An award-winning AI agent for a traditional business is usually not flashy. It sits inside the chat tool staff already use, turns photos and files into one clean record, asks a person when something is unclear, and lets anyone check the status of an order without chasing a colleague. That is what the Hernan Operation Assistant does, the entry that won Agmo and our client Hernan Corporation the Bronze award at the Lark AI Innovator 2026 International Finals in Jakarta this week. This article explains how it works and what Malaysian SMEs can take from it.
The award, briefly
The international finals were held in Jakarta, Indonesia, with a venue in Japan linked live. Gold went to Satudora (Japan) and Silver to Puyo Group (Indonesia). Agmo and Hernan took Bronze and were the only winner from Malaysia. It was also the first time we pitched on stage together with a client at an AI innovator competition: Hernan's team and ours presented side by side.
We also want to thank our fellow Malaysian finalists, myNEWS and ZUS Coffee, for representing Malaysia alongside us. You can read more about the competition and our Lark work on our Lark Training and AI Enablement page.
The business behind the entry
Hernan is a Malaysian durian exporter known for frozen Musang King, shipping to markets across Asia, North America, Europe and Oceania. It is a traditional, physical business: fruit is graded, pre-cooled, packed and loaded into containers, and every step has to be recorded correctly before the shipment leaves.
The line on one of our pitch slides summed up the situation: "Hernan's problem was never demand. It was too manual." Orders, pre-shipment checks, grading, pre-cooling, packing and delivery ran across Lark Sheet, WhatsApp, paper forms and a lot of re-keying. Photos were keyed in by one admin. Sales orders, inventory records and monthly reports all took staff time that added up over a year.
None of this is unusual. Many Malaysian manufacturers, distributors and F&B businesses run on the same mix of group chats, spreadsheets and paper. That is why the entry is a useful example.
How the agent works
The Hernan Operation Assistant is an AI agent built with Claude, by Anthropic, that works inside Hernan's Lark group chats. The flow is simple:
- Staff post what they already have. A photo of a form or label, or an order file, goes into the usual group, and the person tags the agent.
- The agent reads it. It extracts the details that matter, such as grade, weight and quantity.
- It asks when unsure. If a value is unclear or missing, the agent asks in the chat instead of guessing.
- It saves one standard record. The data goes into Lark Base, linked to the right sales order.
- Everyone downstream is told. Lark notifies the teams that need to act next.
- Anyone can ask. Staff can ask the agent for the status of an order, stock or a stage, in plain language.
Today the agent covers the midstream stages, from sales order to container loading. The next stages are distribution and export logistics, the customer interface and traceability.
If the idea of an "agent" is still fuzzy, our explainer on what agentic AI is covers the difference between an assistant that answers and an agent that acts.
Five lessons for Malaysian SMEs
1. Start from a manual workflow, not from the technology
The project did not begin with "where can we use AI?" It began with a specific chain of work that was slow, repetitive and error-prone: recording what happens to each order as it moves from sales to loading. When you can describe the current process step by step, including who re-types what and where, you have a good candidate. If you cannot describe it, you are not ready to automate it. Our guide on how to run an AI pilot walks through choosing that first workflow.
2. Meet staff in the chat tool they already use
Operations staff on a packing floor will not open a new app and fill in a long form. They will, however, snap a photo and post it in the group they already use all day. Putting the agent inside Lark group chats meant no new screen to learn and very little change to habits. Adoption is usually the hardest part of any AI project, and the easiest way to win it is to ask people to change as little as possible.
3. Make one standard record the goal
The real value is not that the agent reads photos. It is that every stage ends up as one consistent record in Lark Base, linked to the sales order. Once the data has the same shape every time, notifications, status checks and reports become straightforward. Before you build anything, agree what that standard record looks like: which fields, which values are allowed, and what links to what.
4. Design the agent to ask when unsure
An agent that guesses quietly is dangerous in operations, because a wrong grade or weight flows straight into shipping and invoicing. The Hernan agent asks a person in the chat when something is unclear. This is a simple form of human-in-the-loop design, and it is what makes staff willing to trust the record. Our article on human-in-the-loop AI explains how to decide where people should stay in control.
5. Go live on part of the chain, then extend
The agent did not try to cover the whole business at once. It went live on the midstream stages first, with logistics, customer interface and traceability planned next. Proving value on one connected section, then extending it, is less risky than a big launch, and it gives the team time to adjust the record and the questions the agent asks.
What this means for your business
You do not need to be a technology company to benefit from an AI agent. You need a workflow that is clearly manual, a chat tool your staff already live in, a shared place to keep records, and people who understand both the business and the AI. Hernan's own team owned the process; our engineers worked alongside them to build the agent. That pairing, business experts plus builders working together in the client's day-to-day tools, is how we approach AI forward deployed engineering.
If you are an SME owner wondering where to begin, read AI for SMEs in Malaysia: where to start, then list the three tasks that cost your team the most re-keying. Those are often the best places to start.
Agmo Studio, a Select Partner of the Anthropic Claude Partner Network, delivers Lark plus AI implementation for manufacturers, retailers and F&B businesses, and Agmo Academy trains teams to use Claude well. If you want your own team to build and run workflows like this, our in-house corporate training can be shaped around your processes, and you can see other client work on our case studies page.