilot works across Word, Outlook, Teams and Excel and now includes agents such as Researcher for multi-step research. Salesforce has folded its Einstein features into Agentforce, now sold as Agentforce 360, which places agents inside its CRM. Google sells Gemini Enterprise, a platform for chat, search and agents that connects to Google Workspace and other business systems. The AI sits where your data and workflows already are, but each vendor's AI works best inside that vendor's walls.

The fourth layer is specialist vertical tools built for one profession or task, such as legal research, contract review, financial reporting or recruitment. Most sit on a foundation model and add domain data, workflows and safeguards. For a firm like Meridian, with legal, financial advisory and consulting service lines, this is the layer where much of the competitive difference will come from.

Every adoption decision across these layers takes one of three forms: buy a tool, embed AI in a platform you already use, or build on a foundation model yourself, a choice Module 3 returns to. In vendor conversations, ask which foundation model sits underneath, because it shapes capability, cost and where your data goes. And treat product names as temporary: Salesforce's AI alone has been renamed several times since 2023, so anchor decisions on what a tool does, not what it is called this year.

The other half of the picture is whether your organisation is ready to use any of it well, which depends less on technology than on leadership literacy, data, current usage, culture, governance and budget. This lesson's readiness assessment measures exactly that.

Most failed AI projects don't fail because the technology is weak. They fail because someone assumed it could do something it can't. Three misconceptions do most of the damage.

Misconception 1: AI is always accurate. Decades of spreadsheets and calculators have taught us that the same input gives the same, correct output. Generative AI isn't a calculator: it produces likely outputs, and likely isn't the same as correct. Predictive models have error rates too, measured on past data that may not match the future. The legal profession learned this in Mata v. Avianca (2023), when a US federal court in New York fined two lawyers and their firm $5,000 for filing a brief that cited court decisions ChatGPT had invented, and for standing by them after the court questioned whether they existed. The judge noted that using a reliable AI tool for assistance isn't inherently improper; the failure was that the lawyers didn't check. So the planning question isn't "is it accurate?" but "how often is it wrong, how would we know, and what does checking cost?"

Misconception 2: AI understands meaning. Because an LLM writes fluently, it is natural to assume it grasps what it is saying. It doesn't, at least not in the way a professional does. It has no goals, no stake in the outcome and no knowledge of your client beyond what you give it. It can summarise a contract elegantly and still miss that a clause is unusual for this particular client, because noticing that takes context and judgement it lacks. Treat fluency as a property of the output, not as evidence of comprehension.

Misconception 3: AI can replace human judgement in high-stakes decisions. AI can inform a decision about a hire, a credit limit, a legal strategy or a client recommendation, but it shouldn't make that decision alone. Accountability can't be delegated to software: regulators, courts and clients will hold a person responsible. The EU AI Act classes AI used in areas such as recruitment and credit scoring as high-risk and will require human oversight of it once those obligations apply from 2 December 2027; Module 5 covers what that means in practice.

None of this makes AI weak. It is very good at first drafts, summaries, classifying large volumes of material, finding patterns in large data sets and doing repetitive knowledge work quickly. The practical rule is to match the task to the tolerance for error. Where mistakes are cheap and easy to spot, let AI do more. Where they are costly or hard to detect, keep a qualified person in the loop and budget for their time, because the cost of checking is part of the cost of the AI.

An agent works in a loop. It perceives: it reads its instructions, its inputs and the results of anything it has done so far. It reasons: it decides which step would move it closer to the goal and which of its tools to use. It acts: it runs a search, opens a document, drafts a summary or updates a record. Then it evaluates the result and goes round again until the goal is met, it reaches a limit or a person steps in. Nobody writes the sequence in advance, and that is what separates an agent from ordinary automation.

Most first agents should be one agent with a few tools. A multi-agent system splits the work between specialists, such as a researcher, a writer and a fact-checker coordinated by an orchestrating agent. It can handle more complex tasks, but every hand-off adds cost, delay and another place for errors to compound, so add agents only when one demonstrably can't cope.

The most important design choice is autonomy. Fully automated agents act and report afterwards. Supervised agents do the work, but a person approves before anything has an external effect. Copilots propose and a person decides. Choose the level by how severe a mistake would be and how reliable the agent has proved, and move up only as evidence builds. Good first use cases have a well-defined goal, an output someone can check, and actions that are reversible or low in consequence. Legal research and data extraction fit, because a reviewer sees the result before it is used; customer communication and financial transactions don't, because a mistake leaves the building before anyone sees it.

Guardrails make autonomy trustworthy: output validation against rules or sources, limits on how often and how much the agent can act, escalation triggers that hand the task to a person, and audit logs of every step. Vendors build some guardrails into their platforms, but the ones that express your risk appetite, such as which actions need approval, are yours to design.

Many agentic pilots stall before they scale. In June 2025 Gartner predicted that over 40% of agentic AI projects will be cancelled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. The usual causes are a vague goal, agents that can't reach the systems they need, cost and response times that surprise people as usage grows, and users who stop trusting the outputs after a few early errors.

You can build intuition without an IT project. A personal AI workbench — an assistant that triages your email, synthesises research or prepares a briefing for your next meeting — shows you within weeks where agents help and where they need watching.