On August 7, 2026, Canva’s CEO Melanie Perkins sent a letter to shareholders that should be required reading for anyone building AI-powered commerce. The design platform, valued at US$42 billion and generating nearly a billion dollars in quarterly revenue, had cut its growth forecast by a third. The reason was not demand. It was cost. Specifically, the cost of running AI features on frontier models from providers like OpenAI.
“The average cost of serving an AI task was too high,” Perkins wrote. “Several of our first-party models were not yet ready for release, and our pricing, consumption model and usage controls had not caught up with the outsized demand we were seeing.”
Canva slowed product rollout while it rebuilt its architecture from scratch. The result was a 90% reduction in the cost of serving an AI task, with its in-house video model running 17 times cheaper than a frontier model and its image model 30 times cheaper. Revenue for the June quarter came in at US$921.9 million, up 25.2% year over year. That is strong growth. But it is below the 30% target Canva set at the beginning of 2026, and the gap is entirely attributable to the economics of AI.
This is a story about a design tool. It is also a story about the future of shopping.
The Unit Economics Problem Nobody Is Talking About
The agentic commerce industry, the ecosystem of AI agents that help consumers discover, evaluate, and purchase products, is being built on an assumption that AI inference is cheap. That assumption is wrong.
Consider what an AI shopping agent actually does. When a user asks “find me the best wireless headphones under $100,” the agent must parse the request, query a product database, retrieve specifications for dozens or hundreds of candidates, analyze review patterns for authenticity and sentiment, compare prices across sellers, evaluate shipping and return policies, weigh all factors against the user’s stated preferences, and produce a ranked recommendation. Each step requires inference. Each inference call costs money.
At the scale of a consumer-facing platform serving millions of queries per day, those costs compound aggressively. Canva discovered this with a design tool that generates images and layouts. The cost of serving each AI task was high enough to force a revenue downgrade at a $42 billion company. A shopping agent that processes real-time product data, review analysis, and multi-factor comparisons faces a cost structure that is orders of magnitude more expensive.
The math is simple and brutal. If Canva, with its relatively constrained set of AI tasks (image generation, design suggestions, text overlay), could not sustain frontier model costs, then a full-stack shopping agent operating across millions of SKUs, thousands of review profiles, and dynamic pricing data has no path to profitability on frontier models. None.
Three Responses to the Cost Crisis, and Why Two of Them Are Dangerous
Companies building AI-powered commerce features will respond to the cost crisis in one of three ways. Two of them are dangerous for consumers.
Response 1: Build Cheaper Models (Canva’s Approach)
Canva’s solution was to invest heavily in first-party models. The acquisition of Leonardo.AI in 2024 gave Canva an in-house generative AI platform. The result: a 90% cost reduction, with video 17 times cheaper and images 30 times cheaper than frontier alternatives. COO Cliff Obrecht explained in April that the company routes tasks intelligently, using cheaper in-house models for most jobs and reserving frontier models for complex tasks that require their capability.
This approach works for Canva because design tasks have relatively clear quality boundaries. An image either looks good or it does not. A layout either works or it does not. The user provides immediate feedback by accepting or rejecting the output. Quality can be evaluated visually and quickly.
For shopping agents, the equivalent is much harder. A product recommendation does not have a visual quality check. The user cannot tell, by looking at a recommendation, whether the underlying review analysis was performed by a frontier model that detected manipulation patterns or by a cheaper model that simply counted star ratings. The quality degradation is invisible. The recommendation looks confident. The reasoning sounds plausible. But the cheaper model may have missed the review suppression that removed authentic negative feedback, or the seller-manipulated rating profile that inflated the score above its true quality level.
Building cheaper models for commerce is not just an engineering challenge. It is a trust challenge. The cost savings come from cutting corners on the analytical depth that makes AI recommendations worth having in the first place.
Response 2: Use Seller-Provided Data (The Cheap Shortcut)
The second response to cost pressure is the most dangerous: skip independent analysis entirely and use data provided by the marketplace or the seller. Amazon product pages already contain descriptions, specifications, review summaries, and ratings. An AI agent can read this data with minimal inference cost. No need to analyze millions of reviews for authenticity. No need to cross-reference pricing history. No need to build independent trust signals. Just read the page and summarize.
