Implementing Agentic AI Ep 1: What an Agent Actually Is

The support team at a Gulf logistics company launched their "AI agent" in March. It greeted customers by name. It apologised beautifully. And when asked to actually change a delivery address, it said: "I understand your frustration." It understood everything. It could do nothing. The word "agent" had been doing a lot of heavy lifting. This series exists so that doesn't happen to you. Implementing agentic AI: the working playbook. Part one of five. For leaders and practitioners who've been told "we should use agents" and now have to actually do it — safely, measurably, without the hype. Five short episodes: what an agent really is, choosing the first process, the mission brief, the guardrails auditors ask about, and proving value to your board. Each episode lives in a different department of one company — support, finance, HR, procurement, sales — because the method is the same everywhere. At the end you'll hold the finished playbook, and the template to fill in for your own organisation. Today, episode one: what an agent actually is — and isn't. We'll spend it in customer support, because that's where most companies meet "agents" first, and where the confusion costs the most. One company hosts the whole series: a Gulf logistics group whose leadership has just been told, by everyone, that agents are the answer. By the end of these five minutes you'll be able to look at any AI tool and place it precisely: chatbot, workflow, or agent — and know which one your problem actually needs. Three very different things get sold under the word "agent". A chatbot talks — retrieves answers, apologises fluently, touches nothing. Our cold open. A workflow does — along a path someone designed. Fixed steps, model inside them. An agent decides — reads the situation, picks its next step, uses tools, checks its work, keeps going until done or a rule says stop. Talk. Do. Decide. Three steps of one staircase — three price tags, three risk profiles — and vendors will happily sell you the word without telling you which step you're buying. Here's a support workflow at our logistics company. A ticket arrives. The model classifies it — delivery delay, damage claim, address change. It drafts the reply from the right template with the right shipment data. A human clicks send. Fixed path, model inside it. The same shape runs everywhere: in finance, an invoice — extract, match, flag mismatches for a person. In HR, a leave request — check balance, check policy, prepare the approval. The path is yours. The model just works the stations. Now a real agent, same department. A customer writes: "my shipment says delivered but nothing arrived." No template survives that. The agent pulls the tracking history, checks the driver's scan, compares the address on file against the delivery pin, spots the mismatch, opens an investigation with the carrier, and drafts the claim — deciding each next step from what the last one revealed. You couldn't draw that flowchart in advance. That's the tell. The same shape appears in procurement, chasing a missing clause across three systems — and in sales, reconciling a pipeline three systems disagree about. Path unknown, progress checkable — agent territory. Now the sentence that will save you the most money this year. Anthropic — the company behind Claude, with every incentive to sell you agents — reports plainly in its engineering guidance that the most successful implementations use simple, composable patterns — and that sometimes the right answer is no agent at all. Use the simplest thing that passes your evaluation; save true agents for paths that genuinely can't be drawn in advance. The support chatbot didn't fail because it wasn't an agent. It failed because nobody asked which step of the staircase the problem needed. The opposite mistake is quieter and pricier: a full agent wired to route password resets — a flowchart job — then a quarter spent reviewing decisions nobody needed made. So here's this episode's pocket rule — the first of five you'll collect. Who chooses the next step? If your flowchart chooses — it's a workflow. Build it with confidence; they're the workhorses. If the model must choose — it's an agent. And an agent needs a leash: boundaries, checkpoints, a human who reviews. That leash is episode four. Ask the question about any tool a vendor shows you, and watch the demo get more honest. And so the playbook opens — stamped by Dana, who runs digital here and will walk us through every floor of this series. Row one, support: ticket handling — workflow, the path is ours. And pencilled beside it, a second candidate: delivered-but-missing investigations — agent, you couldn't draw its path in advance, but you can check its progress. One row per department, and the first column — what is this, really? — now has a method behind it instead of a vendor's word. You'll watch this document fill through the series, and at the end you'll download the same template for your own departments. Episode one, in four lines. "Agent" is a spectrum: chatbots talk, workflows do, agents decide. Simple, composable patterns win in production — a strength, not a confession. The pocket rule: who chooses the next step? And every department has both kinds of problems — we saw four floors today without leaving one. Next time: episode two — choosing your first process. We go down two floors to finance, where an accounts-payable pilot either becomes the win that funds the whole programme, or the mess that ends it. You'll learn the four tests a first process must pass, the one mistake almost every company makes with its first pick — choosing the impressive process instead of the reversible one — and you'll fill the playbook's second column. Five minutes, when you're ready. See you in finance.