Vertical AI in Food Distribution: How a 5-Person Startup Hit $6M ARR — And Why This Is AI's Most Underrated Playbook

Vertical AI food distribution is quietly outperforming most consumer-facing AI products on the metric that actually counts: revenue per employee. In March 2026, New York-based Anchr closed a $5.8M seed round led by a16z Speedrun, with OpenAI leadership participating personally. Berlin's Choco, founded in 2018, now processes 8.8 million orders annually across 21,000 distributors and 100,000 buyers, with backing from Bessemer, Coatue, and Insight Partners at a $1.2 billion valuation. And in Singapore, a five-person startup called Farmio booked $6M in revenue over the past 12 months — automating 55,000+ orders and serving 1,600+ restaurant clients without a single dispatcher, route planner, or warehouse manager.

I've been advising founders on AI agent deployments for the better part of two years. The pattern keeps repeating. The most ambitious teams want to build the next ChatGPT wrapper, the next AI copilot, the next AI-powered SaaS for marketers. The teams quietly printing money are doing something else entirely. They're picking boring legacy industries like egg distribution, seafood wholesale, and restaurant supply, and using AI agents to compress 30-person operations down to 5. This piece breaks down where the money actually comes from, what AI is replacing, and what it costs to build one of these businesses.

The Economics of Food Distribution

US foodservice distribution is a $400 billion industry, according to the International Foodservice Distributors Association. The global foodservice market that Choco operates in is roughly $6 trillion. And the punchline is that most of these transactions still happen through pen, paper, fax, and manually keyed spreadsheets. IFDA's 2025 Technology Benchmarking Report, surveying 32 US distributors, found that integration compatibility with existing systems and budget constraints are still the top blockers for technology adoption. Translation: the industry hasn't been digitized yet, and AI agents are walking into a market that's structurally unprepared to compete.

Three properties make this space unusually well-suited for vertical AI startups.

First, demand is recurring and rate-limited by physics, not marketing budget. A restaurant uses eggs every day. A seafood buyer orders fish at least weekly. There's no "free tier to paid conversion" funnel; once a restaurant becomes a customer, repeat purchases happen by default. Compare that to consumer AI tools where 95% of signups never come back.

Second, every order has gross margin. Distributors don't operate on negative-margin growth loops. The price the restaurant pays minus the price the distributor paid is real money on every line item.

Third, the market is hyperlocal and structurally fragmented. A meat wholesaler in Manhattan doesn't compete with one in Los Angeles. Chef's Warehouse, a $5 billion publicly traded distributor, hasn't consolidated the market. There are still tens of thousands of regional distributors in the US alone. That means winner-take-all dynamics don't apply, and there's room for new entrants in every city and product category.

Choco's existence proves the market ceiling is high enough to matter. The company earned a $1.2B valuation in 3.5 years. Tenten's earlier breakdown in 2025 AI Agent商業革命: the small-team-big-company era made the same argument from a different angle: AI agents create asymmetric operating advantage in industries where labor cost dominates operating expense, and food distribution is exactly such an industry.

What AI Actually Replaces

A traditional mid-size food distributor needs roughly 25-30 people. Procurement negotiates with suppliers. The warehouse team manages cold chain and inventory. A routing team schedules delivery vehicles. Account managers handle customer relationships and payment terms. Finance and admin round out the headcount. AI doesn't shave 10% off the org chart. It eats four of the five core functions.

Order Intake

This is the biggest savings center. Traditional order intake comes through phone calls, faxes, WhatsApp, SMS, voicemails, and handwritten notes, all manually typed into ERP systems by an order desk team. Choco's OrderAgent, built on OpenAI APIs, ingests all of these formats simultaneously and converts them into structured ERP orders. The system handles colloquial phrasing like "more steak" or "same as last time, five units" by matching against each customer's purchase history.

Real-world performance numbers: up to 97% accuracy on order extraction, with onboarding completing in 2-4 weeks. Order processing time drops from roughly 8 minutes to under 30 seconds per order. Lynas Foodservice, a UK distributor, reported that a team of four people handling 240 customers per week scaled to over 1,000 customers per week with just two people on the order desk. The other two redeployed to telesales. The OrderAgent's sibling, VoiceAgent, uses OpenAI's Realtime API to handle live phone orders with sub-second latency.

Procurement Decisions

Which supplier to buy from, in what quantity, on which day. This is a small optimization problem that responds well to LLM-augmented decision systems pulling from inventory levels, historical consumption patterns, and price feeds. Anchr's early customer data is concrete: one distributor reduced aged inventory write-offs by $30,000 in a single month by using AI-driven demand signals to inform purchasing decisions.

