.jpg&w=800&c=0&s=1)
In a controlled AI demonstration, a well-framed question can produce an impressive answer. Everyday business is less accommodating. A customer may leave out a crucial detail, a company policy may contain an exception, or a routine request may become urgent halfway through the conversation.
What strikes me about AI outside controlled environments is how quickly an apparently simple task becomes a question of context, judgement and responsibility when real people enter the picture.
For a traveler caught in a disruption, the distinction is immediate. A prompt response is welcome. Having to explain the same problem again to someone who can actually help is not a welcome experience.
That is where enterprise AI deserves closer scrutiny: at the point where a system’s apparent competence meets someone’s particular circumstances. Speaking with Standard Chartered, FCM Travel, as well as global marketing agency TEAM LEWIS, a common refrain we hear is that making this encounter work demands more than a capable model. It requires an organisation which is able to supply context, manage exceptions, and account for the result.
Context has to be built in
“Understanding a traveler's question is not the same as understanding their situation,” says Abdel Abatouy, CTO and CIO Asia at FCM Travel. Its assistant, Sam, needs company policy, booking information, and the circumstances surrounding a request to respond appropriately. Across Asia-Pacific, he points to differences in languages, market requirements, and customer expectations.
The task is an engineering requirement with a human consequence. If the relevant information sits in a disconnected system, or the assistant cannot access it under the right permissions, fluency alone cannot close the gap.
John Sharratt, Global Head of Technology and Infrastructure at Standard Chartered, identifies the same difficulty from the infrastructure side. “A pilot hides the cost of becoming real,” he says.
During experimentation, teams can clean data manually, work around dependencies, and keep controls outside the process. At a global-banking scale, those accommodations meet the combined demands of integration, capacity, cost, and regulation.
Sharratt’s argument is that reliability must be built into the environment around the model: trusted data, reusable interfaces, workflows, and controls. Modernisation need not mean replacing every legacy system. Governed application programming interfaces can make existing systems accessible while controlling how agents interact with them.
For technology leaders, he said this approach changes the investment question.
“An AI model’s ability to answer a request is only one dependency. The business must also know which records it uses, what actions it can take, and who owns the process when something goes wrong.”
-John Sharratt, Global Head of Technology and Infrastructure at Standard Chartered
Sharratt is describing the foundations of reliability. The financial test comes when those foundations support work whose costs and outcomes the business can measure.
The handover is part of the service
FCM offers an early operational example. Following Sam’s global deployment in June 2026, the company says it handled more than 14,500 traveler conversations across 61 countries and 15 languages in its first two weeks.
FCM reports that it fully resolved 73 percent of conversations, while the average time per request fell from approximately seven minutes to 2.5 minutes. Of the travelers who provided feedback, 88 percent responded positively.
These are encouraging company-reported indicators from a short period. They describe resolution, speed, and feedback; they do not establish longer-term performance or a complete financial return.
Nor can a global aggregate establish reliability in every market or language. APAC leaders should examine resolution, escalation, and repeat requests at that level so an encouraging overall result does not conceal uneven service.
The design behind them matters. FCM included human involvement from the outset, with consultants helping define guardrails and escalation conditions. Sam is designed to recognise when it cannot provide a reliable response and hand the conversation to a consultant with its context intact.
That continuity is easy to overlook in an automation metric. For the person seeking help, it can make the difference between feeling supported and feeling passed around. Escalation should therefore be assessed as part of the service: did the right person receive enough information to resolve the problem?
The same principle becomes more consequential when AI begins identifying issues without waiting for a request. FCM’s Proactive Analytics examines travel data for potential cost and compliance risks, providing recommendations to customers and account managers. People remain responsible for deciding what action should follow.
“We should not confuse AI’s ability to act with its authority to act,” Abatouy says.
That distinction connects directly to Sharratt’s emphasis on governed access. An agent needs a defined remit as well as a technical connection. For enterprises operating across multiple APAC markets, permission to act should be explicit within each workflow, rather than inferred from the system’s apparent confidence.
Deloitte’s 2026 State of AI in the Enterprise report finds that only one in five companies has a mature model for governing autonomous AI agents. That gap makes clear ownership and limits on authority a practical concern as deployment expands.
Faster work still needs a route to value
There is another boundary to respect: the distance between useful activity and demonstrated business value.
TEAM LEWIS reports that approximately one-third of its employees use its internally developed agents daily through its Orbital initiative. In written information supplied for this feature, the company also reports savings from consolidating or retiring third-party subscriptions, alongside a growing commercial pipeline and early conversions associated with its AI & Growth practice.
Those are different kinds of progress. Daily usage indicates adoption, while subscription savings and conversions point to financial benefits. The information supplied does not quantify the savings, revenue, or productivity gains, so it cannot establish the size of the return. That limitation should neither erase the reported benefits nor turn adoption into a substitute for measuring them.
The wider research reinforces the need to keep these measures separate. McKinsey’s 2026 State of AI survey found that 80 percent of respondents reported improved individual productivity, while 37 percent reported a positive enterprise-level impact on earnings before interest and taxes.
The survey covered 1,719 respondents in 97 countries. These global self-reports are not an APAC-specific assessment, but they illustrate why faster individual work cannot simply be counted as financial success.
Make quality part of the economics
A business must decide what happens to the capacity released. Does it reduce backlogs, allow consultants to handle more complex cases, improve retention, or support additional revenue? Without that connection, time saved can remain scattered across employees’ days, valuable to them but difficult to trace through the organisation.
The approach also makes quality part of the economics. A quick response that requires correction, repeat contact, or an avoidable escalation carries costs that a speed metric can miss.
The practical test for APAC leaders is therefore broader than whether AI can complete a task. Can the organisation establish whether the task was completed appropriately, recognise the exceptions, and take responsibility for the next step?
Production is where that discipline must continue. Outcomes need monitoring, failures need investigation, and the boundaries between automated action and human judgment need revisiting as conditions change.
The person at the other end should not have to understand any of this architecture. They should be able to get the right help without repeatedly explaining their circumstances to the organisation.
Delivering that consistently—and showing what it improves—is a more demanding achievement than a successful demonstration.
