Siml Runs 1,000 Online Stores With AI Agents as E-Commerce Automation Shifts From Assisting to Taking Custody
Siml hands store operations to AI agents across 1,000+ merchants. How custody-style agents differ from assistive tools, what Adobe's AI traffic data shows, and where the risks sit.
Key Takeaways
- Siml, built by two 21-year-old founders, runs a set of AI agents covering support, advertising, inventory and returns across more than 1,000 stores
- Unlike assistive tools that stop at suggestions, Siml decides whether an action may execute using accuracy scoring and merchant-defined permissions
- Traffic from buy-side agents keeps growing, yet Gartner expects more than 40% of agentic projects to be scrapped, so how you delegate matters more than whether you delegate
From selling hand sanitizer to running 1,000 stores
Siml takes custody of store operations, not just tasks. Two founders who started selling hand sanitizer at 15 now run AI agents across 1,000+ stores, as Adobe data shows AI-referred shoppers convert 42% better than paid search.
thenextweb.comDuring the pandemic, two 15-year-olds from Kyrgyzstan who had met in high school started selling hand sanitizer online. Orders were fulfilled by hand, and repricing meant opening competitor pages one by one. One of them eventually wrote a scraper to crawl competitor prices, and it broke every time a rival changed its page structure. That experience of automation that keeps breaking became the seed of the product they build today.
Nurtilek Raimzhanov and Zuhayr Abdullazhanov both studied engineering at top U.S. universities before dropping out. Their previous product was Tail AI, a personal finance app for college students that passed 10,000 users within weeks of launch. The two have also published research on data analysis and agent orchestration, meaning how multiple autonomous agents divide work, share state, and stay correct under concurrent execution.
Their current company is Siml. According to The Next Web, agents are already running across more than 1,000 stores, and the number of actions executed on behalf of merchants reaches into the hundreds of thousands. The company's own homepage displays a counter of over 365,875 completed store actions. Customers are spread across New Zealand, Mexico, the United Kingdom and the United States. Siml is described as backed by top Silicon Valley investors, but the amount raised and the names of the investors are undisclosed.
The shape of the product is visible in the company's FAQ. It connects to three channels today, Shopify, Amazon and TikTok Shop, and seven named agents divide the functions between them. Maya handles customer support, Riley writes and fixes listings, Sasha runs marketing, Jordan watches inventory, Quinn tracks finance, Theo manages returns, and Avery works reviews. Pricing is credit based, with tiers charged against the volume of work the agents actually perform, though the unit rates and plan prices are undisclosed. Siml says users report more than 3x time savings compared with manual operation.
What separates assisting from taking custody is the execution design
This is the heart of the story. There is no shortage of AI tools for e-commerce, but most of them stop at a suggestion. They draft ad copy, flag that stock is running low, prepare a reply. A human presses the final button, and if nobody presses it, nothing happens. What Siml claims to do differently is take on that last step. The original wording is that it takes custody of store operations.
Anyone can say that. What deserves scrutiny is what was actually built to make execution safe to hand over.
| Dimension | Assistive AI tool | Custody-style agent |
|---|---|---|
| What triggers it | A person opens the screen and asks | State changes in stock or price trigger it automatically |
| What it produces | Suggestions, drafts, reports | Actual changes to price, stock, ads and replies |
| How correctness is assured | A human reviews and decides | Accuracy scoring and confidence thresholds, with low confidence routed to a human |
| How failures surface | They sit unnoticed until someone looks | Execution traces and audit logs make them traceable |
| Permission design | Access rights to the tool only | Per-action permission scopes and approval checkpoints |
| Irreversible actions | A human judges and executes | Simulated before commit, high risk paused for approval |
| The merchant's job | Doing the work itself | Designing scopes, thresholds and measurement |
Zuhayr's domain was data accuracy. He designed what the company calls the catalog graph, an entity resolution layer that reconciles inconsistent titles, missing identifiers and marketplace variants into a single canonical product reference. It sounds unglamorous, but without it several agents will reach contradictory conclusions about the same item. If the product the inventory agent sees is not the product the ad agent sees, coordination is impossible. He also built the evaluation harness that scores agent outputs and routes low-confidence actions to a human.
If you cannot measure the accuracy of an action, you should not automate it
Nurtilek took execution safety. He built the runtime that lets irreversible actions run safely: consequential actions are wrapped in merchant-defined permissions, effects are simulated before anything is committed, auditable traces are recorded, and high-risk actions pause for approval. The failure mode he points to is not an email with the wrong tone. It is a price that loses money on every unit for three weeks before anyone notices. When the transactions are physical, mistakes accumulate as cash and inventory.