Route Planning

Vehicle routing problems (VRP) have been computationally solvable for over a decade. Google OR-Tools and similar open-source libraries handle most mid-scale routing scenarios. The novelty isn't the algorithm; it's that AI-native distributors are actually deploying this in production from day one instead of treating dispatch as a manual function.

Reconciliation and Inventory Forecasting

Automated invoice generation and demand forecasting based on historical patterns. For thin-margin food distributors, this is less about labor savings and more about cash flow visibility. Accurate inventory forecasting reduces capital tied up in perishable stock.

Add those four together and you've replaced 60-70% of headcount at a traditional distributor. That's the math behind Farmio's five people serving 1,600 restaurants, and behind Choco platform users like Lynas going from 4 to 2 people on the order desk.

The Two Playbooks: Platform vs. Player

There are two distinct routes into this market, and choosing wrong will cost you somewhere between 18 months and $2M.

The Platform Play (Choco)

Sell SaaS to existing distributors. Don't touch inventory, don't deal with cold chain, don't get import licenses. Choco was founded in 2018 by Daniel Khachab and Julian Hammer (with co-founder Rogerio Da Silva Yokomizo). The company has raised $301M to date at a $1.2B valuation set in April 2022.

The wedge: enter through the supplier side, not the restaurant side. Convince one large supplier to onboard, and their hundreds of restaurant customers come along automatically. Restaurants pay nothing; suppliers get core functionality free and pay for the AI modules. Choco now serves over 21,000 distributors and 100,000 buyers across the US, UK, Europe, and the GCC.

Platform mode has a clean cold-start story. It solves a real pain: distributors want to digitize but don't know where to begin. The trade-off: margins are SaaS-thin, and the technical moat compresses as foundation LLMs improve. OrderAgent's core capabilities were nearly proprietary infrastructure when Choco started in 2018; today, similar systems can be assembled in weeks using OpenAI or Anthropic APIs.

The Player Play (Farmio)

Become the distributor. Buy product, sell product, deliver product. AI runs the back office. Farmio was founded in 2023 by Paco Chan and Andrian Kanta in Singapore, with $125K in seed funding from Antler and GreenBridge Venture, plus incubation support from Hong Kong's Cyberport.

The trailing-12-months numbers are striking for a five-person team: $6M in revenue, 55,000+ orders automated, 1,600+ B2B customers. That works out to roughly $109 per order and about 150 automated orders per day. The order density is what makes the unit economics work.

Player mode is harder to start. Food supply runs on trust: restaurant owners have years-long relationships with existing suppliers, credit terms, payment habits, and reliability data. Switching suppliers means rebuilding all of that, and a single missed delivery means the restaurant doesn't open the next day. Nobody changes suppliers to save a few cents per case.

But once customers do switch, player-mode unit economics are dramatically better than platform-mode. The revenue isn't SaaS subscriptions; it's product gross margin plus the labor savings from running on five people instead of thirty. And the long-term moat is deeper, because supply chain relationships, import licenses, and cold-chain partnerships aren't replicable by a well-funded competitor in 90 days.

Side by Side

Dimension	Platform (Choco)	Player (Farmio)
Role	SaaS vendor	Actual distributor
Cold start difficulty	Medium	High
Unit economics	SaaS subscription, thin	Product margin + labor savings, thick
Customer acquisition wedge	Supplier-led network effect	Backup supplier → primary supplier
Moat	Product depth, integrations	Licenses, supply chain, cold chain
Geographic ceiling	Global	Single city or metro
Capital required to launch	$3M–$10M	$300K–$500K
Typical customer	Mid-to-large distributors	Restaurants, retail, F&B

Which one fits depends on what you actually have. Technical founders without industry relationships should go platform. Operators with a decade in food supply chains have a structural advantage on the player side that no amount of venture funding can replicate.

The Concrete Path In

Pick the Market and Category

Small and dense. Singapore is 280 square miles; in 2021, the average person there consumed about 390 eggs per year, and delivery radius is short enough that one truck handles a full day's customer list. New York City, the SF Bay Area, and a single dense suburb of a major metro all qualify. The wrong choice is "the Northeast" or "Southern California." Too sparse, too much driving, unit economics break.

Pick a category that's high-frequency, standardized, and used by every restaurant. Farmio picked eggs because purchase frequency is high, individual order sizes are small, and repeat behavior is near-automatic. Once the customer relationship is established, expanding into cooking oil, rice, condiments, and other staples has near-zero marginal acquisition cost. Farmio is already doing this.

