Why AI Projects Fail in Indonesian Companies Even With Advanced Technology

The first mistake often happens before a company chooses its AI model. Teams see what ChatGPT, Gemini, Claude, computer vision, speech-to-text, or other AI technologies can do and immediately imagine automating major business processes. The problem is that an AI system can only perform reliably when its context, data, business rules, and integrations are reliable. A powerful model can still produce an operationally useless answer when the underlying business data is inconsistent.

Consider an Indonesian distributor with 20,000 SKUs. The same product may appear as "PVC 3 mm" in one system, "PVC-003" in the warehouse, and a completely different marketing name on an online marketplace. When the company wants AI to answer stock questions or prepare quotations, the main problem is not the language model. The AI first needs to understand that these different records refer to the same product.

The same problem appears in WhatsApp customer service. A company may want an AI assistant to respond 24/7, but prices are stored in Excel, inventory sits in an ERP, shipping information comes from a marketplace or logistics provider, and return policies are known only by a supervisor. The AI can produce a fluent response while still lacking the operational data required to give the customer a correct answer.

For AI implementation, we usually separate three questions: Is the business problem clearly defined? Is the required data available and trustworthy? Can the AI be connected to the systems employees already use? If one answer is "not yet," buying a more expensive model is rarely the best first step.

NIST's AI Risk Management Framework similarly treats AI as an organizational risk-management problem involving governance, mapping, measurement, and management rather than simply selecting a model.

Many companies already have large amounts of data, but that does not mean they have AI-ready data. A five-year transaction database is not automatically a high-quality dataset. It may contain duplicate customers, duplicated products, cancelled transactions, historical prices without context, or free-text fields maintained differently by different employees.

For example, a CRM may contain 50,000 customer records. Before using AI to identify customers likely to purchase again, the company needs to determine whether one customer can have multiple records, whether WhatsApp numbers are consistent, whether marketplace and offline transactions can be connected, and whether customer status is actually maintained. Without this work, AI scoring may look sophisticated while being based on unreliable inputs.

SOPs create another major problem. Business owners often say, "The AI should follow the way our sales team works." The next question is: which sales process? Salesperson A may offer a discount of up to 10% without approval, salesperson B may require a supervisor's approval, while salesperson C uses special pricing for long-term customers. If these rules have never been formally defined, there is no single business rule for the AI to follow.

For that reason, before starting AI development, we often recommend a process audit. Take one repetitive workflow, such as quotation creation. Document the input, price validation, discount rules, approval process, PDF generation, WhatsApp delivery, and follow-up recording. Once the workflow is clear, AI can be placed where language understanding, document processing, or recommendations actually provide value.

This approach also aligns with the broader AI risk-management principle that measurement, oversight, and management should continue throughout the AI system lifecycle.

One common AI implementation mistake is treating the chatbot as the finished product. A chatbot is easy to demonstrate: enter a question and receive an answer. The real challenge begins when the chatbot has to perform actual operational work.

Imagine a customer asks through WhatsApp, "Do you still have 80 boxes of 4 mm Oak Brown SPC flooring?" To answer correctly, the AI needs access to the product master, relevant warehouse inventory, product units, and potentially incoming purchase orders. If the customer then asks, "What is the price if I buy 60 boxes?", the system needs the correct price list for that customer or sales channel. If the next question is, "Can you deliver to Bekasi on Friday?", the AI may also need access to shipping information and relevant cut-off rules.

This is no longer just a prompt-engineering problem. It is an application architecture and integration problem.

In a real implementation, AI may need APIs connecting to an ERP, CRM, database, ticketing system, WhatsApp gateway, marketplace, or internal application. For certain use cases, retrieval-augmented generation can allow the AI to retrieve information from company sources before generating an answer. For tasks that require actions, an AI agent can be given controlled tools and permissions—for example, permission to read inventory or create a quotation draft, but not change prices without approval.

The trade-off matters. If AI only answers a small number of infrequent FAQs, a complex integration may not make economic sense. If it helps a sales team process hundreds of inquiries every day, integration with the CRM and quotation system may create substantially more value than a standalone chatbot.

Do not use metrics such as "the AI is live" or "the chatbot has a certain accuracy rate" as the only definition of success. Management needs measurable business outcomes that can be compared before and after deployment.

Suppose a sales team spends an average of 20 minutes preparing a quotation. If AI can prepare a draft using customer information, products, pricing, and the company's template and reduce the employee's work to 7–10 minutes, that is a meaningful business result. Another example is customer service. If a company receives 500 questions per day and 60% are repetitive, AI does not necessarily need to handle every conversation. It may be more valuable to handle routine questions and escalate complex cases to employees together with the full conversation context.

We usually recommend establishing a baseline before development starts. Measure workload volume, processing time, error rates, escalation rates, and relevant labor costs. Once the pilot is running, compare the same metrics.

AI implementation should also be evaluated according to risk. An AI system that drafts marketing copy has a very different risk profile from one that recommends credit limits, determines discounts, or processes customer data. NIST identifies characteristics such as validity, reliability, safety, security, transparency, privacy, and bias management as important elements of trustworthy AI.

Companies do not need to wait until every system is perfect before using AI. However, they should avoid starting with the largest possible project simply because the technology is popular. A safer approach is to choose one process with sufficient volume, reasonably clear rules, available data, and measurable outcomes.

For example, a distributor could start with AI that reads WhatsApp inquiries and prepares quotation drafts. An e-commerce business could start with customer-question classification and suggested responses. A professional services company could use AI to summarize meetings, prepare follow-ups, and update CRM records. A document-heavy organization could start with an internal knowledge assistant before deploying an AI agent that can execute transactions.

At Sahabat Teknologi, the approach we recommend is to start with the business problem, then audit the process and data, and only afterward select the technology. The audit may show that a company does not actually need an AI agent yet and would benefit more from conventional automation. Tasks such as inventory synchronization, invoice generation, notifications, or moving structured data between systems are often cheaper and easier to control with deterministic workflow automation.

If the pilot works, expand gradually: from one department to several, from reading information to creating drafts, and then from drafts to approved actions. This staged approach limits risk while giving the company time to improve its data, SOPs, and governance before AI receives broader operational access.

Indonesia is also developing policy and ethical guidance around artificial intelligence. In 2025, the Ministry of Communication and Digital Affairs published a public consultation concerning the National AI Roadmap White Paper and a proposed AI Ethics Guideline.