This is the path of least resistance. It is also the path that replicates every existing problem with online shopping. Seller-provided descriptions are marketing copy. Platform-provided ratings are algorithmically filtered and manipulated. Review summaries are shaped by suppression systems whose criteria are opaque. An AI agent that simply reads and summarizes this data is not adding intelligence. It is adding a confident voice to a compromised dataset.
The FTC’s July 2026 enforcement actions illustrate the risk concretely. On July 15, the Commission finalized an order against TruHeight, a supplement company that used employee-written reviews, incentivized 5-star reviews, and fake bot profiles to manufacture credibility. The company’s products appeared highly rated on Amazon. Consumers, and any AI agent analyzing those listings, would have seen a trustworthy product. The deception was real. The $4 million judgment, partially suspended to $750,000 based on inability to pay, reflects the FTC’s assessment of the harm.
An AI shopping agent that uses seller-provided data without independent verification is structurally incapable of detecting this kind of manipulation. It reads the reviews, counts the stars, and recommends the product. The cost is low. The accuracy is zero. The consumer gets a recommendation that is worse than no recommendation at all, because it carries the false authority of AI analysis.
Response 3: Independent Trust Infrastructure (The Hard Path)
The third response is the one that actually works for consumers: build independent trust infrastructure that does not depend on marketplace data. This is architecturally similar to what Canva did with its AI costs, but applied to data integrity rather than inference compute.
Independent trust infrastructure means analyzing review patterns from sources beyond the seller’s control. It means maintaining historical pricing data that reveals whether a “discount” is genuine. It means filtering fake reviews using detection models trained on known manipulation patterns. It means scoring products on the quality of authentic reviews, not the quantity of potentially fake ones.
This approach costs more than reading seller-provided data. It requires maintaining databases, running detection models, and continuously updating trust signals. But it produces recommendations that are actually worth the consumer’s trust. The cost is an investment in accuracy. The alternative is a cheap system that produces confident nonsense.
The MCP Protocol: The Plumbing for Trustworthy Agents
The Model Context Protocol (MCP), the open-source standard for connecting AI applications to external systems, provides the technical infrastructure for this independent trust layer. MCP functions as what the protocol’s documentation calls “a USB-C port for AI applications,” a standardized way for AI agents to connect to external data sources, tools, and workflows.
For agentic commerce, this architecture is critical. An AI shopping agent that connects only to Amazon’s product API is trapped in the marketplace’s data ecosystem. It sees what the marketplace wants it to see. Its recommendations are shaped by the marketplace’s ranking algorithm, sponsored placements, and review suppression systems. The agent may be intelligent, but its intelligence is applied to a curated, commercially biased dataset.
An MCP server that provides independent trust data breaks this constraint. An agent that consults an external trust scoring service, one that has analyzed review authenticity across sources the marketplace does not control, gains access to signals that the marketplace’s data does not contain. The agent’s recommendation shifts from “the highest-ranked product on this platform” to “the product with the strongest evidence of genuine quality.”
This is the architecture that agentic commerce needs. Not agents that are smarter at reading compromised data, but agents that have access to uncompromised data. The intelligence of the system is bounded by the integrity of its inputs.
The Grokipedia Lesson: Silent Degradation Is Already Here
The Canva cost crisis is a warning about what happens when AI economics are unsustainable. But there is an equally important warning about what happens when AI-maintained systems degrade silently.
In August 2026, Lawfare published an investigation revealing that Grokipedia, xAI’s AI-generated encyclopedia with over 6 million articles, has not processed a single edit since April 24. Over 13,000 suggested corrections sit in a dead-end queue. The site still loads. The articles still display. But the editorial system is frozen. SpaceX’s June 12 IPO, one of the most significant business events of the year, is not mentioned on Grokipedia’s SpaceX page despite multiple users submitting the correction.
The Tow Center for Digital Journalism documented that by December 2025, 57.8% of Grokipedia’s edits were AI-authored, creating a closed loop where the system edited itself. A mass rewrite on March 14 broke the anchoring system and retroactively reclassified previously accepted edits as “rejected.” The audit trail became unreliable. The live transparency feed went dark between January and March.
The parallel to commerce is direct. Amazon product pages that have not been updated in months look identical to current ones. A product whose reviews were written for a previous manufacturing run, with better components, still displays a 4.7-star rating. The information is frozen, but the product has changed. An AI agent that reads this data has no way to detect the freeze.
