They started selling online at 15. Their company, Siml, now runs over 1,000 stores.

Nurtilek Raimzhanov and Zuhayr Abdullazhanov met in high school and have been best friends ever since. They were fifteen when the pandemic hit, and they decided to sell something online. In their case, hand sanitizer.

It worked well enough to become an education. They wrote the listings, ran the ads, and answered the messages. Because Zuhayr got tired of opening the same twelve competitor pages every morning, he started writing scrapers to do it for him. They broke constantly, and that detail is the seed of everything they have built since.

They Got Into the Best Schools. Then They Left.

Both left Kyrgyzstan to study engineering at top universities in the United States, and kept building together across a two hour time difference. Their last project before Siml was Tail AI, a personal finance app for college students, which passed 10,000 users within a couple of weeks of launch. Alongside it they published research on data analysis and agent orchestration, the question of how multiple autonomous agents divide work, share state, and stay correct when they act at the same time. Nurtilek was separately building machine learning systems for freight and logistics companies, working on routing, delivery prediction, and the classification problems that sit between a shipment and a customs officer.

Then they both dropped out. Siml is now backed by some of the top investors in Silicon Valley.

Asked what has actually changed since they were teenagers selling online during the pandemic, Zuhayr’s answer is flat. “Nothing, really. We were figuring out listings and ads and customer messages then, and millions of people are still figuring out the same things now. The tools got prettier. The work did not get easier.”

The Platforms Solved Distribution. They Never Solved the Work.

Opening a store takes an afternoon. Running one takes your life. Worldwide retail e-commerce is projected to reach $6.9 trillion in 2026, and anyone with a laptop can be selling by the weekend. But the first order brings a stack of jobs with it: support, advertising, listings, creative, pricing, returns, compliance, analytics. An entire economy grew up around not doing them, and the most reliable way to make money in e-commerce is still to sell people the promise of making money in e-commerce.

Then the ground moved twice.

Tariffs Turned Margin Into a Daily Calculation

The United States suspended the de minimis exemption in 2025 and confirmed in February 2026 that it continues with no restoration date. Roughly four million packages a day now require formal customs treatment. Every product needs a classification code, country of origin, and material description, and landed cost has to be recalculated per shipment rather than assumed per season. Most sellers do not have someone who understands duties. They have a spreadsheet and a bad feeling.

Your Next Customer Is an AI Agent Shopping for a Human

The second shift is bigger, and it is the one Siml was built around. Buyers are increasingly not browsing stores. They ask an assistant, and the assistant browses for them.

Adobe Analytics found traffic from AI sources to United States retail sites grew 393% year over year in the first quarter of 2026, and the quality flipped even faster than the volume. In March 2025, AI referred visitors converted 38% worse than paid search and email. By March 2026, they converted 42% better, with revenue per visit 37% higher. Then the finding that should keep every store owner awake. Adobe scored retail product pages at just 66% machine readability, meaning a third of the content on the page where the purchase decision happens cannot be read by the models sending that traffic.

Two centuries of commerce went into making products findable by people. Almost none of it made them findable by agents. A merchant can be discoverable to every human shopper alive and invisible to the systems those shoppers now delegate to. McKinsey projects agentic commerce will drive $3 to $5 trillion globally by 2030.

What Siml Actually Does

Siml is an AI that runs your online store, and the distinction that matters is between assistance and custody. Most AI commerce tools assist. They draft a description, suggest a bid, answer one email. Siml takes custody. It treats store operations as a single autonomous control loop, which means the agents hold the state of the whole business and act on it continuously, without being asked.

Support runs itself across every inbox, with full context on the order, the product, and the customer’s history. Advertising becomes a decision rather than a second job. Creative is produced and tested continuously rather than commissioned. Pricing and inventory decisions arrive with the reasoning attached instead of as forty-two metrics on a dashboard.

