Connecting artificial intelligence to your systems is the hard part
Most businesses still use artificial intelligence as a chat window, cut off from their own records. The real risk begins once it is given access to look things up and act on your behalf.
You probably already use artificial intelligence (AI) as a chat window. You type a question, it answers, then forgets everything. That is useful for quick drafting, but it is limited. The AI model only knows what you told it in that message. It cannot check your stock levels, confirm an order or search your customer records for you.
Giving an AI model access to tools changes that. It means connecting the model to your own systems, so it can look things up, and in some cases take an action, instead of guessing from general knowledge. A customer asks when an order will arrive, and the model checks your order system rather than inventing a date. A manager asks for last month's sales figure, and the model pulls the real number from your accounts rather than estimating one.
This sounds straightforward, but few organisations have built it properly. A recent survey of business leaders found that just 24% of organisations run AI agents able to decide on a goal, reason about it and act across more than one system, and only 15% pair that with the coordination needed to see a customer request through from first contact to resolution (SiliconANGLE). Most AI use today still stops at the chat window.
Where it goes wrong
The main worry is not that the model will misbehave. It is that it will sound certain while being wrong. Concern about the accuracy or quality of AI outputs is already the top barrier to wider AI use at work, named by workers more than any other issue (The Business Times). If you connect that same model to a live system, a wrong answer is no longer just an awkward reply. It becomes a wrong order, a wrong refund, or a wrong figure sent to a customer.
Only 15% of organisations pair AI agents that can set a goal and act with the coordination needed to resolve a customer request from start to finish ([SiliconANGLE](https://siliconangle.com/2026/10/02/ai-in-customer-experience-has-an-orchestration-problem-not-an-adoption-problem)).
The gap between a model that can chat and one that can act safely across several systems is wide. Closing it needs clear limits on what the model can see, what it can change and when a person must check its work before anything reaches a customer.
What to ask a supplier
Before you let any AI model touch your systems, ask exactly what it is allowed to read, and ask separately what it is allowed to change. Ask what it does when it is unsure. A model that can say it does not know, and hand the question to a person, is safer than one built to always produce an answer.
Adoption already varies by sector in Malaysia. Financial services businesses report 53% AI adoption and manufacturers 50%, against 38% across all businesses (The Vibes). Malaysia's ministry responsible for digital policy hopes the 2027 national budget will include measures to make AI tools more accessible for small and medium enterprises (The Sun Malaysia). That kind of support may lower the cost of getting started, but it will not remove your responsibility to check, before anything is connected, exactly what the model can see and do on your behalf.
Sources
- AI in customer experience has an orchestration problem, not an adoption problem, SiliconANGLE, 3 October 2026
- Singapore workers embrace AI faster than global peers, but fewer feel secure in their jobs: PwC survey, The Business Times, 12 October 2026
- Malaysian workers race ahead of employers in AI adoption, The Vibes, 20 September 2026
- Budget 2027: Digital Ministry hopes for affordable AI tools, MSME adoption, The Sun Malaysia, 28 September 2026
A question this article does not answer? Email hello@yunaristech.com.