What an AI agent actually does inside a business — a working example from a Malaysian F&B group
“AI agent” has become one of those phrases that means everything and nothing. Depending on who’s selling to you, an agent is a chatbot, a scheduling tool, a spreadsheet plugin, or a digital employee that will apparently run your company while you sleep.
We build agents for a living, so let me offer a more boring — and more useful — definition. An agent is software that takes a real input from your business, works through a series of steps the way a trained staff member would, and produces an output your operation actually uses. The key word is steps. Not one clever answer. A process.
The best way to explain is to walk through one that’s been running in production for a while now, inside a Malaysian F&B group with around twenty outlets across several brands and a central kitchen.
The problem: supplier invoices nobody wants to type
Every outlet in the group receives supplier invoices daily. Vegetables, meat, dry goods, packaging. Many of these invoices are handwritten — and this being Malaysia, they arrive in Malay, Chinese, and English, sometimes on the same page.
Before automation, someone had to key every line of every invoice into the accounting system. Across the group that’s tens of thousands of invoices a year. It’s slow, it’s error-prone, and it’s the kind of work that burns out exactly the people you want focused on running outlets.
It’s also — and this matters — work with a shape. Same input type, same decisions, same output, repeated thousands of times. That’s what agent-shaped work looks like.
What the agent does, step by step
The pipeline is deliberately simple to describe:
Read. An outlet staff member photographs the invoice. The agent extracts every line item from the image — including the handwritten ones, in all three languages. This alone was impossible to automate reliably until a couple of years ago.
Match. Each extracted line is matched against the group’s actual SKU list. “Bawang besar 10kg” needs to become the same inventory item every time, regardless of which supplier scrawled it or how.
Code. Each line gets assigned to the right general-ledger account, so the finance team’s chart of accounts stays clean without anyone manually classifying anything.
Flag. The agent checks prices against history. A supplier quietly charging 15% more for the same item this week is exactly the kind of thing that hides in manual data entry — and exactly what software is good at catching.
Then the important part: anything the agent is not confident about is held for a human to review before it posts to the ledger. Low-confidence lines don’t get guessed. They get queued.
Why the human-in-the-loop bit is the whole point
When people ask us whether AI is accurate enough to trust with their books, they’re usually imagining a system that’s either fully manual or fully automatic. The deployments that actually work are neither.
A well-built agent knows its own confidence. If the handwriting is clear and the SKU match is exact, it posts. If the photo is blurry or the item is ambiguous, it asks. In practice this means a task that used to consume hours of staff time per outlet per week becomes a review queue that one person clears in minutes — while the quality of the data goes up, because the price-anomaly checks catch things humans never had time to look for.
That’s the honest pitch for agents. Not “replace your team.” More like: give the repetitive 90% to software that never gets tired, and keep humans on the 10% that needs judgment.
How to spot agent-shaped work in your own operation
You almost certainly have processes like this. The tell-tale signs:
- Same input, over and over. Invoices, claim forms, delivery orders, booking requests, CVs, quotations.
- Rules a new hire could learn in a week. If you can write the SOP, an agent can probably follow it.
- A clear “done” state. The output lands somewhere concrete — a ledger, an inventory count, a CRM entry.
- Volume that makes people sloppy. Anywhere accuracy drops on Friday afternoons.
What’s not agent-shaped: negotiations, relationships, taste, strategy. Anything where the interesting part is the exception, not the rule.
A note on data, because someone will ask
Whenever AI touches operational data in Malaysia, PDPA questions follow — rightly. Supplier invoices are relatively low-sensitivity, but the same architecture questions apply to any deployment: what data does the model see, where is it processed, and what gets logged. We’ve written about that separately in our piece on AI and data privacy for Malaysian businesses, and about the PDPA compliance checklist we walk through with clients.
The short version: an agent doesn’t need access to everything to be useful. Good ones are scoped tightly — this one reads invoices and writes ledger entries, and that’s all it can do.
Where this goes next
The invoice reader is one worker inside a larger ERP we built for that group — the same platform reconciles two different POS systems into one daily ledger and handles e-Invois submission, which no restaurant operator should be doing by hand in 2026.
That’s generally how agent adoption works in practice. Not a grand AI transformation programme. One well-chosen process, automated properly, with humans kept where they matter. Then the next one.
If you’re trying to work out which process in your business should go first, that’s a conversation we genuinely enjoy having — it’s the same question we asked ourselves before we automated our own studio.