When Forbes published its Asia 100 to Watch 2026 list this August, the cohort skewed heavily toward deep tech: GaN chip fabs out of Taiwan, humanoid-hand robotics from Korea, brain-computer interface startups from China [Forbes]. Sitting alongside them, less flashy but arguably more immediately investable, were three names from Insignia Ventures Partners’ portfolio: Cekat.AI, Eezee, and Pollen Tech, none of them building chips, none of them building robots. All three are Southeast Asian, all three are AI-agent companies, and all three are solving unglamorous, offline-heavy commerce problems: talking to customers, buying industrial supplies, and recovering value from stock that’s about to go to waste.

At first glance, they look like three unrelated bets scattered across a long list. Look closer and they form something more interesting: a functioning AI factory for commerce, not the data-center kind Nvidia’s Jensen Huang has spent the last several years evangelizing, but a regional, SME-native version of the same underlying idea.

The distinction matters and deserves precision before drawing the parallel too loosely.

First, What “AI Factory” Actually Means

Huang’s framing, repeated at every Nvidia GTC keynote since 2023, is deceptively simple: a data center should be understood not as a cost center that stores information, but as a factory that manufactures intelligence [Nvidia GTC]. Raw inputs (data and electricity) go in one end; a measurable, sellable output (tokens) comes out the other. The KPI that matters in this model isn’t uptime or utilization in the traditional IT sense. It’s throughput: tokens produced per dollar, per watt, per hour, continuously reinvested into more inference that in turn generates more revenue.

It’s a reframe, and reframes coming out of a chipmaker’s marketing department deserve some skepticism. But the underlying logic holds up once you strip away the GPU literalism: an AI factory, at its core, is any system where a repeatable input gets run through an intelligent process and comes out the other side as a measurable, monetizable unit of output, continuously, at a KPI you can actually track and optimize against.

That logic doesn’t require a data center. It requires a workflow: a repeatable, high-frequency process with a clear before-state and after-state. And Southeast Asia, it turns out, has been quietly industrializing exactly this kind of workflow, one commerce function at a time. Cekat.AI, Eezee, and Pollen Tech each own one link of what amounts to the full operational chain that commerce runs on: acquiring a customer, transacting with a supplier, and recovering value from what didn’t sell the first time around.

The Three Stations on the Line

Station One: Acquiring the Customer (Cekat.AI)

Founded in Tangerang in 2023, Cekat.AI runs AI-powered sales and customer-service agents across WhatsApp and Instagram, the channels where the overwhelming majority of Indonesian commerce actually happens, far more so than on a branded website or app [Cekat.AI]. This is not a peripheral detail; it’s the entire thesis. Indonesia’s retail and SME economy is conversational and mobile-first, and any AI layer that wants to matter at scale has to live inside the chat thread, not bolt onto a storefront nobody visits.

The problem Cekat is solving, in the company’s own framing, is a familiar one to anyone who has run a WhatsApp-based SME operation in Southeast Asia: manual replies costing roughly Rp150,000 per transaction and climbing an estimated 35% year over year, more than 30% of leads lost outright to slow response times, over 120 hours a month burned by staff answering the same repetitive questions, and closing rates languishing under 10% [Cekat.AI].

Recast in AI-factory terms, the raw input is an unstructured, often poorly punctuated inbound message: a DM, a WhatsApp text, a comment on an Instagram post. The “machine” is an AI agent that replies instantly, asks qualifying questions, and, in one documented case study, automatically recovers abandoned carts, lifting conversion by 35% [Cekat.AI case study]. Crucially, the output isn’t “a chatbot response.” It’s a structured, sales-ready lead logged into a CRM with intent, product interest, and contact details already captured, at a throughput no human support team could match during a midnight flash sale or a viral product moment.

