What If Your Business AI Tools Could Work Together?

You may already use several digital tools in your business: one for customer enquiries, another for stock records, a separate accounting system, and perhaps an AI assistant for drafting replies. The problem is that these tools often work in isolation. You still need to copy information between them, check whether the answer is reliable, and step in when one system does not understand the situation.

That creates a familiar burden for Malaysian SME owners. Your team may be small, but every customer expects a quick response, accurate information, and consistent service. You need automation that helps people make decisions without forcing you to rebuild your entire business around one expensive platform.

A recent development from Russian AI startup Mostik points towards another approach: allowing different AI models to share useful internal information without one model first converting everything into ordinary text. The idea could eventually help smaller businesses combine specialised AI tools more efficiently.

TL;DR: Different AI models may soon cooperate through their internal mathematical representations instead of passing text from one to another. For your business, this could mean smaller, specialised tools working together while keeping workflows faster and easier to manage.

The approach is still emerging, so you should not treat it as a ready-made solution for every SME. However, it offers a useful way to think about your next automation project: do not ask one AI model to do everything when a small group of focused tools may do the work more effectively.

What This Means

AI models are systems trained to identify patterns and produce useful outputs. Their internal “weights” are mathematical values that influence how they respond to a prompt. Mostik’s approach attempts to connect models through these values, rather than requiring one model to produce text that another model must read and interpret.

Think of two employees with different strengths. One understands technical product details, while the other is good at communicating with customers. In a normal workflow, the first employee writes a message, and the second reads it before responding. That works, but it takes time and can introduce misunderstandings.

A model-to-model connection is more like sharing structured internal signals directly. The models may not “talk” using sentences, but one model’s capabilities can influence another model’s response. According to Wired, Mostik demonstrated a hybrid system combining a very large model with a much smaller model designed to run on a mobile device.

The example involved a 753-billion-parameter GLM model and a 4-billion-parameter Qwen model. The source reports that the hybrid system cost one-twentieth of the full GLM model and achieved performance halfway between the two. These figures come from the source article and should be treated as reported results, not a guarantee for your own business application.

The broader idea is similar to combining several opinions before making a decision. The Mostik team compared it with estimating the weight of a pig by combining guesses from several people. One person may be wrong, but a group of independent estimates can produce a better result. In machine learning, this is commonly known as an ensemble approach.

Key insight: You may not need one giant AI system for every business task. A practical combination of smaller, specialised models could be more suitable for specific workflows.

How This Applies to Malaysian SMEs

Consider a local distributor serving customers through WhatsApp, email, and a website. One AI tool could classify incoming enquiries, another could check product availability, and a third could prepare a reply in Bahasa Malaysia or English. Today, these tools may pass information through copied text or separate software integrations. A more direct model-to-model connection could help the systems share meaning with fewer handovers. You would still need approval rules, but your staff could spend less time sorting routine enquiries.

For a restaurant, café, or food manufacturer, different AI models could handle different parts of daily operations. A forecasting model could identify likely demand for ingredients, while another could analyse customer feedback and highlight recurring complaints. A language model could then prepare a simple management summary. Instead of asking one general AI tool to understand purchasing, operations, and customer service at once, you could assign each model a narrow responsibility.

Service businesses could also benefit. Imagine an air-conditioning contractor receiving a photo, location, previous service history, and a short customer message. A visual model could identify possible equipment issues, a scheduling system could check technician availability, and a communication model could prepare a polite appointment message. The systems would need to work from accurate records and follow human approval, but the concept shows how specialised tools can support the full workflow.

For Malaysian SMEs, language and context are especially important. Your customers may switch between English, Bahasa Malaysia, Mandarin, Tamil, or local expressions in the same conversation. A general model may produce a grammatically correct response that still sounds unnatural or misses business context. A smaller language-focused model, combined with a customer-service model, could potentially provide better results than relying on a single general-purpose system.

