From oracle to operator.

Picture the owner of a small events business on a Monday morning. She asks a chatbot to help her organise a conference for 120 people, and within seconds it produces an admirable plan: confirm numbers, fix a budget, shortlist venues, check dates, book catering. Then it stops. The plan sits politely on her screen, and the real work — ringing venues, comparing quotes, chasing replies, paying the deposit — is still hers.

For most of its short public life, generative AI has behaved like an oracle. You bring it a question, it gives an answer, and then it waits for the next question. The next wave is different in kind, not just in degree. It does not simply tell you what to do. It goes and does it. It is the difference between asking for directions and handing over the keys.

IBM's course on generative AI draws the line crisply. Generative systems are reactive: they wait for a prompt, produce content from patterns learned in training, and their work ends at the point of generation. Agentic systems are proactive: they usually start from a prompt too, but then pursue a goal through a series of actions with minimal human intervention. The oracle answers. The operator acts.

That sounds like a technical footnote. It is not. Handing over the keys changes what these systems are worth, what can go wrong and, above all, where people stand. With that in mind, here are four things every business leader should understand before giving the keys to an agent.

01An agent works in a loop, not a line

A chatbot's work is a line: prompt in, content out. An agent's work is a loop. The IBM material describes four steps — perceive the environment, decide on an action, execute it, learn from the result — and then round again. IBM's introductory AI course defines agents in the same spirit, as software that interacts with its environment and performs tasks on its own.

What makes the loop possible is that the large language models that write fluent prose can also plan. Asked to organise that conference, an agent uses the model to talk itself through the job: first the requirements (size, duration, budget), then venues that fit, then which of those are free on the day. Researchers call this "chain-of-thought" reasoning, and the course describes generative AI as the "cognitive engine" behind an agent's decisions. In 2022 a team from Princeton and Google showed, in a method they called ReAct, that alternating reasoning with actions — think, search, check, think again — produced more grounded results than reasoning alone, because the model could test its thinking against the world rather than against its own memory.

Insight: a loop compounds. A good step makes the next one easier; a wrong fact early on gets acted upon later. Judge an agent on the whole journey, not on one clever answer.

02A good driver needs a reliable map

An operator is only as good as the facts it works from: your price list, your refund policy, last week's emails. A model on its own knows only what it absorbed in training, which may be out of date and is never specific to your business. The remedy, which IBM's introductory course covers, is retrieval-augmented generation, or RAG: before answering, the system fetches the relevant documents and writes with them in hand.

The technique was set out in 2020 by researchers at Facebook AI Research, University College London and New York University. They highlighted two practical virtues. Knowledge can be updated by swapping the documents rather than retraining the model, and the retrieved sources can be inspected. For a business the second matters more than the first. An agent that can show where it found a fact is an agent you can audit.

Insight: before you connect an agent to anything, decide which documents are the source of truth, and ask it to cite them every time.

You can delegate the task; you cannot delegate the responsibility.

03The human moves from the driver's seat to the control tower

Here is the crux. The IBM course contrasts two examples. In the first, a YouTuber uses generative AI to review scripts, suggest thumbnails and compose background music, but a person reviews and refines every output. In the second, a personal shopping agent hunts for stock across platforms, watches prices, handles checkout and arranges delivery, asking its owner for input only when it needs to.

In the first case the human is the author and editor, seeing everything. In the second the human becomes an exception-handler, seeing only what the system chooses to escalate. That is a profound change in the job description. Editors catch mistakes because every draft crosses their desk. Exception-handlers catch only the mistakes the machine recognises as mistakes.

Accountability does not move with the keys. When an agent pays the wrong supplier or sends the wrong message to a customer, "the AI did it" is not an answer a customer, an auditor or a regulator will accept. You can delegate the task; you cannot delegate the responsibility.

Insight: write down the stop signs before you write down the goal. Payments, anything sent to a customer and anything deleted should need a human "yes".

04Start with one journey, not the whole fleet

None of this is a reason for timidity. The dull middle of most jobs — chasing, comparing, copying between systems, following up — is exactly where agents shine, and every hour they take back is an hour for judgement, relationships and creative work. The course's own conclusion is measured and, I think, right: the most powerful systems will be neither purely generative nor purely agentic, but intelligent collaborators that know when to explore options and when to commit to a course of action.

The risks are real, though, and the researchers say so themselves. The ReAct authors warned that connecting a language model to real-world actions brings dangers, such as looking up private information or taking harmful actions, and they deliberately limited their experiments to safe environments. Businesses should do the same: begin with one bounded, measurable task, run it with tight permissions, read the logs, and widen the agent's licence only as it earns trust.

Insight: the most valuable skill of the agentic era is supervision: setting clear goals, reading what the agent did and spotting the odd escalation. It can be taught, and teams that learn it early will move faster.

In short

Generative AI gave us a brilliant oracle. Agentic AI offers an operator that perceives, decides, acts and learns in a loop. To make the most of it, give it a reliable map in the form of your own documents, mark the stop signs before it sets off, and start with one journey. The keys can be handed over. The responsibility stays with us.

"The AI generates possibilities, but the human curates them." — from IBM's Generative AI: Introduction and Applications course

The agentic era asks us to keep the spirit of that line even when we are no longer curating every step.

(MCP: Model Context Protocol, the standard that lets AI tools plug into other apps.)

So, which job would you hand over to your first agent, and where would you put the stop sign? What are your insights?

— Will

Written by Will, with Bloggy, his AI editor