Grokipedia received 6.7 million visits in June 2026 and was cited as a source across roughly 356,000 references in AI systems including ChatGPT and Google AI Mode, according to an Ahrefs analysis. The stale information propagated downstream. AI systems that cited Grokipedia did not know it was frozen. They presented its content as current and authoritative.
In commerce, the same propagation happens when AI shopping agents read stale product data and pass it to consumers as analysis. The agent does not know the review profile was manipulated. It does not know the price was artificially inflated before a “discount.” It does not know the product specifications were changed by the seller after the reviews were written. It reasons perfectly over bad data and produces a recommendation that inherits every flaw.
Why This Week Matters
The convergence of these stories in the first week of August 2026 is not coincidental. It is the emerging shape of a problem that the industry has been postponing.
Canva’s disclosure proves that AI costs are a material business constraint, not a theoretical concern. A $42 billion company with nearly a billion dollars in quarterly revenue could not sustain the cost of AI features built on frontier models. The agentic commerce industry, which faces significantly more complex inference requirements, is building on the same unsustainable foundation.
The FTC’s TruHeight order proves that review manipulation is not a minor nuisance but a law enforcement priority. The agency required $750,000 in payments and a permanent ban on fake and incentivized reviews. The order sends a signal to the market, but enforcement is retrospective. The fake reviews accumulated for months before the FTC acted. During that period, any AI agent analyzing TruHeight’s products would have produced recommendations based on fabricated trust signals.
The Grokipedia investigation proves that AI-maintained information systems degrade silently. There is no alarm bell when an AI system stops processing updates. The data looks alive. The consumer has no way to detect the freeze. And downstream systems that cite the data inherit the staleness without knowing it.
The MCP protocol’s continued adoption, with support from Claude, ChatGPT, Visual Studio Code, Cursor, and other major platforms, provides the technical foundation for a solution. But the protocol is only plumbing. What matters is what flows through the pipes.
The Path Forward for Agentic Commerce
The companies that succeed in agentic commerce will not be the ones with the most capable models. Model capability is table stakes. The winners will be the ones that solve the data integrity problem at economically viable cost.
This requires three things:
Independent data sources. AI shopping agents must consult trust infrastructure that is not controlled by the marketplace. Review authenticity analysis, pricing history, and quality scoring must come from sources with no commercial incentive to misrepresent. The MCP protocol makes this technically feasible. The GoBuy MCP server at gobuy.ai/api/mcp provides exactly this kind of independent trust data, a Smart Score from 0 to 100 that weights review quality over review quantity, filters fake reviews, and surfaces only the top products rather than thousands of undifferentiated options.
Cost-aware architecture. Trust verification cannot depend entirely on frontier models. Like Canva’s approach of routing simple tasks to cheaper in-house models and reserving frontier models for complex reasoning, trust infrastructure must use tiered processing. Mechanical review filtering can be done cheaply. Deep authenticity analysis requires more compute. The system must know which products warrant deep analysis and which can be reliably scored with lighter methods.
Transparency about limitations. The most dangerous AI system is one that presents confident recommendations without disclosing the limits of its analysis. An agent that analyzed 500 Amazon reviews using a cheap model should not present its conclusion with the same confidence as one that cross-referenced independent databases, checked pricing history, and ran review authenticity detection. Consumers need to know what the agent actually verified versus what it assumed.
Conclusion
Canva’s cost crisis is a preview of the crisis coming for agentic commerce. The companies building shopping agents are about to discover what Canva discovered: AI at scale is expensive, the easy cost-cutting paths compromise quality, and the hard path requires fundamental architectural choices made early, not late.
The question is not whether AI shopping agents will be built. They will. The question is whether they will be built on independent trust infrastructure or on seller-provided data optimized for marketplace revenue. The first serves consumers. The second serves everyone except the consumer.
At GoBuy, we have built the trust layer that AI agents need before they recommend products. Our MCP server provides independent Smart Scores, fake review filtering, and curated product rankings that do not depend on marketplace data. AI agents that consult GoBuy before purchasing get the independent verification that Canva’s cost crisis proves is essential. AI agents that do not are just reading seller marketing with a confident voice.
Build with trust, or build without it. The economics will determine the outcome.
Start building with GoBuy’s trust infrastructure at gobuy.ai/agent-docs. Connect your AI agent to our MCP server at gobuy.ai/api/mcp and give your recommendations the independent verification they need.