Your AI may not be the problem—your workflow might be
You may have tried an AI tool for customer replies, document preparation, sales follow-ups, or internal reporting and felt disappointed. The tool produces a decent answer, but it forgets important details, repeats work, misses instructions, or stops halfway through a task.
It is easy to conclude that you need a smarter AI model. However, recent research highlighted by TechCrunch suggests that the software layer around the model—the harness—can matter just as much, or even more, for tasks involving several steps. For a Malaysian SME, this is a practical lesson: better automation often comes from designing a better process, not simply selecting the most advanced AI brand.
TL;DR: An AI model is the reasoning engine, but the harness gives it memory, tools, rules, feedback, and supervision. Before changing models, review how your workflow guides the AI from start to finish.
For your business, this means focusing on clear instructions, approved data sources, human checks, and sensible escalation rules. These improvements can make an existing AI system more reliable for everyday work.
What This Means
A raw AI model is similar to a capable new employee who has general knowledge but does not yet understand your company. It can draft text, classify information, summarise documents, and suggest actions. But it may not know which customer record to check, which policy to follow, or when it should stop and ask for help.
The harness is the software wrapper around that model. It connects the model to tools, business data, memory, rules, and feedback. It controls what information the AI receives, what actions it can take, and how it responds when a task becomes unclear.
TechCrunch reported that Nvidia researchers used a custom harness with memory management and a supervising component to help Claude Opus 5 achieve a 100% score on the ARC-AGI-3 interactive reasoning benchmark. Without that harness, the same model scored 30%. Source: TechCrunch
The supervising component worked like a senior manager. It checked whether the main agent was moving in the right direction, encouraged it to reconsider dead ends, and helped it avoid repeating an unproductive path. This is important because long tasks are not just about producing one good response. They require a sequence of decisions.
For example, “prepare a monthly customer follow-up report” may involve collecting sales records, checking outstanding replies, grouping customers, drafting messages, identifying sensitive cases, and sending the final list to a manager. If the AI loses context after the second step, the final result can be incomplete even if the model itself is powerful.
Key insight: An AI model generates the thinking, but the harness determines how that thinking is organised, checked, and connected to real business work.
Why Long Tasks Need More Than a Prompt
A simple prompt can work well for a one-off activity such as “summarise this meeting.” It is less suitable for work that involves multiple systems, decisions, and approvals.
Long-horizon tasks create several risks. The AI may forget an earlier instruction, use outdated information, repeat an action, misunderstand a customer’s situation, or continue confidently after making an error. Research cited in the source article found that 19 large language models produced errors when handling long-horizon document-editing tasks. Source: TechCrunch
A good harness reduces these risks by breaking work into stages. It can tell the AI to collect information first, validate it second, prepare a draft third, and request approval before taking an external action. It can also retain a record of what has already been done.
This does not mean your SME needs to build a complicated research system. The principle is simpler: do not ask AI to “handle everything” without defining checkpoints, permissions, and a clear completion standard.
How This Applies to Malaysian SMEs
1. Customer service and WhatsApp enquiries
Suppose you operate a renovation company, tuition centre, clinic, online retailer, or service business. An AI assistant could receive enquiries through WhatsApp, identify the customer’s request, check your service information, suggest a reply, and record the next action. The model is only one part of this process.
The harness should tell the assistant which information is approved, how to handle Bahasa Malaysia and English messages, when to ask for missing details, and when to transfer the conversation to a human. If a customer asks about a customised quotation or makes a complaint, the assistant should not improvise. It should flag the conversation for review.
2. Sales follow-up
Many SME owners lose opportunities because follow-ups depend on memory. An AI workflow can review enquiries, identify leads that have not received a response, draft a personalised message, and place the conversation in a follow-up queue.
A useful harness would remember the customer’s previous questions, avoid sending duplicate messages, and apply different rules to a new enquiry, an existing customer, and a customer who has declined. You can also require approval before any message is sent. This gives you consistency without allowing the system to contact people blindly.
3. Invoicing and administration
For a trading company or professional service firm, an AI assistant might extract details from purchase orders, compare them against delivery records, prepare an invoice draft, and identify missing information. The harness should define the order of these steps and require a check before the document is issued.
This matters because administrative errors are rarely caused by one difficult question. They often happen when several small details are missed. A workflow that checks customer name, service description, tax treatment, payment terms, and supporting documents is more dependable than a single instruction asking AI to “prepare the invoice.”
4. Staff onboarding and internal knowledge
You can use AI to answer questions about leave procedures, sales scripts, product information, or operating procedures. But the assistant must have access only to current, approved documents. The harness should identify the source document, show when it was updated, and escalate questions that are not covered by company policy.
This is especially useful when your team includes part-time employees or staff working across branches. Instead of relying on informal explanations, you can provide a consistent first point of reference while keeping sensitive decisions with a manager.
Model Versus Harness: What Should You Improve First?
| Business need | Model contribution | Harness contribution |
|---|---|---|
| Drafting a customer reply | Produces clear language | Uses customer history, tone rules, and approved information |
| Preparing a report | Summarises patterns | Collects the right records and checks missing data |
| Following up leads | Writes a suitable message | Tracks timing, avoids duplicates, and requests approval |
| Handling internal questions | Explains information | Restricts answers to current company documents |
| Managing a multi-step task | Makes decisions | Maintains memory, checkpoints, tools, and escalation rules |
The source article also reported that Databricks research found the harness could significantly affect AI usage costs, with the wrong harness potentially doubling costs for the same model. Source: TechCrunch The practical message is not to chase the biggest model automatically. A poorly designed workflow may waste processing, repeat tasks, and create unnecessary reviews.
Practical Takeaways for Your Business
- Choose one repetitive workflow first. Start with lead follow-up, enquiry triage, document summaries, or staff questions.
- Write the process as steps. List what must happen first, what information is needed, and what counts as complete.
- Separate drafting from sending. Let AI prepare messages or documents, but require approval for customer-facing or sensitive actions.
- Give the AI approved sources. Use current product lists, operating procedures, service details, and internal policies.
- Build memory carefully. Store only the context needed for the task, and avoid giving broad access to unrelated records.
- Add an escalation rule. Tell the system when to stop and ask a person, especially for complaints, unusual requests, or uncertain information.
- Track failure points. Record where the workflow gets stuck, repeats work, or produces incorrect results.
- Test with real examples. Use common, incomplete, and difficult cases before allowing wider use.
- Review access permissions. An assistant should have only the tools and data required for its job.
- Improve the process before changing the model. Clearer steps may deliver more benefit than moving to a different AI provider.
The Bigger Picture
AI adoption among SMEs is likely to move from isolated chatbots towards connected assistants that can complete defined business processes. The important capability will not be simply answering a question. It will be remembering the right context, using the correct tools, checking its own work, and knowing when a human decision is required.
This places more responsibility on business owners. You will need to decide which tasks can be automated, which data is trustworthy, and which actions require approval. These are management decisions, not purely technical decisions.
Open harnesses may also give businesses more control over how an AI system behaves. Nvidia’s research argued that open agent systems allow users to adjust the tools, runtime, infrastructure, and accuracy controls around the model. Source: TechCrunch For an SME, the practical benefit is the ability to shape automation around your actual workflow instead of adapting your operations to a generic chatbot.
Start by asking a straightforward question: What does my team repeatedly do that involves several steps and clear rules? Then document that process, add checkpoints, connect only the necessary information, and measure the result. The best AI system for your business may not be the one with the most impressive model. It may be the one with the clearest workflow around it.
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 →