Ask a founder what protects their AI product and you will usually hear about the model, the data, or the volume of something. Those are the easiest things to say and the easiest things for someone else to acquire. Models are rented. Data can be bought or licensed. Volume is a fact about today.
What tends to hold a position is duller and harder to copy. It is the work that makes an output trustworthy, and the workflow that makes a customer stop paying somebody else to do the job.
Capability is not the constraint
For most knowledge work tasks, raw capability stopped being the bottleneck some time ago. Summarising a long document, extracting figures from a report, comparing two contracts: these are close to solved and getting cheaper every quarter. A team that wins on capability alone is holding a position that erodes with the next model release.
This is why so many products that demo brilliantly do not get adopted. The demo shows the capability. The adoption decision is about something else.
The checking problem
Here is the pattern we keep seeing inside real companies. A team adopts a tool that produces an answer in seconds. Then somebody senior asks where the answer came from, and a person spends forty minutes going back through the source material to confirm it.
The tool saved seconds and created an hour. Nobody says this out loud, because the tool is new and the team is meant to be enthusiastic about it, but usage quietly drops off and the renewal conversation goes badly.
Anything that automates knowledge work runs into this. In regulated or professional contexts it is not a preference, it is a hard requirement. A compliance officer cannot act on an answer they cannot trace. Neither can an analyst whose name goes on the report.
What verification actually buys
The interesting engineering problem, then, is not producing the answer. It is producing an answer somebody is willing to stand behind.
K-Link, a Thai company we backed, is a clean example. It runs an AI powered omnichannel contact centre, which it describes as one platform for every customer conversation. Its Kai agent answers from the company's own documentation rather than from general knowledge, and takes real actions through the company's own APIs. Every conversation is recorded, transcribed and searchable, and the platform reports first response time, resolution time, SLA breaches and sentiment. The company puts Kai's resolution rate at 80% of conversations.
Read as a feature list, those look ordinary. Read as strategy, they are the whole thing. Grounding an answer in the company's own material is what makes it defensible. Recording and measuring every conversation is what makes it checkable afterwards. Together they are what lets a support director allow software to speak in the company's name at all.
We would rather point at the mechanism than at the resolution rate. The mechanism is that the cost of checking fell far enough that the work could move.
From outsourced to in-house
The second defensible layer sits on top of the first, and it is the one that changes a customer's budget line rather than their afternoon.
Customer support is the clearest case in Southeast Asia. It has been outsourced by default for years, not because answering a customer is hard, but because staffing it around the clock, across languages and channels, meant building an operation most companies did not want to own. So it went to a third party, on a contract, priced per seat or per ticket.
When answers become grounded and every conversation becomes measurable, that calculation changes. The team that owns the customer can own the conversation, on its own timeline, without standing up an operation to do it. The work stops being a line in a vendor contract and becomes a workflow inside the company.
That is a much stickier position than a capability advantage. It is not a better version of what the customer was buying. It is a different shape of spending, and going back means rebuilding an outsourcing relationship the company has already exited.
What this means for a first check
We write first cheques of one hundred to two hundred and fifty thousand dollars into B2B and enterprise technology companies across Southeast Asia, usually before the metrics can tell us much. So what we are underwriting is mostly judgement about where a company chose to spend its hardest engineering.
A few things we look for, in the order we look for them.
Did the team go at the trust problem or the capability problem? Capability is where the demo lives. Trust is where the renewal lives.
Is the output checkable by the person who has to defend it? Not auditable in principle. Checkable in practice, in seconds, by the analyst whose name is on the work.
Does adoption move a task from outside the company to inside it? If it only makes an existing internal task faster, it is a feature. If it changes who does the work, it is a business.
Is the unglamorous part of the product the part they are proudest of? Founders who lead with verification, cleaning, validation and workflow tend to have spent real time with the buyer. Founders who lead with the model tend to have spent real time with other engineers.
None of this requires a large fund or a late stage to assess. It is the kind of judgement a first cheque is for, and it is one of the reasons we think seed in this region is underpriced rather than crowded. Early capital across Southeast Asia is scarce, and that scarcity is what lets a disciplined manager pay a rational entry price for exactly this sort of company.
The takeaway
When you look at an AI company, discount whatever it leads with. Ask instead what a customer would have to do to check its output, and what that customer stopped outsourcing once they trusted it.
Those two answers are usually the whole investment case, and they are almost never the thing on the front page.