Solve Supply Chain and Licensing

This is the real moat. For Singapore egg imports, you need an SFA import license, SFA-inspected cold storage, and supplier relationships with farms in one of the 42 countries/regions accredited by SFA to export poultry, meat, or eggs to Singapore. In the US, the equivalent is FDA-registered facilities, state-level licensing for cold-chain operations, and direct relationships with USDA-inspected suppliers.

These are both your moat and your barrier to entry. Farmio's approach: don't own trucks, don't own warehouses, partner with existing cold-chain infrastructure. Asset-light is the only sane way to start.

Build the AI Stack

You don't need to build from scratch. WhatsApp Business API plus a foundation LLM handles order intake. Google OR-Tools handles routing. Inventory forecasting can start in Excel with a few Python scripts. Choco's OrderAgent ships with 2-4 week onboarding for customers, which means the systems integration work is genuinely tractable.

The hard part isn't picking models. It's the integration glue. Getting a WhatsApp message that says "same as last time, 30 cases" to correctly trigger an order against the right customer's specific SKU, then auto-route a delivery and flag exceptions for human review. That's months of iteration, not weeks. Choco hit 97% accuracy after years of building ground-truth datasets, evaluation infrastructure, and A/B testing harnesses.

On model choice: OpenAI's GPT family currently leads on multimodal input (voice, images, PDFs), which is why Choco and Anchr both selected OpenAI as their primary provider. Anthropic's Claude has comparative advantages on long-context reasoning and tool calling, making it well-suited for the procurement-decision and forecasting agents that need to synthesize across multiple data sources. The right call is usually a multi-model stack, not a single-vendor commitment.

Customer Acquisition: The Hardest Part

Farmio has never publicly disclosed how they landed their first 100 customers. Paco Chan's background (distribution sales at P&G, last-mile logistics at Pickupp) suggests warm intros from industry contacts likely drove early traction. The probable playbook from there is a classic foot-in-the-door: become the backup supplier, prove reliability over 60-90 days, then earn primary status.

New restaurants opening in any given metro are another low-friction wedge. They have no incumbent supplier relationships and are explicitly shopping for distributors. Farmio is signing 6-7 new customers per business day on average across 1,600+ accounts, which suggests they've found a repeatable acquisition channel, though the specifics aren't public.

Plan on 12 months minimum to build customer count and trust if you're going the player route. The quality of your first 100 customers determines whether the business lives or dies. Their repeat purchases are your cash flow base, and their word-of-mouth is your acquisition flywheel. This is the same lesson from YC's startup ideation guide that Tenten has analyzed: solve one person's problem completely before scaling.

Why Now

Timing matters more than business model here. Three factors compound.

Foundation Models Hit the Required Capability Tier

When Choco started in 2018, GPT-3 didn't exist. The OrderAgent stack was largely proprietary, requiring years of in-context learning infrastructure, customer SKU mapping tables, unit preferences, delivery patterns, all encoded from order desk knowledge that lived in human heads. Today, the same capability is accessible through OpenAI or Anthropic APIs in weeks. The technical entry barrier has collapsed.

That means food distributors don't need to become AI companies to deploy AI. They just need to integrate it. This is precisely the wedge Choco is selling, and it's why Anchr booked seven-figure revenue in the 12 weeks of its a16z Speedrun residency.

Capital Has Confirmed the Direction

The Anchr seed round on March 10, 2026 is a meaningful signal. The $5.8M check itself isn't large by venture standards, and a16z Speedrun is an early-stage accelerator rather than a16z's main fund, so I'd be cautious about over-reading the institutional commitment. But Anchr's seed-stage customer mix already includes a $5B publicly traded distributor, reported to be Chef's Warehouse. OpenAI leadership investing personally is a credible signal. The people closest to the model capabilities believe this segment is real.

Incumbents Will Not Self-Transform

A distributor founder who's been running the business for 20 years isn't going to wake up tomorrow and rewrite their stack. This isn't disrespect; it's organizational physics. Their reaction speed defines the window for new entrants. And because food distribution is hyperlocal and fragmented, no single competitor will lock the market. Choco operated for years without a Singapore competitor; Farmio walked in and reached top-10 status in eggs within 24 months.

There are equivalent openings in Southeast Asia, the Middle East, Africa, Latin America, and dozens of US metro areas. Anywhere that still runs on phone orders and fax sheets has a vertical AI opportunity attached.

The Real Cost of Starting

The "five people, $125K seed, $6M revenue" framing makes this sound easier than it is. Three things get glossed over.