Cekat now serves more than 3,000 businesses across 15 industries, with enterprise clients including Siloam Hospitals and Indomaret [Cekat.AI]. That range (a private hospital network on one end, a nationwide convenience-store chain on the other) is itself evidence for the factory thesis: the underlying input-output shape (message in, qualified customer out) is identical regardless of what’s actually being sold, which is exactly the kind of horizontal scalability a genuine production system needs to be worth building.

Station Two: Buying the Inputs (Eezee)

If Cekat is the factory’s front door, Eezee is the loading dock. Founded in Singapore in 2016, Eezee runs a B2B procurement platform built specifically around what the company calls “tail-end spend”: the low-value, high-frequency industrial purchases (gaskets, PPE, spare parts, consumables) that are individually too small to justify a formal sourcing process, but collectively numerous enough to quietly consume enormous amounts of procurement-team time [Eezee].

Eezee’s flagship AI tool, RFQBot, automates the request-for-quotation cycle that has traditionally required a procurement officer to manually email multiple suppliers, wait for responses, and reconcile quotes in a spreadsheet, a process that can stretch on for days even for a purchase order worth a few hundred dollars. RFQBot compresses that cycle dramatically by parsing the request, matching it against qualified suppliers, and returning comparable quotes automatically.

The company closed a $5 million pre-Series B round in February 2026, led by Korea Investment Partners, with participation from Wavemaker Ventures and continued backing from existing investor Kickstart Ventures [Eezee funding announcement]. The capital is earmarked explicitly for scaling RFQBot and a companion tool, ProcureFlow, across an expanded regional footprint spanning Thailand, the Philippines, Malaysia, Indonesia, and Singapore. Eezee’s own claim about the impact is direct and unhedged: procurement cycles cut from days to minutes, with total procurement costs down by at least 20% [Eezee]. The company’s enterprise customer roster (Chevron, ExxonMobil, Shell, and Zuellig Pharma among them [Eezee]) is telling in its own right: these are organizations with the resources to run sophisticated in-house procurement functions, and they are choosing instead to hand the long tail of purchasing to an agent.

In factory terms: the raw input is a fragmented, manual RFQ process scattered across email threads and phone calls; the machine is an agent that parses the request, matches suppliers, and returns a structured, comparable quote; the output is an awarded purchase order, produced at a cycle time collapsed by one or two orders of magnitude relative to the manual baseline.

Station Three: Recovering the Value (Pollen Tech)

The most counterintuitive station on this production line (and arguably the most data-rich) belongs to Pollen Tech. Pollen didn’t start as an AI company. It started in 2018 running manual excess-inventory recovery operations for FMCG brands across Southeast Asia: physically collecting aging stock from warehouses and distributors, finding buyers willing to take it off a brand’s hands, negotiating price, and arranging logistics [Pollen Tech]. It was, for years, a services business built on relationships and spreadsheets.

But eight years and more than 100 brand partnerships later, that operational grind had quietly produced something far more valuable than the recovery revenue itself: a first-party dataset covering 1,000-plus buyers and over 1 billion recovery signals (granular information on what excess stock actually sells for, who reliably buys it, at what discount, and how quickly its value decays once it’s flagged as surplus) [Pollen Tech].

That proprietary dataset became the training ground for Lily AI, Pollen’s “controlled recovery” agent, working alongside a demand-forecasting layer called Dahlia that scores which SKUs are at risk of becoming excess inventory and routes them for recovery before their value collapses entirely [Pollen Tech].

The scale of the underlying problem is stark: Pollen’s own figures show that inventory can lose up to 65% of its value by day 90 if left unmanaged, and when brands are forced into uncontrolled liquidation, stock frequently leaks into the grey market at discounts of 30-60% below retail, triggering distributor conflict and eroding brand pricing power across every other channel the brand sells through [Pollen Tech].

