Kyndryl says 72% of Singapore firms struggle to scale their AI pilots. Tech in Asia reported it on 12 November 2026. The same coverage notes that only 17% of business and technology leaders surveyed in Singapore said they were completely ready for external risks. The global figure is 24%.
I could not load the full article, so I am working from the headline and the summary only. I won’t quote anything beyond those numbers.
The excitement around it
The usual reading goes like this. The models are ready, but companies lack skills, data or governance, so pilots stall. Buy better tools, hire more specialists, and the pilots will scale.
I test that reading with one question first. Is there a real need? For most of these companies, yes. They have work that is slow, repetitive and expensive. The need was never the problem.
So what actually changed? The models got good enough to do real work. What didn’t change is how the company runs. The same people do the same jobs in the same order. A pilot sits next to that work and never replaces any of it.
A pilot is a demo of a tool
A typical pilot picks a tool, gives a few enthusiastic people access, and measures whether they like it. Everyone agrees it is impressive. Then someone asks what happens next, and nobody can answer. That is because the pilot was never attached to a process.
Technology is the easy part. Changing how people think and work is the hard part, even for a tech team. A pilot avoids that hard part by design. Nobody has to change a habit to run a demo.
The unit of work is the workflow, not the prompt. A prompt produces one good answer. A workflow takes a request in, does the steps, checks the result and hands it to the next person or system. Scaling means the second thing happens every week without someone nursing it.
What a scalable pilot needs
Three things have to be named before the work starts.
- The process. Write down each step, from the trigger to the finished result.
- The owner. One person is responsible for how it runs and for fixing the brief when the output is wrong.
- The number. Pick one measure the workflow should move, such as reply time, bookings caught or hours spent on admin.
If any of the three is missing, the pilot has no way to succeed or fail. It can only be admired. That is why I think many of the 72% are stuck. Their pilots never had a way to graduate.
Take a clinic that wants AI to help with appointments. A pilot might let the front desk try a chatbot. A workflow has a name: “new enquiry to confirmed booking.” It has an owner, perhaps the practice manager. It has a number, such as how many after-hours enquiries become bookings. Only the second version can be switched on for good, because everyone knows what it is supposed to do.
The same applies to a shop answering order-status questions, a service business chasing quotes, or a marketing team producing weekly reports. In each case the job already exists and already runs every week.
Where this shows up in what I build
AeroChat is built around this idea. You describe the customer service or sales workflow you want automated, and the system translates it into a set of instructions and runs it. It connects to your own data systems, so it can perform actions and not only answer questions. I built it that way because a chat window that answers well still leaves the process untouched.
What a business should do
Pick one workflow that already runs every week. Choose something with a clear start and finish, where the cost of a mistake is low enough to learn from.
Write down the steps, name the owner and choose the number. Automate it end to end, so a person is no longer the glue between the steps. Rules where a wrong answer is costly stay in the system. AI handles the parts that need judgement.
Run it for a few weeks and check the output against the number. Fix the brief when it falls short. Only when that workflow runs reliably should you start a second pilot.
Source: Tech in Asia: Kyndryl says 72% of Singapore firms struggle to scale AI Pilots