His landed cost engine follows the same logic. The United States suspended the de minimis exemption for low-value imports in 2025, and the White House confirmed the continuation in February 2026. Customs and Border Protection then codified an indefinite suspension in June 2026. According to the source article, roughly 4 million parcels a day now require formal customs treatment as a result. Classification codes must be assigned per product and landed costs recalculated per shipment. The unit of margin management dropped to a granularity that a monthly human process cannot follow.
In other words, what makes the custody model viable is not model intelligence. It is three properties: the action can be measured, it can be stopped, and it can be traced afterwards. Automation missing any of the three is simply a mechanism for being wrong faster.
Agents are arriving from both sides of the transaction
The other current matters too: buying agents such as ChatGPT that search, compare and send shoppers to a store.
Adobe Analytics, drawing on more than a trillion visits to U.S. retail sites, found AI-referred traffic in Q1 2026 up 393% year over year. The quality is the more interesting part. As of March 2026, visitors arriving via AI converted 42% better and produced 37% higher revenue per visit than visitors from other channels. They also spent 48% longer on site and viewed 13% more pages. Given that AI traffic converted 38% worse in March 2025, that is a fast reversal.
One caveat is worth stating. The source article frames the 42% as an advantage over paid search, but checking closer-to-primary reporting including TechCrunch shows Adobe's comparison is against all non-AI traffic, including organic, email and affiliate, not paid search alone. In practice, knowing the denominator is more useful than the headline number.
Growing traffic means nothing if the destination is unreadable to machines. In the same Adobe research, retail product pages averaged just 66% on machine readability. We covered that thread in detail in U.S. Retail Sites Average 61% Machine Readability.
The seller side is moving in parallel. Shopify now provides a Storefront MCP server for each store, giving AI agents a direct path to retrieve products and complete purchases (Shopify Sets the Standard for AI Commerce with Storefront MCP Servers and Universal Commerce Protocol). What makes Siml interesting is that it advertises marketing to AI shopping agents rather than only to humans. When the buyer is an agent and the seller's operations are agents too, a large share of commerce becomes machine-to-machine. What the human store manager used to contribute was judgment in between. How far that judgment can be delegated is the question of the next few years.
Do not evaluate the risk using only the words of the people selling it
The other side of the enthusiasm deserves a cold look. In June 2025 Gartner predicted that more than 40% of agentic AI projects will be canceled by the end of 2027. The forecast rests on a survey of more than 3,400 organizations, and the cause is attributed not to immature technology but to decisions made by the organizations deploying it. Anushree Verma, senior director analyst at Gartner, has said that most current projects are early-stage experiments or proofs of concept driven by hype and often misapplied.
The same research names a phenomenon it calls agent washing: existing chatbots and automation tools rebranded as agentic without delivering autonomy. Of the thousands of vendors claiming agentic capability, Gartner estimates only around 130 offer the real thing. Placed in that context, the way to judge Siml is not by the list of things it says it can do, but by whether the mechanism for returning low-confidence actions to humans, the pre-execution simulation and the audit trail genuinely function.
Liability is unsettled as well. If an agent sets a losing price, or auto-generated creative damages a brand, who carries the loss? Merchant operating rules, vendor contracts, and national advertising and labeling regulation have not yet been written with this arrangement in mind. The founders themselves acknowledge that surviving a bad operational week will be the real test. Merchants who remember how rules-engine automation failed silently in the past have earned their skepticism.
Where Japanese merchants should start
Siml is not usable as-is in Japan. Its connectors cover Shopify, Amazon and TikTok Shop, which leaves Rakuten Ichiba and Yahoo! Shopping out. The design principles, however, can be adopted ahead of any tool.
Start by deciding permissions and thresholds. Which mistakes in your operation are still recoverable? How deep a discount may be applied automatically? Up to what order value may a return be auto-approved? Those lines can be drawn on paper today, regardless of whether you adopt AI. Work with no line drawn cannot safely be handed to a part-time employee either.
Measurement comes next. Ask whether an automated action can be scored after the fact: repeat contact rate for support replies, gross margin for pricing, stockout rate and excess inventory days for inventory. Automating where you cannot measure means losses accumulate without ever surfacing. Zuhayr's line works perfectly well as a vendor selection criterion.
Then prepare the inbound side. Since AI-referred visits are becoming high-quality traffic, adding structured data to product pages and verifying that price and stock appear in the HTML before JavaScript executes are among the clearest returns available. The groundwork for being found costs less to start than automating operations.
Conclusion
A thousand stores is still a small number. What matters is that the focus of AI support in e-commerce is shifting from making work faster to holding the judgment and the execution, and Siml has shown one implementation of that shift. If the pattern of agents choosing on the buy side and agents operating on the sell side takes hold, a merchant's job moves from doing the work to designing permissions and measurement.
Three things are worth watching next: when Siml discloses its pricing and its funding, whether the wave of cancellations Gartner described actually arrives through 2027, and how liability rules settle once an automated action goes wrong. The craft of delegating is about to matter more than the technology that makes delegation possible.
