An AI agent is a system that takes a goal, works out the steps, uses tools to carry them out, and reports back. MT Labs builds and deploys AI agents for Singapore companies on private infrastructure, so the agent runs against your own data, on a server you rent or buy from us, and nothing is handed to a third-party cloud.
The problem most Singapore businesses are hitting
Most teams we meet have already run an agent pilot. Someone connected a chatbot to a company inbox, it worked in the demo, and it quietly stopped being used. The reason is almost never the model. It is that the agent had no access to the company’s real knowledge, no memory between conversations, and no safe way to touch the systems where the work actually happens.
The second problem is where the data goes. An agent is only useful once it can read your documents, your customer history and your internal notes. On a public cloud service, that is exactly the material you are sending to someone else’s infrastructure, and it is the material your PDPA obligations care most about. Plenty of Singapore firms have stalled a rollout at precisely this point, and they were right to.
The third is cost shape. Cloud agent platforms bill per seat or per action, so the moment an agent becomes genuinely useful and the whole team starts using it, the bill grows with adoption. That is a strange incentive to build a business process on.
How we build AI agents
We deploy agents onto a private AI server that you rent or buy from us. The models run locally, so prompts, documents and conversation history stay on that machine. We work with current open models such as Mistral, Qwen and Gemma, chosen per task rather than defaulting to the largest one available.
An agent becomes useful when three things are in place. It needs retrieval over your own documents, so answers come from your material and not from general web knowledge. It needs memory, so a conversation on Thursday knows what was agreed on Monday. And it needs tools, which is where an agent stops being a chat window and starts doing work: searching, scheduling, reading a file, updating a record.
Because the cost sits with the server rather than a per seat licence, every person in the company can use the agents on it at the same flat cost. Adding your whole team does not change the price.
What it looks like in practice
Three of our own agents are the clearest illustration.
Agent Kenzo is a WhatsApp agent. It connects to a business number, answers from a knowledge base built out of your own documents, transcribes voice notes locally, remembers previous conversations with each contact, and can send scheduled or proactive messages. It is the shape most Singapore SMEs want first, because WhatsApp is where their customers already are, and because the conversations never leave the building.
AgentsCommand is the console for running several agents at once. Workflows are built as nodes, so you can see what each agent is doing, where a task is waiting, and which step failed. Once you are past a single assistant and into a handful of them, the operational question stops being capability and becomes visibility.
ContentMachine is the same idea applied to publishing. Agent tasks run on a schedule, each one picking up the next brief, drafting it and reporting back through a mission control view that shows what ran, what is in progress and what needs a person to look at it. It is the clearest example of agents doing repeated work without someone driving every step.
Both come preloaded on the server alongside the rest of the suite, so an agent deployment is not a separate procurement exercise.
Where an AI agent is not the answer
If a process is fully deterministic and already scripted, an agent adds latency and a failure mode for no benefit. Write the automation instead. We will say so.
If the work requires a judgement call with legal or financial consequences, an agent should prepare the decision, not make it. The useful pattern is an agent that gathers, drafts and routes, with a person approving.
And if your documents are scattered, contradictory or years out of date, retrieval will surface that faster than anything else. That is worth knowing, but it means the first phase of the project is your knowledge base, not the agent.
Agents also need supervision in the first weeks. Anyone promising a system you configure once and never look at again is selling something we would not.
MT Labs helps companies across Singapore deploy AI tools they actually own. Private infrastructure, no recurring cloud subscriptions, and a setup built around how your team already works. Whether you want a single WhatsApp agent answering customers or a set of agents running an internal process, we will size it to your team. Get in touch and let’s figure out what makes sense for your business.
Agents run on the same infrastructure as the rest of the suite. See our private AI server plans.
FAQ
What does an AI agent actually do for a business?
It takes a goal, works out the steps, uses tools to carry them out and reports back. In practice that means answering customer messages from your own documents, preparing drafts, scheduling follow-ups, pulling information out of files and updating records, rather than simply chatting.
Where does our data go if we deploy an AI agent?
On an MT Labs deployment it stays on the private server you rent or buy from us. The models run locally, so prompts, documents and conversation history are not sent to a third-party cloud. That is usually the deciding factor for teams with PDPA obligations.
Do you charge per user for AI agents?
No. The cost sits with the server, so every person in the company can use the agents on it at the same flat cost. Adding your whole team does not change the price.
Can an AI agent work over WhatsApp?
Yes. Agent Kenzo connects to a business WhatsApp number, answers from a knowledge base built out of your own documents, transcribes voice notes locally, remembers previous conversations with each contact and can send scheduled messages.
How long does an AI agent deployment take?
A single focused agent is usually a matter of weeks rather than months. The variable is rarely the model. It is how ready your documents are, because retrieval quality depends entirely on the material the agent is given.
When is an AI agent the wrong tool?
When the process is fully deterministic and already scripted, a normal automation is better. When the decision carries legal or financial consequences, the agent should prepare it and a person should approve it. We will tell you when that is the case.