Startup capital. Farmio raised $125K from Antler at seed. The realistic capital requirement to get from launch to stable operations is closer to $300K–$500K. That accounts for working capital, cold-chain partnerships, system development, and customer acquisition subsidies. The platform path requires more. Choco raised over $300M before reaching its current 8.8M-order-per-year scale.

Team composition. Five people is real, but you need at least one founder who understands B2B supply chain sales and one who can build the AI stack. CEOs in player-mode startups spend most of their early months on licensing applications, supplier negotiations, and customer relationships. None of these can be done from a laptop.

Timeline. Farmio was founded in 2023 and reached its $6M revenue run rate over roughly 24 months. Choco took close to seven years to reach 8.8 million orders annually. Supply chain trust isn't an asset you can buy faster by deploying more capital.

Author Insight

When I sit down with founders to talk about AI agent deployment, the most common misconception is that the green field is in new technology rather than old industries. The Farmio and Choco stories are clean rebuttals to that assumption.

Over the past two years, my team at Tenten has worked with manufacturing, finance, and distribution clients on AI agent pilots. The pattern that consistently produces order-of-magnitude operational gains isn't AI for marketing copy or AI for customer service chatbots, since those use cases are saturated, and the unit economics rarely justify the integration cost. What works is plugging AI agents into the workflows that have been manual for decades: procurement, dispatch, reconciliation, inventory replenishment. The compression from 30 people to 5 isn't a marketing pitch; it's the actual delta when AI eats a function that was previously labor-bound.

What I tell founders considering this space: if you're not from the industry, start with the platform model. The supply chain trust, licensing, and physical infrastructure required for the player model take years to build and aren't compressible. But if you've spent a decade inside food distribution, manufacturing logistics, or commercial cleaning supply, the player-mode unit economics will eventually beat any pure-SaaS play in the same vertical.

For mid-sized distributors and B2B operators reading this in the US, the question isn't whether to adopt AI. It's whether you'll be the incumbent that AI-native operators disrupt, or the operator deploying AI first. Tenten has been running AI agent integration pilots with distribution and operations clients in the past several quarters, typically reducing back-office headcount requirements by 40-60% across order intake, dispatch, and reconciliation workflows. If you want to talk through how AI agents map to your existing ERP and operations stack, book a session with the Tenten team.

FAQ

How does Farmio operate with only 5 employees serving 1,600 restaurants?

The team replaces four labor-intensive functions with AI agents. Order intake uses LLM parsing of WhatsApp, SMS, and email. Procurement decisions are data-driven. Route optimization uses Google OR-Tools or similar. Reconciliation and forecasting are automated. Farmio reports automating 55,000+ orders over the past 12 months, averaging around 150 orders per business day.

What's the difference between Choco and Farmio's business models?

Choco is a SaaS platform that sells AI-powered order processing tools to existing distributors. The restaurant side is free; suppliers pay for premium modules like OrderAgent. Farmio is itself a distributor; it buys, sells, and delivers eggs and other staples, using AI as the back-office system. Platform mode is easier to cold-start but has thinner margins. Player mode is harder to launch but has structurally better unit economics and deeper moats over time.

How much capital do you need to start a vertical AI food distribution business?

Player mode typically requires $300K to $500K in capital to cover working capital for inventory, cold-chain partnerships, AI system development, and early customer acquisition. Platform mode requires more upfront technical investment but avoids the inventory and licensing burden. Both paths need at least one founder with B2B supply chain experience and one with AI engineering capability.

How does Choco's OrderAgent actually work?

OrderAgent uses OpenAI APIs to ingest multimodal inputs (email, SMS, image attachments, PDFs, voicemail transcriptions) and converts them into structured orders ready for ERP ingestion. The core technical innovation is dynamic in-context learning: the system resolves ambiguous phrases like "same as last time" by referencing each customer's historical order history and SKU mappings. Accuracy reaches up to 97% on production traffic, with order processing times dropping from approximately 8 minutes to under 30 seconds.

Which markets and product categories are best for vertical AI food distribution?

Small, dense geographies with high restaurant density and high food-away-from-home rates. Manhattan, Brooklyn, the SF Bay Area, downtown Boston, Singapore, Tokyo, and Hong Kong all qualify. Categories should be high-frequency, standardized, and used by virtually every restaurant: eggs, cooking oil, rice, dry goods, specific protein cuts. The strategy is to establish the customer relationship with one high-frequency category, then expand horizontally at near-zero marginal acquisition cost.