The part with no precedent is the other side of the counter. Siml lets a seller sell to agents rather than to humans. Listings and pages are structured so an assistant shopping on a person’s behalf can find the product, read its real attributes, compare it honestly, and buy it, without a human ever scrolling the page. Every previous era of commerce built a place where demand could find supply: the market square, the catalog, the search result. None were built for a buyer that reads structured data and never scrolls. The merchants who are legible to agents first will hold the advantage early sellers held on every previous platform.

Sellers do not have to move to get it. They can migrate onto Siml or keep selling where they are and let Siml run the operation behind it. Siml reports over 1,000 stores, hundreds of thousands of agent actions executed on behalf of merchants, and measured time savings of more than 3x against the same operation run manually, with customers from New Zealand and Mexico to the United Kingdom and across the United States. What struck the team in early interviews was how identical the failure was everywhere. The inbox is full, the ads are not converting, the listings are stale, the margin math changed and nobody explained why.

Why This Was Hard to Build, and Who Built Which Half

Autonomous store operation had not shipped before, and not because language models were not good enough. Three problems underneath it are systems engineering, not text generation. A store’s truth is scattered across marketplaces, catalogs, invoices, and customs forms with no shared identifier. Commerce actions are irreversible, because an agent that issues a refund or moves a price has spent real money. And a customer writing “it never came, this is the third time” is describing a delivery exception, an unresolved prior ticket, and an escalating emotional state at once. Partially right is worse than silent. The two split those problems the way they always have.

Zuhayr owns the data core and the accuracy bar. He is the architect of the catalog graph, the entity resolution layer that ingests listings, variants, supplier records, and order history from every connected channel and resolves them into single canonical products, matching across inconsistent titles, missing identifiers, and near duplicate variants. Every agent action reads from that layer rather than a raw marketplace API, which is what lets one agent answer a customer, adjust a price, and update a listing while referring to the same object. He also built Siml’s evaluation harness, which scores agent outputs per action type against labeled outcomes and routes anything below threshold to a human instead of executing it.

“The harder question came later, which is proving a system is allowed to act. If you cannot measure the accuracy of an action, you should not automate it. Most of the industry automated first.”

Nurtilek owns execution and cost. He is the architect of the execution layer, the runtime that lets agents take irreversible actions safely. It wraps every consequential action in a merchant defined permission model, simulates the effect before committing, records an auditable trace of what was done and on what evidence, and pauses anything above a configured risk threshold for approval. It is what allows a merchant to sleep while agents issue refunds and change prices, and it is harder to build than the agents themselves, because the platforms it acts on offer inconsistent guarantees about what can be undone. When the de minimis exemption disappeared, he built Siml’s landed cost engine, which classifies products, computes duty exposure, and pushes the margin consequence directly into pricing, using the same classification work he had done for freight companies two years earlier.

“Logistics teaches you that the expensive mistakes are physical. Software people talk about hallucination as a text problem. In commerce a hallucination is a package going to the wrong country, or a price that loses money on every unit for three weeks before anyone notices.”

Not the Next Storefront Platform

Plenty of companies tried to build the anti-Amazon or the anti-Shopify and died on an axis where the incumbent was never beatable. Siml’s read is different. The opportunity is not another place for sellers to arrive, but to solve the problems they already have, on whatever platform they are already standing.

“We had offers to join both of the companies people assume we are competing with. We think they will keep operating and keep winning at what they do. But AI is changing how commerce works underneath all of them, and that change does not belong to any single platform.”

The Case Against

Merchants have been sold automation before, usually as rules engines that broke silently, so handing pricing authority to a model is a large ask. Siml’s answer is measurement and permission thresholds, which still has to survive a bad week. The founders do not dispute it. Their argument is that the direction is not in question, even if the timing is.

What Comes Next

A store used to be a place a person visits. Then a website. Now it is a structured set of claims about products, prices, and availability that agents query on behalf of people. Running a store is no longer something a person should do by hand. It is something a person should direct.

The seller still makes the product. She just stopped doing the other five jobs.

Get the TNW newsletter

Get the most important tech news in your inbox each week.

Contributed article. Not produced by the TNW newsroom and does not reflect the editorial stance of TNW.