There is also a useful operational benefit. Smaller models may be easier to run on internal computers, mobile devices, or controlled business systems. The source article describes a 4-billion-parameter model capable of running on a mobile device, while noting that the hybrid performance sat between the smaller and much larger models. For an SME, this suggests a future where some tasks can happen closer to your staff or customers, rather than every request needing to go to a large external system.

A Simple Comparison

Approach	How It Works	Possible SME Use	Main Concern
Single general model	One AI handles many tasks	Drafting, summaries, basic customer replies	May lack depth in specialised work
Text-based model chain	One model writes output for another to read	Enquiry classification followed by reply generation	Extra processing and possible misunderstanding
Model ensemble	Several models contribute to one result	Risk checks, forecasting, document review	Requires careful coordination and testing
Internal model bridge	Models share mathematical representations	Future specialised automation workflows	Still an emerging technical approach

The source article reports a 753-billion-parameter model, a 4-billion-parameter model, and a hybrid cost ratio of one-twentieth. See the original Wired article for those specific figures and the wider explanation of Mostik’s demonstration.

Practical Takeaways for Your Business

• Map the workflow before choosing AI. Write down what happens from customer request to completed task. Identify where staff repeatedly copy, check, rewrite, or approve information.
• Separate tasks by capability. Use one tool for classification, another for document extraction, and another for communication when that produces clearer responsibilities.
• Start with low-risk processes. Trial AI on internal summaries, enquiry sorting, meeting notes, or product descriptions before applying it to sensitive decisions.
• Keep a human approval step. Customer replies, stock commitments, compliance statements, and technical recommendations should be reviewed until the system proves dependable.
• Test Malaysian language needs. Include real examples involving Bahasa Malaysia, mixed-language messages, local addresses, abbreviations, and customer names.
• Measure useful outcomes. Track response time, correction rates, unresolved enquiries, and staff workload rather than judging AI by impressive demonstrations.
• Protect business information. Decide what customer, supplier, employee, and financial data may be processed by each tool.
• Choose systems that can connect. Ask vendors whether their tools support application programming interfaces, exports, webhooks, or other practical integration methods.

What You Should Do Next

Choose one repetitive workflow that currently involves at least two people or software systems. Customer enquiries are often a good starting point because the process is easy to observe. Document the steps, collect a sample of real messages, and mark where mistakes happen.

Then decide which parts require judgement and which parts are repetitive. AI can classify an enquiry, extract details, or draft an answer, but your team should define when a person must intervene. This creates a safer foundation than buying a tool and hoping it will discover your process automatically.

When speaking with an automation provider, ask whether the system can use different models for different tasks. Ask how information moves between systems, where data is stored, how staff can correct mistakes, and whether you can review an audit trail. These questions matter more than whether a tool claims to use the latest model.

The Bigger Picture

The long-term direction suggested by Mostik is a move away from treating AI as one enormous system that must handle every task. Instead, businesses may use networks of smaller models, each trained or configured for a particular job. A customer-service model could work alongside a product model, a compliance model, and a forecasting model.

The source article also reports comments from experts who believe specialised models could work with frontier models and help approach large-model quality without requiring the largest model to handle the entire process. Read the full discussion at Wired.

For you, the practical lesson is not to wait for “machine telepathy” to arrive. It is to design your business systems so that useful information can move between specialised tools in a controlled way. The best automation may not look like a single clever chatbot. It may look like a quiet team of digital assistants, each handling one responsibility and passing reliable signals to the next.

That approach can make your automation easier to improve. If the inventory model makes mistakes, you can review that component without replacing the customer-service system. If your business expands into new languages or product categories, you can add a specialised capability rather than rebuilding the entire workflow.

Model-to-model communication remains a developing area, and technical claims still require independent testing. But the business principle is already useful: specialisation, coordination, and human oversight can be more practical than asking one system to do everything.

Ready to Streamline Your Operations?

Your business should run itself. AutoRunBiz deploys AI agents to automate your daily operations — WhatsApp orders, invoicing, customer follow-ups, and accounting. Book a free 15-min ops audit to see where automation fits your business →