When Algorithms Shop for You, Whose Side Are They On?
Meta is building an AI agent called Hatch and an agentic shopping tool for Instagram. What it does not yet have is an answer to the hardest question in the race to build AI shopping agents: can a company that makes its money auctioning attention build something that genuinely serves the user?
Hatch is Meta's codename for an agent built on top of Muse Spark, the model the company launched April 8. Hatch is designed to let users delegate tasks across multiple apps and services, learning from the data those tasks generate. Meta is targeting an internal testing milestone soon, The Information reported. Separately, Meta is building an agentic shopping tool directly into Instagram, targeting launch before Q4, Reuters reported.
Muse Spark already ships the technical capabilities that make Hatch possible. Meta's own announcement describes the model as natively multimodal with tool-use, visual chain of thought reasoning, and multi-agent orchestration. The model deploys subagents in parallel: one to draft an itinerary, another to compare destinations, a third to find activities. That infrastructure is live. What Hatch adds is a consumer-facing wrapper aimed at everyday transactions.
The pressure to make it work is financial. Meta raised its 2026 capital expenditure guidance to $125B-$145B in late April, up from $115B-$135B, nearly double the $72.2B it spent last year, according to Fortune. The stock fell 7% in after-hours trading. Meta halted Q2 share buybacks to preserve cash, BigGo Finance reported. On the earnings call, CEO Mark Zuckerberg called ROI a very technical question and declined to offer specific return timelines. Hatch and the Instagram shopping tool are the product-level answer: proof that the infrastructure buildout generates real engagement and commerce revenue, not just benchmark scores.
Here is the problem. Meta makes money by auctioning attention. Its ad auction system ranks bids, but also factors in predicted engagement and user fit. The auction winner gets shown to the user. Sponsored results already appear throughout Instagram's shopping surfaces. When Hatch recommends a product, it will operate in the same structural environment: an auction that has historically favored what advertisers pay, not what users need.
Meta's own history with Instagram Shopping suggests the tension is not theoretical. The platform has run sponsored product placements alongside organic recommendations since at least 2023. Meta has not disclosed what role its ad auction plays in default shopping recommendations. Meta did not respond to questions about whether Hatch's recommendation logic will be subject to the same auction mechanics.
Shopify's Sidekick and Amazon's Alexa handle the same underlying conflict differently by business model. Shopify makes money on commerce volume, so its agent has structural incentive to find genuine best matches. Amazon's retail business similarly aligns algorithm quality with revenue. Neither has published disclosure standards for how sponsored products appear in agent recommendations.
Regulatory guardrails are nascent. The EU AI Act classifies AI systems that manipulate human behavior as high risk but does not yet specifically address AI shopping agents or mandate disclosure when algorithmic recommendations are influenced by paid placement. The FTC has brought cases against deceptive AI-generated reviews but has not ruled on AI-driven product recommendations that blur organic and sponsored content. No US law requires Meta or its competitors to disclose when an agent's recommendation reflects a paid placement rather than a genuine best match.
The disclosure gap matters because the stakes are asymmetric. A search result that favors a sponsor is annoying. An AI agent that consistently routes purchases toward advertisers rather than optimal matches is harder to detect and harder to appeal. Users delegate decision-making to the agent precisely because they trust it to process more information than they could themselves. That trust is valuable. Meta knows it.
Until Hatch ships with disclosed recommendation mechanics, the answer to whose side the algorithm is on is: the evidence is inconclusive, and Meta has strong financial incentive to keep it that way. The $145B capex, the halted buybacks, and the near-term testing target all point to one thing: Meta needs this to work as a product, regardless of whether it works as intended.
The broader shopping agent race makes this more than a Meta story. Amazon, Google, and Apple are all building variations. None has published how its ad mechanics interact with agent recommendations. The industry standard is silence. Hatch is the highest-profile test case yet of whether the companies building AI shopping agents will disclose their conflicts, or whether the trust that makes agents valuable will be spent serving the business model that funds them.