NovAutomate the handoff before you automate the voice
AI is often introduced through content generation. The first automation target is usually a place where information is repeatedly copied, delayed or lost.
This article reflects the named expert’s practical perspective. See NovAsia’s editorial policy for how material is prepared and reviewed.
When a company says it wants AI, content is usually the easiest demo: generate posts, draft emails, write replies. It is visible, impressive and immediately understandable. That makes it a natural place to start a presentation and, in my view, often the wrong place to start the work.
The first automation target I look for is less glamorous: where does information get copied, delayed, dropped or reconstructed by hand? A lead arrives in one tool and somebody retypes it into another. A call ends and nobody creates the next task. A customer asks a recurring question whose answer exists somewhere, but the team searches for it again every time.
Those failures are boring. Fixing them can remove hours of labour without asking software to impersonate the company.
Map the handoff before choosing the technology
A process diagram usually shows the clean version of work. The useful version is the one a real request follows on a normal day.
Take one inquiry and trace it. Where did it arrive? Which fields came with it? Who noticed it? What did they have to copy? Which system became the source of truth? What happened after the first conversation? How did the next person know they owned the task?
This exercise often reveals that the obvious problem is downstream of another one. A team may ask for automatic customer replies because response times are slow. The actual delay may be that staff spend ten minutes finding previous messages, checking a spreadsheet and asking a colleague who owns the account. Generating a faster sentence does not repair that workflow.
A useful first automation can usually be described without technical vocabulary. “When this form is submitted, create the task for the right owner and include the customer’s original question.” “After this call, create a follow-up with the agreed date.” “If the same request has an approved answer, surface it before someone searches manually.” The value is clear even if nobody mentions AI.
Moving information is different from making a decision
There is a meaningful risk boundary between transporting context and exercising judgment.
Software can classify an inquiry, attach relevant records, draft a response or flag a missed deadline. Deciding what promise to make, which customer to prioritise, whether an exception should be granted, or what commercial term to offer is a different category of action.
The closer automation gets to money, reputation or a relationship, the more clearly I want a human control point. That is not a claim that automated systems are inherently unreliable. It is simply an acknowledgement that the cost of an unnoticed error rises as the decision becomes more consequential.
A misapplied internal tag may be irritating. A confident but incorrect commitment to a customer can trigger days of repair work.
For that reason, I like to separate low-cost reversible actions from high-cost decisions. Automate the first group aggressively if the process is stable. Make the second group transparent before trying to remove the person from it.
The best automation removes work without hiding accountability
Teams sometimes count the number of automated steps as proof of progress. That is a weak measure. The useful question is what repetitive work disappeared and what failure became less likely.
Suppose a manager previously copied a lead into the customer system, assigned an owner and set a reminder. Automation now performs all three. That sounds successful. But if attachments are occasionally lost or the owner is wrong in one out of ten cases, the team may spend the saved time cleaning up exceptions.
So after launch I care about failure rates as much as time saved. How many requests reached the correct person with complete context? How often did a human have to repair the workflow? Did time to the next meaningful action fall? Did fewer tasks vanish between systems?
Those measures tell me whether the automation improved the operation rather than merely making the diagram look modern.
Voice comes later because voice depends on boundaries
Content generation becomes safer and more useful once a company has decided what its voice actually means. Many teams have never documented that. One employee writes formally, another is casual, another makes promises the business would rather not make. A model introduced into that environment does not flatten a stable brand voice; it scales an unresolved one.
The more useful sequence is to stabilise the facts, the approved claims, the ownership of decisions and the handoff of customer context. Then generated drafts can save time because a person is reviewing something grounded in the right information.
That is a very different use of AI from asking it to “sound like us” before the company can explain what “us” means.
My favourite first automation is the one nobody celebrates a month later. The team has simply stopped noticing a repetitive failure because it no longer happens. That is a better foundation for sophisticated automation than a hundred generated posts built on top of a broken handoff.