Never marry your models: Kata.ai's decade of outcome pricing

The first .ai domain in Indonesia

Muzamil opens the episode by framing Irzan as someone who has spent a decade in conversational AI in a market where the frontier labs still barely operate. Irzan takes it further. “Kata.ai is the first AI company in Indonesia. We bought the first .ai domain, I think, for less than $10 back then.”

The origin story is not an AI company. It is YesBoss, a 2015 SMS concierge modelled on a US service called Magic, which had made headlines by letting customers order almost anything by text, including a tiger. Irzan’s read on Indonesia was cultural. It was a chatty country, a heavy BlackBerry Messenger user base, and a population that loved services.

But he knew from before the first line of code that the human concierge model would not scale. So they ran two tracks in parallel: a B2C concierge with almost 100 people across Jakarta and Manila, and a small AI lab buying GPUs, scraping Indonesian forums and social media, and working with linguistics students to build a corpus. At the time OpenAI was not there. Anthropic was not there. Google, Amazon and Microsoft had no Indonesian NLP model. Irzan calls the decision to build one themselves a product of “our youthful energy and, you know, foolishness.”

The pivot Unilever paid for

By 2016 the money was gone. Revenue was essentially zero. Irzan had to let go of 92 staff. The one asset that survived the cull was the NLP model, which happened to reach production at the moment Unilever came looking for someone in Indonesia who could build a female-assistant chatbot for their Gen Z and millennial customers.

The deal size, in Irzan’s phrase, was “a high figure of uh six digits in US dollar.” His investors were impressed. His logic was straightforward: if Unilever needs this, so does FMCG generally, and so do telco, retail, e-commerce and banking. In October 2016, YesBoss died and Kata.ai was born.

A decade later, Kata.ai works with 300+ enterprise clients across telecom, financial services, retail, healthcare and government, reaches 70 million end users, and facilitates $2.5 billion in transaction volume annually.

The 100x account that changed the pricing model

The most important story in the conversation is a repricing.

For years Kata.ai sold customer service automation on a cost-saving pitch. It worked for the largest customers, the ones with hundreds of agent seats, but Irzan is blunt about the ceiling: “when we talk about cost saving, there’s always a ceiling.” Human in the loop does not disappear. The savings math eventually stops selling itself.

The uptick came from flipping the value proposition. Same product, same technology, new frame. Irzan tells the story of Indonesia’s largest national pawn brokerage, a Kata.ai customer since 2018. For three to four years the relationship was a customer service chatbot handling 70-80% deflection. Then Kata.ai noticed the client’s app was not landing with its Gen X and baby boomer base.

They proposed turning the assistant into a transactional channel on WhatsApp: pay bills, make payments, buy gold. Year one facilitated half a million dollars in transaction volume. By last year it was close to $10 million. And the account itself, as Irzan puts it, “their ARR contribution to us from 2018 to last year, it grew by 100 times.”

The lesson he draws is one of the sharpest lines in the episode: “the future of AI business that will build most out there is not charging by tokens, not charging by usage, but if you can build it by outcome models.” Customers, he adds, do not care which model is running. They care about the outcome.

Never marry your models

Muzamil pushes on a natural question. Given Kata.ai originally built its own Indonesian NLP model, is it still training its own frontier?

Irzan’s answer is direct. “My mantra to my team is never marry your models. So, instead of we fight them, we join them.” Kata.ai uses everything: US frontier labs, Chinese open weights, whatever fits the use case. Around 70% of workload still runs on US frontier models, mostly because SaaS-style inference APIs are simpler for customers to consume.

But the open-weight pull is real and it is not primarily a cost story. Irzan estimates the driver split at “60 40” in favour of sovereignty over cost. Regulated banking, insurance and telecom clients want their data on their own cluster. They read the frontier lab terms of service, but hosting it themselves is, in his words, “something for their peace of mind.”

To operationalise this, Kata.ai partners with a local Indonesian data center running a token factory with multiple open-weight models. Data residency stays in Indonesia. Models can be swapped as the frontier moves.

Sovereign AI is existential, not nice-to-have

Muzamil brings the conversation to the wider sovereignty debate, the one recently pulled into the mainstream by Alex Karp of Palantir and Satya Nadella of Microsoft. Irzan gives the most declarative line of the episode.

