AI Employees in Production for SMBs

“I ran an update on my demo and I broke it,” John said. “I told my AI agents to go fix it, but I don’t know if it’s fully fixed.”

Michael had just introduced him, noting this talk was a major update to his last visit in January. Now, John was about to demo Hachi, his company’s AI employee, live over WhatsApp. He explained that the system waits 20 seconds to process messages, a small accommodation for how people actually talk: in short bursts.

As he spoke, a new message appeared on the screen from Hachi.

John looked at the screen, surprised. “Oh, hello. Wow. Nice. It fixed itself.” The demo worked. The moment landed the audience inside a real builder’s session. The system wasn’t being sold as magic. It was being shown mid-repair, a testament to the messy reality of shipping real AI infrastructure. It was a system John’s own company, Hasky, uses to run administration, with the goal of maintaining zero administrative staff even as the company scales.

"We are a small company and cannot afford to hire an entire team of administrative staff. For John, he believes that AI can do the work, and if they setup from the get-go for the company to rely on AI-native processes, the goal is very achievable.

From 100x Cheaper Software to a New Bottleneck

When Opus 4.5 with its 200k context window arrived in November of last year, John saw that the economics of enterprise software had changed forever. He had spent years in the field, moving from finance to AI Singapore to becoming director of AI engineering for a Temasek portfolio company. He knew the traditional project model was slow and expensive. The new models could build custom software that was “at least 100 times faster” and “at least 100 times cheaper.”

In his mind, he believed that this shift made enterprise-grade solutions economically viable for small and medium-sized enterprises.

The room he is speaking to is filled with people who have already navigated this shift. There are builders working on AI governance, advanced manufacturing software, and tools for the sustainable materials industry. Around them are public equity investors, fractional CFOs, and operators building autonomous companies.

For John, the sweet spot was clear: SMEs with five to twenty million dollars in revenue, whose operations live in Google Drive and who have no legacy system. For them, solving the cost problem created a new one. With software cheap and abundant, the bottleneck moved from price to architecture.

SME problems typically fall into four domains: CRM, ERP, operations, and WhatsApp as the communication backbone. The challenge was no longer just building a single tool, but designing a unified system where a company’s projects, documents, and history could be organized into a single source of truth.

The goal was to build a second brain for the entire company. This system needed to be more than a passive knowledge base. It had to be active. “AI employee to us is really just a proactive second brain,” John explained.

The Three Walls of Enterprise AI

The path to building that proactive second brain, what John’s team calls Hasky OS, meant hitting three distinct walls.

The first was the trap of single-user architecture. John’s team tried using tools powerful for individuals, like OpenClaw, but they failed at the company level. When every employee has their own personal agent, John noted, “all the data kind of gets stuck in a silo, and you don’t get any shared knowledge.” An institutional memory can’t form. A multi-user system requires a fundamentally different architecture, including role-based access control. As John noted, if you are putting in HR documents, only HR should see them. The same goes for legal.

Sitting in this room, watching builders nod along, I see the evidence for why we gather. The senior operator managing a legacy system and the founder just starting out are wrestling with the same questions of architecture and control. This is the conversation we show up for.

The second wall was the document-as-file problem. Most of a company's knowledge exists in documents, but treating them as files means information gets lost. The solution was to redefine what a document is. “A document is not a PDF,” John argued. “A document is an entity that has information inside.” This entity can be encoded as a PDF or Word document, but at its core, it is structured data. This allows for lifecycle management and reliable parsing by LLMs without the high token cost of reading raw files.

The third wall was the brittleness of business automation. Processes are rarely simple. As John described, a typical SME workflow involves sales handing off to contracting, which hands off to procurement, which hands off to operations, often over WhatsApp and with many exceptions. A simple rules-based system breaks. An LLM is needed for judgment, but that introduces its own challenges. The key is a framework that allows an LLM to operate reliably, with the ability to audit and fix issues quickly.

Hasky's Design: The Kernel, Graph, and WhatsApp Layer

To break through these walls, Hasky OS was built on three core architectural ideas.

The first is a shared knowledge platform with a kernel, an entity graph, and access control. The kernel is a small, open core that can be customized. Business logic lives in pluggable apps, not the kernel itself, because as John said, “everybody's business logic is different” and might change every three months. Every piece of information, from a client to a project, is an entity, and an LLM dynamically links them into a knowledge graph.

The second idea is a more sophisticated approach to memory. John’s team modeled their system on human memory. As he explained it while looking at me, when "I look at Michael, my brain does not parse through a list."

Instead, “15 or 20 things just appear” at once. He believes this is analogous to an embedding. They use a platform called Graphiti to build and manage a knowledge graph that allows their agent to retrieve context quickly from a vast memory store. For managing workflows, they use LangGraph to define the multi-step tracks an agent follows.

Hachi, their agent, runs on four tracks:

1. respond to direct messages

2. monitor group chats for key information

3. synthesize new knowledge into the graph

4. patrol the company’s infrastructure.

The final piece was the interface: Where do the staff already live? John’s clients were clear: “my staff don't use Claude code. They use WhatsApp. It's in my pocket 24-7,”. So HaskyOS was designed to bring the technology to live within existing usage habits.

Where AI Belongs and Where It Does Not

A panel discussion just days earlier on enterprise AI adoption surfaced a parallel finding: the hardest part of transformation is not the technology itself, but aligning the people and processes around it. In the Q&A, one builder in the room shared a story that echoed this. He had built a system for what was “like a shipping company,” but the more he tested the AI, the more mistakes he found. “I end up just switching to deterministic non AI components,” he said.

John agreed. “Don’t use bazooka to kill mosquito,” he said, repeating a phrase he loves. His team’s process is to build with “dumb deterministic software and only when rules don't work then you put an AI layer on top.” Building the deterministic software first forces a rigorous definition of the business logic, which in turn narrows the choices for the AI and improves its reliability. He also clarified that his team is not selling a product, but a custom service.

For builders, this is the new frontier: making architectural decisions about what becomes deterministic infrastructure, what becomes graph memory, and what gets handed to an LLM for judgment. The most practical guidance came from the numbers.

The cutoff for agentic tool use is a 200 billion parameter model. Hachi runs on something at the level of DeepSeek 3.2, which can be hosted for around $1,500 a month. Even on bigger parameter model like a GLM 5.2, John estimates the cost to be around $15,000/month. A very acceptable cost, given the increasing cost of labour in the region.

The work is not just about what is possible, but what is practical, reliable, and lives inside the tools a team already uses.

---

SQ Collective hosts Coworking Fridays for founders, operators, and AI builders working through real product questions in Singapore.

Join an upcoming Coworking Friday: https://lu.ma/sqc-friday Explore SQ Collective: https://www.sq-collective.com

Michael