Why Local AI Agents May Matter to Your Business
If your business uses AI to review documents, analyse sales data, prepare reports or connect information across different applications, you may already have noticed a practical problem: useful AI work is rarely limited to one short question. An agent may need to read many files, check its own work, search for information and update another system.
That creates two concerns for you as an SME owner. First, sensitive business information may need to leave your computer and travel to a cloud service. Second, long-running AI tasks can consume substantial usage credits or API allowances. A new approach from Perplexity, developed with Nvidia, points towards a different model: running AI agents directly on suitable local hardware.
TL;DR: Local AI agents process models, files and tasks on your own computer, which can improve data control and reduce dependence on cloud processing. For Malaysian SMEs, the most sensible starting point is a small, controlled workflow involving private documents or repetitive analysis—not an immediate company-wide rollout.
Perplexity’s Portable Computer is available for Linux systems with Nvidia RTX GPUs containing at least 24GB of video memory, with Windows support announced for September. Source
What This Means
A cloud AI tool sends your request and, depending on the workflow, related files or data to remote servers. The remote model performs the work and sends the result back to you. This arrangement is convenient because you do not need to manage powerful hardware, but it means your information passes through an external environment.
A local AI agent works mainly on your own machine. The model, agent instructions, files and processing remain on the device. In Perplexity’s example, the system can review documents, analyse a CSV file and prepare an output without sending every step to a cloud model. It can still ask permission to use a more powerful cloud model when a task exceeds local capability.
This is different from simply installing a chatbot. An agent can perform a sequence of actions. It may inspect a folder, identify relevant files, compare figures, produce a report and send the result to an approved business application. Portable Computer packages the model, tools, connectors, inference engine and security sandbox into one application rather than requiring users to assemble each part separately.
The product initially supports Qwen 3.8 27B and PPLX 27B, with Nvidia Nemotron 3.5 Lightning planned for a later release. Source
The important question is not whether local AI is better than cloud AI. It is whether a particular business task should keep its data and processing on your own device.
Local processing does not mean every task becomes private or risk-free. If you connect Gmail, Google Drive, GitHub, Slack or another online service, information may still move through those connectors. You also remain responsible for access controls, backups, malware protection and the security of the computer running the agent.
How This Applies to Malaysian SMEs
1. Handling sensitive customer and financial documents. A Malaysian accounting firm, property agency, distributor or professional services company may regularly work with invoices, bank statements, tax documents, contracts and identity information. You may want AI to classify documents, identify missing pages or prepare a summary, but uploading everything to an external service may not fit your internal policies or your clients’ expectations. A local agent could review documents on a designated workstation, with staff checking the final output before it enters your records.
For example, you could create a workflow that reads a folder of supplier invoices, extracts invoice numbers and dates, flags duplicates and produces a spreadsheet for approval. The agent should not automatically approve payments. Its role should be limited to preparation and checking, while a responsible employee remains in control of the financial decision.
2. Analysing sales and operations data. Many SMEs already keep useful information in spreadsheets, but the owner may not have time to examine it every week. A local agent could review sales by branch, product or salesperson; compare actual results with targets; identify unusual transactions; and draft questions for your next management meeting.
This can be particularly useful for wholesalers, retailers, restaurants and service businesses. You might ask the agent to compare weekday and weekend performance, detect items that are frequently returned or identify delays between order receipt and delivery. Since the analysis can begin with a local CSV or spreadsheet, you can test the workflow without redesigning your whole technology setup.
3. Preparing internal reports and standard operating procedures. If your business has employees who repeatedly search through folders for policies, product specifications or customer-service instructions, a local agent can help organise and summarise that material. It could draft a new SOP from several existing documents, compare two versions of a procedure or create a checklist for onboarding staff.
This is suitable for companies that want practical assistance but do not want an employee pasting confidential operational information into a public chatbot. You should still review every draft. An AI-generated procedure may omit an important exception, misunderstand a Malaysian regulatory requirement or use outdated company information.
4. Supporting technical or semi-technical work. Some SMEs have an internal administrator, e-commerce manager or developer who works with website files, product catalogues and scripts. A local agent may help explain code, clean data, review configuration files or prepare a draft change. Keeping the work on the company machine can reduce unnecessary exposure of proprietary systems.
However, avoid giving the agent unrestricted access to production systems. Use a sandbox, separate test environment and a clear approval process. Perplexity says its local harness disables itself when the operating-system sandbox is unavailable, which illustrates an important principle: an agent should not run business tools with full computer permissions by default. Source
What the Reported Results Tell You
Perplexity reported results from its internal Local Knowledge Work Bench, which contains 53 tasks covering research, financial analysis and document creation. Qwen 3.8 27B running through its Computer harness scored 82.6%, while the same model using the Pi harness scored 77.6% and Hermes scored 74.0%. PPLX 27B scored 85.4%. Source
| Reported measure | Computer with Qwen 3.8 27B | Comparison |
|---|---|---|
| Local Knowledge Work Bench | 82.6% | Pi: 77.6%; Hermes: 74.0% |
| Local Knowledge Work Bench with PPLX 27B | 85.4% | Perplexity’s reported result |
| BrowseComp research benchmark | 66.7% | Pi: 50.2%; Hermes: 43.9% |
| Multimodal document understanding | 65.1% | Hermes: 34.6%; Pi: 13.9% |
These are company-reported results, so you should not treat them as a guarantee for your own documents or workflows. They do show why the surrounding agent design matters. A smaller model may perform better when it has a focused set of tools, shorter instructions and strong verification steps, rather than being connected to everything at once.
Practical Takeaways
- Start with one low-risk workflow. Choose document classification, report drafting or spreadsheet analysis before attempting automated financial transactions.
- Separate private data from general data. Identify which documents must remain on the local machine and which information is acceptable for cloud processing.
- Check your hardware first. The launch threshold is an Nvidia RTX GPU with at least 24GB of VRAM, approximately the level of an RTX 3090 or newer. Source
- Use a dedicated business computer. Avoid running an agent on a personal laptop containing unrelated files, browser sessions or saved passwords.
- Keep human approval in the process. AI may prepare, compare and recommend, but your team should approve payments, customer messages, legal documents and changes to live systems.
- Require a sandbox. The agent should not receive unrestricted access to your operating system or company applications.
- Measure useful outcomes. Track review time, correction rates, missed issues and staff adoption—not just whether the model produces an impressive demonstration.
- Document the workflow. Record which files the agent can access, which tools it can use and who reviews its output.
The Bigger Picture
Cloud AI will remain useful, especially when you need the strongest available model, broad web research or no local hardware management. Local AI is not a replacement for it. The more practical direction is a hybrid setup: routine and sensitive processing happens on your device, while selected complex tasks are sent to a cloud model only with approval.
For SMEs, this could make AI easier to govern. Instead of allowing staff to use disconnected tools without clear rules, you can define which workflows run locally, which applications can be connected and when escalation is permitted. That gives you a clearer way to manage privacy, access and business continuity.
The hardware requirement also matters. A local agent is not automatically suitable just because the software is available. You need reliable hardware, sufficient storage, updates, user permissions, backup procedures and someone responsible for maintaining the setup. If your team cannot support those basics, a managed cloud workflow may still be the more practical choice.
The strongest opportunity is therefore not chasing the newest AI application. It is identifying work that currently consumes staff time and deciding where that work should happen. For your business, the right starting point may be a private document review process, a weekly sales analysis or an internal knowledge assistant. Test one workflow, protect the data, verify the results and expand only when the process proves dependable.
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 →