"Agentic AI" Is Everywhere. Most of It Isn't Agentic.

"Agentic AI" Is Everywhere. Most of It Isn't Agentic.

If you've sat through a vendor pitch in the last few months, you've heard the word. Agentic AI. Every product brochure has it, every conference keynote leads with it, and every automation tool that makes a single call to an LLM now claims it. The word has gone the way of "cloud" circa 2010 — universally claimed, rarely understood.

Having spent over three decades building AI and decision systems — long before generative AI made the field fashionable — I find the confusion costly, not just annoying. Businesses are paying agentic prices for ordinary automation, and worse, deploying genuinely autonomous systems in places where they shouldn't be. So let's draw the line clearly.

What Agentic Actually Means

An agentic AI system is given two things: a goal and a set of tools. What it is not given is a script.

The system looks at the current situation, reasons about it, chooses an action, executes it, observes the result, and decides what to do next. Then it does that again, and again, until the goal is met. The defining property is this: the path is not predetermined. Nobody drew the flowchart in advance, because the flowchart doesn't exist until the agent runs. Give it the same task twice and it may take two different routes, depending on what it encounters along the way.

The number of steps, the order of steps, which tools get used, whether to retry, whether to stop early, whether to escalate to a human — these decisions belong to the model, made at runtime, not to a developer who made them months ago at design time.

What Agentic Is Not

This is where most of the market sits, so it deserves equal attention. The following are useful, legitimate technologies — but they are not agents:

A chatbot is not an agent. Answering questions, however fluently, is conversation, not autonomous action toward a goal.

An automation pipeline with an LLM inside is not an agent. This is the most common impostor. Extract the document, classify it with an LLM, route it based on the label, send the email. The LLM is doing intelligent work — but it's doing that work inside fixed boxes. It never chooses the boxes. Step 1 always feeds step 2 feeds step 3. That is a workflow, and the developer made every decision about its shape before the first document ever arrived.

RPA with AI bolted on, scheduled scripts, keyword bots, and decision trees are not agents. All of them execute predetermined paths, however many branches those paths contain.

Here is the cleanest way I know to say it: in a workflow, intelligence sits inside the steps; in an agent, intelligence decides the steps.

Five Real Applications — Workflow vs. Agent

Definitions are abstract until you see them in your own business. Here are five scenarios any Malaysian SME or enterprise will recognise, each shown both ways.

1. Collections and Accounts Receivable

Workflow: Invoice hits 30 days overdue, the system sends reminder template A. At 45 days, template B. At 60 days, it flags the boss. Every customer, same treatment, forever.

Agent: You give it a goal — "improve collections on these 40 overdue accounts" — and access to your invoicing system, email history, and payment records. It reviews each account and notices things. One customer historically pays only after a phone reminder, never email — so it schedules a call task instead of sending another ignored message. Another went quiet after disputing a delivery; the agent finds the dispute thread and decides the dispute must be addressed before any chasing happens, or the chasing will backfire. A fifteen-year customer gets a gentler tone than a new account with no history. Forty accounts, potentially forty different paths — none of which you scripted.

2. Customer Service on WhatsApp

Workflow: A keyword bot. "Price" triggers the price list. "Hours" triggers opening hours. Anything outside the decision tree gets "Sorry, I didn't understand that."

Agent: A customer writes: "The blender I bought last month stopped working, and I'm travelling next week." There is no keyword for that. The agent looks up the order, confirms it's within warranty, registers the travel constraint, and reasons through the options: an expedited replacement that arrives before the trip, or a pickup arranged for after the customer returns. It offers both, books whichever the customer picks, and logs the case. It handled a compound, messy, real-world request by deciding — in the moment — what information it needed and what actions to take.

3. Procurement and Vendor Sourcing

Workflow: Purchase request comes in, system checks budget, routes for approval based on amount, issues the PO. Classic approval routing — valuable, and entirely predetermined.

Agent: The goal: "Source three quotes for industrial safety gloves meeting the required SIRIM standard, delivered to Pengerang by month-end." The agent searches the approved vendor list and finds only one supplier carries the specification. A workflow would stop or fail here. The agent decides to look beyond the list, identifies two additional candidates, verifies their SSM registration and certifications, sends out RFQs, and when responses arrive, compares them — flagging one quote whose pricing looks suspiciously low against market rates. It hands the procurement officer a decision-ready comparison, having navigated obstacles nobody anticipated.

4. Stock and Supplier Operations for Retail and F&B

Workflow: Stock drops below threshold, system fires a reorder to the default supplier. Simple, dependable, blind.

Agent: The goal: "Make sure we don't stock out for the Raya season." The agent examines last year's sales and finds the festive spike started earlier than the standard reorder logic assumes. It checks supplier records and notices the main supplier had delivery delays during the same period last year. So it decides to split the order between two suppliers, bring the timing forward by three weeks, and while doing so spots that one supplier has quietly raised prices — which it flags to the owner. The reorder threshold never saw any of this coming, because thresholds don't reason.

5. Sales and Marketing Follow-Up

Workflow: Every new lead enters the same three-email drip sequence. Day 1, day 4, day 10. Identical for everyone.

Agent: The goal: "Revive our dormant leads." The agent reads through past conversations. One lead had asked about a feature that didn't exist then — but has since launched; the agent decides to lead its outreach with exactly that news. Another lead went cold after a pricing objection; the agent checks whether the current promotion applies and tailors the message around it. A third showed no real intent at all, so the agent decides not to contact them and pollute the brand. It schedules sales calls only for the leads it judges warm. Same goal, different treatment for every lead — based on judgment, not a sequence.

The Litmus Test

Strip away the marketing and one question separates agents from everything else:

Can the system take a path you didn't program?

Can it decide to stop early? Retry a different way? Use a tool out of order? Escalate when it judges the situation ambiguous? If every possible route through the system was drawn by a developer before deployment, you are looking at a workflow — whatever the brochure says.

Ask your vendor this question. The answer, and how quickly it arrives, will tell you a great deal.

The Honest Conclusion: You Often Don't Want an Agent

Here is what the hype omits. Workflows are frequently the better choice. They are deterministic, auditable, cheaper to run, and predictable — exactly what you want for high-stakes, regulated, or compliance-bound processes where "the AI improvised" is not an acceptable audit finding. An agent trades that reliability for flexibility.

The decision rule is simple: use a workflow when the path through the task can be known in advance; use an agent when it cannot. Investigating why something went wrong, handling messy compound requests, navigating situations with obstacles nobody anticipated — that is agent territory. Processing a thousand identical transactions identically — that is workflow territory, and proudly so.

The businesses that win with AI over the next few years won't be the ones that bought the loudest buzzword. They'll be the ones that understood the difference — and demanded precision from their vendors instead of vocabulary.

The author has over 30 years of experience in AI and data engineering, from predictive analytics across Asia-Pacific telcos and banks to building modern GenAI solutions for Malaysian enterprises.