Lily’s job is to price every affected SKU dynamically, route it exclusively to pre-approved buyers and channels, and log the entire decision trail for audit and compliance purposes, work that previously took weeks of manual coordination and now happens in minutes. In factory terms: the raw input is inventory that is losing value with every day it sits unsold; the machine is an AI agent making real-time pricing and channel-routing decisions; the output is cash recovered on the brand’s own terms: margin protected, retail pricing intact, and none of it leaking into a grey market that would otherwise erode the brand permanently.

Put Together, It’s a Full Production Line

Line the three companies up end to end and a different shape emerges: not three disconnected startups, but three stations on a single commerce production line.

This is the same production-line logic Nvidia applies to a data center (input, machine, measured output, reinvested into more revenue) applied instead to the actual lifecycle of a commercial transaction rather than to a GPU cluster. Independent market data on vertical AI agents supports the pattern: research from Bessemer Venture Partners and Gartner indicates the category is already capturing roughly 80% of the contract value that legacy SaaS used to command, while growing more than 400% year over year [Bessemer/Gartner]. The market is already re-pricing these tools as production systems that generate measurable throughput, not as software subscriptions sold per seat.

That distinction (throughput-priced versus seat-priced) is precisely the shift Nvidia has been trying to force on how the industry values data-center infrastructure. Cekat, Eezee, and Pollen are executing the identical reclassification one level down the stack, at the SME and mid-market layer of Southeast Asian commerce, largely without anyone framing it in those terms.

Why This Is a Southeast Asia Story, Not Just an Insignia Story

For readers scanning the broader Forbes list for regional signal, the contrast is stark: almost every other company on it is capital-intensive by design (chip fabs, humanoid robotics platforms, biotech ventures requiring years of clinical validation), the kind of “AI factory” that genuinely requires billions of dollars in data-center or fabrication capex just to attempt. Southeast Asia’s three entries needed none of that.

Cekat plugs into WhatsApp, a channel that already exists and is already the dominant commerce interface across Indonesia. Eezee plugs into email-based RFQ processes that already exist inside every procurement department it serves. Pollen plugs into FMCG distributor networks that already exist, refined over years of manual recovery operations before any AI model was involved. None of the three asked the region to build new infrastructure before they could operate. Each was wired into infrastructure that already existed, doing the work more intelligently and at a fraction of the previous cycle time.

This is the more transferable insight for the region’s next generation of founders and the investors backing them: you don’t need to build the factory from scratch: you need to find the workflow that’s still manual, wire an intelligent agent into its existing pipes, and start measuring throughput as if it were a production line. Southeast Asia’s AI advantage isn’t going to look like Nvidia’s chips or Korea’s humanoid robotics. It’s going to look like three founders who each picked one unglamorous, offline-heavy corner of commerce (sales conversations, tail-end procurement, and decaying inventory) and quietly industrialized it, using rails that already existed and capital far smaller than what deep tech requires.

This also explains why Southeast Asia’s presence on lists like Forbes’ 100 to Watch tends to be systematically undercounted relative to the region’s actual commercial impact. Deep tech photographs well; a WhatsApp-native sales agent recovering a 35% conversion lift for an SME distributor in Surabaya does not. But the economics of the latter are, in aggregate, arguably more consequential for a region where SMEs represent the overwhelming majority of GDP and employment, and where the “last mile” of digitization has always been the binding constraint rather than access to compute.

The Investor Implication

If vertical AI agents really are capturing SaaS-level contract value while growing at roughly four times the historical rate of legacy software categories [Bessemer/Gartner], the moat question for investors evaluating the next Cekat, Eezee, or Pollen shouldn’t be “how good is the underlying model.” Every serious agent in this category is drawing on broadly comparable foundation models, fine-tuned and orchestrated differently but not fundamentally differentiated at the model layer. The moat, instead, is who owns the workflow at the exact point where a token gets converted into a completed transaction: the RFQ that becomes a purchase order, the WhatsApp thread that becomes a sale, the pallet of aging stock that becomes recovered cash.