“Sovereign AI is not a nice-to-have for Indonesia. It’s existential.” He lays out the argument: the fourth largest country by population, over 700 languages, a digital economy heading toward hundreds of billions of dollars. If the AI running Indonesian banks, government services and daily conversations is built on foreign models, foreign data and foreign infrastructure, then Indonesia is not participating in the AI economy. “We’re just renting it.”

He points to Indosat, a major Indonesian telco and Kata.ai customer, which has pivoted into building an AI data center with Nvidia. Kata.ai is their local token factory partner. Sovereignty, he clarifies, is not isolation. “It’s about making sure Indonesia captures the value, controls the data, and shapes the technology on our own terms.”

Corporate excitement, bureaucratic drag

Muzamil describes the strange geography of AI attitudes inside the US: San Francisco billboards screaming AI, corporate America two hours east barely mentioning it. He asks Irzan for the Southeast Asian read.

Consumer adoption among Indonesian white-collar workers is high. Irzan estimates 70% of them bring their own AI tools to work, unpaid by the employer, free or paid. Corporate excitement exists too, but the decision cycle does not match the model release cycle. He describes recent voice AI deals where clients want to launch next year with specifications locked to today’s technology, in a market where the frontier moves every month.

Muzamil pushes on the structural problem this creates. Any architecture built today provides clean value for about two weeks before something better arrives. Irzan’s answer runs on three layers. SaaS deployments update silently on Kata.ai’s cloud. Containerised Kubernetes deployments allow controlled updates on client clusters. And most enterprise clients do not actually need every generational jump. “We may still be able to ship what they need without kind of like updating or asking them to to switch the model.”

Latency, memory, and the moat nobody talks about

Muzamil raises a technical thread: chats gave way to agents, then loops, then graphs. Multi-agent ping-pong dramatically improves output quality. But Kata.ai runs 95% of its deployments on WhatsApp. Latency matters. You cannot run a twelve-agent swarm on a customer waiting for a balance query.

Irzan’s framing is customer-experience-first, not architecture-first. “If we work from technology, it’s just too wild.” Kata.ai runs a cocktail of models: smaller models for chat on the back end, reasoning models for agentic functions, deliberately not the longest reasoning chains, because latency compounds with every loop.

Then there is UX craft. Typing indicators. Cues like “please wait, I’m checking this and that.” The interface becomes part of the answer. And underneath it, memory, which Irzan flags as the real moat: persistent memory tied to CRM and CDP, plus lightweight session-level context. Portability of that memory, he notes almost in passing, is what will decide vendor lock.

Forward-deployed engineers, and 60 people doing more than 100

Muzamil brings up the current fashion for forward-deployed engineers. Irzan’s response is that Kata.ai has run this model for ten years, not as a trend but because the work demands it. Data cleansing, integration across enterprise systems, workflow design, prompt tuning. “It takes a village to make AI work for customers.”

Roughly 80% of Kata.ai’s deployments involve FDEs. The definition has widened beyond software engineers to include people who can speak the customer’s business language.

Internally the company runs on around 60 people, materially more productive than the 100 it had a decade ago. Irzan estimates 70% of work is now AI or agent assisted, and role boundaries are dissolving. “There’s no more product managers. There’s like builder. Like either you have, you’re an engineering with business understanding and product sense or you’re a product person that can code.”

On the wider jobs question he is measured. “AI exposes people, organization, country, the strong ones that can adapt and the weak ones that are not willing to adapt.” He expects role shifts more than mass unemployment over a three-to-five year window, while conceding the pace makes prediction unreliable. His individual-level framing is stark: “now the gap of, you know, you you’re having the ideas, you want to do something, and the real results are two things, ambitions and tokens. You lack two of those, it’s game over.”

Signal over noise

Muzamil closes with a question about the three biggest learnings from the last three years. Irzan collapses it to one discipline: remove the noise, focus on the signal. Twitter will always show a new framework, a new agent architecture, a new benchmark. The anchor has to sit somewhere else. For him it is customer success.

By the end of the conversation that discipline is what the whole Kata.ai story reduces to. A decade of not marrying models. Building for outcomes, in Indonesian, on infrastructure the country controls, priced against transaction volume rather than token count. Top line doubled in 2025 with bottom line tripling, and revenue growing three times year-on-year through the first half of 2026. Profitable for three years running.

In a market where most AI companies are still trying to work out what they actually sell, that clarity is the story.