The prevailing vision for AI-enabled commerce is that a personal agent will shop on your behalf. Tell it you need new running shoes, and it already knows your size, budget, style, preferred brands, and injury history. It searches across retailers, compares inventory, reviews, prices, and return policies, then places the order without you ever visiting a storefront.
AI shops so you do not have to. That works well when you have a specific need to fulfill. But much of shopping is discovery. People browse, compare, change their minds, and develop preferences as they go. Sometimes the thing you buy is not the thing you set out to find. Rather than remove the shopper from the experience, AI can make the store itself more intelligent, adapting what it shows each person based on what they do in real time.
Sidd Motwani met Ian Anderson at Spotify, where they built the systems behind more than 90% of its recommendations. Those systems learned from more than 800 million listeners across a billion-item catalog and made Discover Weekly feel like it knew you. Now, Sidd and Ian have teamed up with Shivaditya Sinha to build Malachyte, the missing intent layer for commerce.
Today, Malachyte announced $10 million in funding co-led by Gradient and Bessemer Venture Partners.
Where the intelligence went
For the past decade, most of the machine learning investment in commerce went toward finding shoppers, not understanding them once they arrived. Google, Meta, and TikTok accumulated enormous amounts of behavioral data and monetized it through increasingly sophisticated advertising systems. Brands got better at buying the right traffic, while the storefront itself changed far less.
Building Spotify-quality personalization in-house remained out of reach for nearly every retailer. I spent years at Meta building infrastructure for machine learning teams, and the experiences people remember as effortless are anything but. They depend on large teams defining features, maintaining pipelines, and continually retraining models as behavior changes. That investment makes sense for a platform serving billions of users, but it does not make sense for a retailer.
As a result, the page receiving a precisely targeted shopper often still runs on the software of 2015: batch models retrained overnight, rules based on purchase history, and rows of “customers like you also bought.” That approach works tolerably for logged-in regulars and poorly for everyone else, which matters because most shoppers are neither.
At Jordan Craig, one of Malachyte’s customers, 80% of traffic comes from first-time visitors with no purchase history. Most of the shoppers the brand pays to acquire are effectively invisible to the software responsible for converting them. A decade of progress in targeting has been landing on storefronts that barely changed.
What changed?
Two things make this fixable in 2026 that were not true in 2019.
The first is that the technology now travels. The systems Sidd and Ian helped build at Spotify represented each listener as a continuously updated set of preferences, revised as that person listened, skipped, searched, and saved. The models and infrastructure required to do this have become cheap and fast enough to run on someone else’s storefront.
Malachyte trains a model on each brand’s first-party behavioral data and updates its understanding of a shopper in under 200 milliseconds, using the clicks, dwell time, comparisons, and abandoned carts from the session underway. It does not depend on a login or prior purchase history, and a brand can be onboarded in roughly a week. What was once an in-house luxury available to perhaps a dozen companies is becoming something any consumer brand can use.
The second is that the old workaround is breaking. The commerce stack of the past decade relied heavily on identity: knowing who a shopper was and connecting that person to a history of previous behavior. Privacy restrictions and logged-out browsing have made that identity less reliable. For an anonymous visitor, behavior within the current session is often the most immediate signal available.
The results are directly measurable. Malachyte runs three-week trials against a brand’s incumbent vendor using the brand’s own traffic. Those trials have so far produced a 31% lift in revenue per visitor at Fun.com, an 80% lift in add-to-cart click-through at Brunt Workwear, and a 17% lift in revenue per visit from new visitors at Jordan Craig, while holding latency below 200 milliseconds through Cyber Monday traffic.
The Future of Recommendations
Search and recommendations are the entry point because their return can be measured within weeks. Once Malachyte understands what a shopper wants in real time, the same intelligence can rank category pages, sequence email and SMS, inform ad spend, and eventually connect online behavior with what happens inside a physical store.
Agents make the timing more urgent. McKinsey projects that AI agents could orchestrate as much as $1 trillion in US consumer retail by 2030. An agent shopping on someone’s behalf will not navigate a storefront through banner ads, brand habit, or visual merchandising in the same way a person does. Retailers will need systems capable of interpreting intent and exposing the right products in real time, whether the buyer is a person or software. Malachyte is building that merchant-side intelligence layer.
Google turned a few typed words and the structure of the web into a remarkably accurate understanding of which page someone wanted. Malachyte is applying the same underlying principle: use sparse behavioral signals to proactively understand what someone is actually looking for.
Brands that understand intent at the exact moment should convert more of the demand they already pay to create. If you run a consumer brand you should be talking to Malachyte.