Generative, predictive, agentic: the three kinds of AI in plain English

"AI" has become a single word for several different technologies. One writes text and summarises documents. Another forecasts numbers from your historical data. A third carries out tasks across your systems with limited supervision. They behave differently, cost differently and carry different risks, and most of the confusion in management conversations about AI comes from treating them as one thing.

This explainer covers the three types you are most likely to meet: generative, predictive and agentic. We have put agentic last, because it is built on top of generative and is much easier to understand once that is in place.

Generative AI: it produces things

Generative AI is the type most people have now used. ChatGPT, Claude, Gemini and Copilot are all generative: you give them a request in ordinary language and they produce new content in return.

The important thing to understand is how it works. A generative model has been trained on a very large amount of text and has learned the patterns in it. When you ask a question, it does not look up the answer in a database. It produces the most plausible response given everything it has seen. That is why it is so fluent, and also why it can be fluently wrong.

Where it turns up in a business

• Summarising a sixty-page tender or contract into the ten points the estimator or director needs to know.
• Producing first drafts of method statements, job descriptions, board papers and meeting notes, for a person to edit.
• Answering staff questions from your own manuals, policies and specifications, with a reference to the page the answer came from.
• Turning a rambling technical enquiry email into a structured quote request the sales team can act on.
• Writing and explaining spreadsheet formulas, database queries and small scripts.

What to watch

Generative models fill gaps with plausible-sounding invention, usually called hallucination. That is acceptable for a draft someone will read and correct. It is not acceptable for anything published, sent to a customer or relied on without a check. The practical fixes are to point the model at your own documents so it answers from them rather than from memory, to ask it to cite where each answer came from, and to keep a person responsible for the output.

The other thing to watch is where your data goes. Consumer accounts on the big tools may use what you type to train future models. Business accounts, properly configured, should not. If your staff are already using these tools, that setting matters more than any feature.

---

Predictive AI: it estimates what is likely

Predictive AI is the oldest of the three, and it has quietly run parts of business for decades. Credit scoring, demand forecasting and the spam filter on your email are all predictive. The idea is simple: take historical data, find the patterns in it, and use those patterns to estimate what is likely to happen next.

Unlike a generative model, a predictive model does not produce prose. It produces a number, a probability or a category: this machine is 80% likely to fail in the next fortnight; this invoice will probably be paid late; this image contains a defect. It is usually built to answer one specific question and trained mainly on your own data.

Where it turns up in a business

• Forecasting demand for stock lines, so purchasing orders the right quantity at the right time.
• Scoring which invoices are likely to be paid late, so credit control chases the right ones early.
• Predicting when a machine needs maintenance from vibration, temperature and run-hours, before it stops.
• Flagging unusual transactions in the accounts or expenses that deserve a second look.
• Spotting defects, missing safety equipment or unexpected movement in camera footage on a production line, at a gate or in a yard. Vision systems like this are a form of predictive AI: they classify what they see.

What to watch

A predictive model is only as good as the history you give it; patchy or badly recorded data produces unreliable forecasts. It also assumes the future will resemble the past. A new product line, a lost major customer or a supply shock breaks the pattern, and the model will not know. For many smaller businesses the honest conclusion is that the data is not yet clean enough for this to be worth doing, which is a good reason to start somewhere else and fix the record-keeping first.

---

Agentic AI: it carries out tasks

Agentic AI is the newest of the three and the most talked about. An agent is a generative model that has been given a goal, access to tools, and permission to take a sequence of steps on its own. Instead of handing you back a draft, it reads the mailbox, looks up the record, decides what to do next, does it, checks the result and moves on.

The word "agent" gets stretched a long way in sales material, so a useful test is this: does the system decide its own next step, and does it act inside your systems (send, update, create, book) rather than return text for a person to act on? If the answer to both is yes, it is agentic. If not, it is generative AI with a nicer interface.

Where it turns up in a business

• Receiving a supplier invoice, matching it to the purchase order and delivery note, flagging discrepancies and queuing the clean ones for approval.
• Watching a shared inbox, sorting enquiries, pulling the relevant customer or job record, drafting a reply and holding it for sign-off.
• Chasing missing paperwork: checking which subcontractors have not returned certificates, sending reminders and updating the tracker.
• Building the weekly operations report by pulling figures from finance, job costing and site logs, then writing the commentary.
• Onboarding a new starter across several systems: creating accounts, sending forms, booking induction.

What to watch

Because an agent acts, its mistakes compound. A wrong judgement early in the chain becomes a series of wrong actions later. The controls that matter are clear rules on what it may do without asking, what needs a person's approval, and a log of every action it took. Cost is the other surprise: an agent might make twenty or thirty model calls to complete one task, so a job that costs pennies as a single question can cost pounds as an agent. The sensible route is to start with agents that gather, prepare and recommend, and let them act unsupervised only once the pattern has proved itself.

---

How the three fit together

In practice, a useful deployment usually mixes them. A predictive model flags an invoice that is likely to be paid late. An agent gathers the account history and recent correspondence. A generative model drafts the chaser. A person approves it and it goes. Or a camera spots a fault on the line, an agent raises the maintenance ticket and the shift lead gets a plain-English summary.

The common thread is that all three depend on your information being somewhere the model can reach it: documents that are organised, systems that can be connected, and the way you do things written down. The model itself is increasingly the interchangeable part. Get the knowledge, the connections and the rules right once, and you can choose the model for each job and swap it when a better or cheaper one arrives.

---

Which should you start with?

That depends on where the problem is, not on which type is most fashionable. Generative is the cheapest to trial and shows results in days, which makes it the usual starting point. Predictive is worth it where a measurable decision is made repeatedly and the history behind it is clean. Agentic pays off where a task is well defined, repetitive and currently eats skilled people's time, but it needs the other two foundations and a proper set of guardrails first.