Own that workflow long enough, and the data flywheel compounds in a way a challenger cannot easily replicate. Pollen spent eight years running manual recovery operations before Lily AI ever existed, accumulating the 1,000-plus buyer relationships and billion-plus recovery signals that now make Lily’s pricing and routing decisions meaningfully better than a generic model could produce cold. Eezee’s years of enterprise procurement relationships with the likes of Chevron and Shell generate a supplier-matching dataset that a new entrant would need years to reconstruct. Cekat’s 3,000-plus customers across 15 industries generate a conversational dataset spanning verticals that a narrower competitor simply won’t have access to.

That is the reproducible pattern worth watching for across the rest of Insignia’s portfolio and the broader Southeast Asian venture landscape: proprietary, transaction-level signal that compounds with every cycle through the machine, generated not by scraping the open web but by sitting directly inside a workflow that used to be entirely manual.

Conclusion: A Factory Built on WhatsApp, Not H100s

The headline buried under three names in a hundred-company list is this: Southeast Asia isn’t building Nvidia’s version of the AI factory. It’s building its own: cheaper to construct, less photogenic in a press release, and running on WhatsApp threads and RFQ inboxes rather than racks of H100s.

For the region’s investors and founders, the implication is a useful corrective to a global AI narrative still dominated by compute scarcity and chip nationalism. The next wave of durable Southeast Asian AI companies is unlikely to be defined by whose model is marginally better. It will be defined by who has embedded an agent so deeply into an existing, high-frequency workflow (commerce’s messy, offline, SME-heavy middle) that ripping it out becomes more costly than the software itself ever was. Cekat.AI, Eezee, and Pollen Tech got there first, at three different stations of the same production line. The rest of the region’s AI-agent economy is likely to be built by founders finding the next station nobody has automated yet.

References

1. Wehbe, R. “Forbes Asia 100 To Watch 2026.” Forbes, August 24, 2026.
2. Nvidia GTC keynotes (2023-2025): Jensen Huang’s “AI factory” framing of data centers as production systems for tokens.
3. Cekat.AI corporate website and product materials: WhatsApp/Instagram-native AI sales and customer experience agent positioning.
4. Cekat.AI website: stated efficiency figures (cost per manual reply, lead-loss rate, hours spent on repeat queries, baseline closing rates).
5. Cekat.AI case study materials: documented 35% conversion lift via automated lead-qualification and cart-recovery workflow.
6. Cekat.AI website: customer base figures (3,000+ businesses across 15 industries) and named enterprise clients, including Siloam Hospitals and Indomaret.
7. Eezee corporate website: B2B procurement platform positioning and definition of “tail-end spend” as a product category.
8. Funding announcement: Eezee $5 million pre-Series B, February 2026, led by Korea Investment Partners with participation from Wavemaker Ventures and Kickstart Ventures; stated geographic expansion across Thailand, the Philippines, Malaysia, Indonesia, and Singapore.
9. Eezee website and press materials: RFQBot performance claims (RFQ cycle time reduced from days to minutes; procurement cost reductions of approximately 20%).
10. Eezee website: enterprise customer roster, including Chevron, ExxonMobil, Shell, and Zuellig Pharma.
11. Pollen Tech corporate website and founder materials: company origin as a manual FMCG excess-inventory recovery operation, founded 2018.
12. Pollen Tech website: proprietary dataset figures (1,000+ buyers, 1 billion+ recovery signals) drawn from eight years of recovery operations across 100+ brands.
13. Pollen Tech product materials: description of Lily AI (“controlled recovery” agent) and Dahlia (demand-forecasting and at-risk SKU scoring layer).
14. Pollen Tech website: stated inventory value-decay statistics (up to 65% value loss by day 90) and grey-market discounting figures (30-60% below retail).
15. Bessemer Venture Partners and Gartner research on vertical AI agents: contract-value capture relative to legacy SaaS (approximately 80%) and year-on-year category growth (400%+).
16. Insignia Ventures Partners portfolio materials and public statements on Cekat.AI, Eezee, and Pollen Tech.