A company can have everyone using ChatGPT and still have no AI capability at all. Individual use becomes a business capability only when a recurring piece of work is owned, repeatable, measurable and able to survive the end of a chat.
That distinction matters in Malaysia now. Digital access is widespread: the Department of Statistics Malaysia reported that 97.2% of establishments used computers and 95.3% had internet access in 2024.1
But access is the beginning of the story, not the outcome. A recent ISEAS perspective makes the same broader point: digital use in Malaysian operations remains lower than front-end digitalisation, while generative-AI use is uneven.2
ChatGPT use is not the same as a changed workflow
I do not mean that individual use is worthless.
Someone who uses ChatGPT to improve a difficult email, prepare a meeting, review a proposal or understand a document may save time and make better decisions. That is real value. It is also personal productivity. The person remains the process.
The value disappears when they leave, forget their prompt, change their preferred tool or simply have a busy week.
A business capability is different. It changes the way a known piece of work gets done.
Take lead follow-up. If each salesperson asks ChatGPT to improve a reply when they remember, the company has enthusiastic AI users. It does not have a lead-follow-up system.
A capability begins when the business can answer less glamorous questions:
- What counts as a new lead?
- Where is the record of truth?
- Who owns the first reply?
- What should happen if there is no response after three days?
- Which messages require human approval?
- How will we know whether the process improves conversion, response time or customer experience?
The model may help write the message. But the capability is the agreed operating design around it.
The useful unit of adoption is a workflow
This is why I would stop measuring AI adoption by licences purchased, training sessions delivered or the number of people who say they use a chatbot.
The more useful unit is a workflow: a repeated sequence with an input, a decision, an owner, an output and a failure mode.
A small services business might choose one of these first:
- qualifying enquiries before a human callback
- assembling a weekly pipeline exception report
- extracting known fields from client documents for review
- preparing first-draft follow-ups from approved context
- monitoring renewals that would otherwise be missed
That also makes the project safer. A team can decide what data the system may read, where it may write, which outputs need review and who handles exceptions. If the process is not clear enough to describe in those terms, it is probably too early to automate or agentise it.
Capability has to survive ordinary business life
The strongest argument against this approach is reasonable: formalising every useful use of AI creates bureaucracy. Sometimes the fastest route is simply to let a capable person ask a capable model for help.
I agree. Not every task needs a workflow, an automation or a dashboard. Novel problems, difficult thinking and one-off writing often belong in a chat. The mistake is insisting that a chat is enough after the work becomes repetitive, customer-facing or consequential.
That is where informal use becomes fragile.
The useful prompt is not shared. Two people give a prospect contradictory answers. A customer record is updated in one system but not another. A report arrives late because the only person who knows how to produce it is on leave.
The remedy is operational discipline: define the work, keep the important record in the right place, make approval visible and build only as much automation as the process can support.
That is the work behind AI adoption. It is also why automation delivery should begin with a problem that already happens often enough to deserve a reliable answer.
Start with one failure you can name
I would not begin by giving an AI assistant access to every inbox, drive and customer system. I would begin with one recurring failure that a team can describe precisely.
“Leads are sometimes followed up too late” is a better starting point than “we need AI.” So is “monthly status reports take three days to assemble” or “we cannot see which client documents are incomplete.”
Then run a simple test:
- Describe the current input, owner and output.
- Decide what good looks like in a measurable way.
- Specify the exceptions and approval points.
- Test the smallest change that could improve it.
- Keep a record of what actually changed.
The last step matters. A full calendar, a polished demo or a team using the latest tool are not proof of a capability. A capability earns the name when it makes a known piece of work more dependable without creating an unowned mess somewhere else.
Malaysia does not need fewer people experimenting with AI. It needs more teams that can turn a useful experiment into a process worth keeping.
If you have an AI idea but cannot yet name the workflow, the starting problem is not the model. It is the work.