On July 9, OpenAI launched GPT-5.6 and ChatGPT Work. The announcement was framed around productivity: an agent inside ChatGPT that can gather information across your apps, create finished materials, and stay with complex projects for hours. Over five million people use Codex every week. One million of them use it for work outside software development.
The subheadline that matters: agents are now mainstream infrastructure. They connect to Slack, Google Drive, SharePoint, email, calendars, CRMs. They take actions. They work independently for extended periods. OpenAI’s own guidance frames this as a portfolio investment problem: broad access for everyday productivity, function-specific workflows, and strategic bets built around proprietary context.
But there is a category of agentic work that OpenAI’s launch materials barely mention, and it is the one with the highest stakes: commerce decisions. When an agent inside ChatGPT Work is asked to “find the best project management software for our 50-person team and purchase three annual licenses,” it is making a commerce decision. When a finance agent reconciles vendor spending and recommends switching suppliers, it is making a commerce decision. When an office manager agent restops supplies through Amazon Business, it is making a commerce decision.
These decisions already happen inside enterprises every day. The difference is that humans make them, slowly, with all the cognitive biases and inefficiencies that entails. Agents will make them faster. They will also make them on corrupted data unless the infrastructure problem is solved.
The Agentic Commerce Gap
OpenAI’s “managing AI investments in the agentic era” guidance includes five pillars: sharpen visibility into usage and spend, evaluate model efficiency by outcome ROI, govern advanced workflows, fund workflows that compound, and match capacity to proven demand.
It is excellent guidance for enterprise AI adoption. It is silent on the question of what happens when agents interact with external commerce systems that are actively designed to manipulate them.
Amazon’s marketplace is the most obvious example. Search results are shaped by paid placement algorithms. Review systems are gamed by sellers who purchase fake reviews, suppress negative feedback, and hijack ratings from unrelated products. Reference prices are fabricated to create artificial discounts. A human shopper brings skepticism to these signals. They read reviews critically. They cross-reference prices. They notice when something feels off.
An agent does none of this. When ChatGPT Work queries Amazon through a plugin or API, it receives structured data: star ratings, review counts, pricing, feature lists. It processes this data as input for a recommendation. It does not pause to wonder whether 4,000 of the 12,000 reviews were purchased from a review farm. It does not cross-reference the star rating against independent review fraud detection. It computes an answer and presents it with confidence.
This is the agentic commerce gap. The model is excellent at reasoning. The data it reasons over is corrupted. Better models do not fix this. GPT-5.6 is more capable than GPT-5.4. But a more capable model processing corrupted data produces more confident wrong answers, not better ones.
Why Token Economics Make This Worse
OpenAI highlighted that GPT-5.6 achieves better performance with 54% fewer output tokens and 57% less time per task compared to previous models. Token prices have fallen 97% from GPT-4 to GPT-5.4. This means agentic workflows are getting dramatically cheaper to run.
Cheaper agents mean more agents. More agents mean more commerce decisions made without human oversight. The cost of an agent researching and recommending a product purchase is approaching zero. Enterprises will deploy agents to handle procurement, vendor selection, supply chain optimization, and purchasing at a scale that humans simply cannot match.
Each of those agent-led transactions depends on the quality of the underlying product data. If that data comes directly from Amazon or other marketplaces without a verification layer, every agent-led purchase carries the same risk of review manipulation that human purchases do. But at agent scale, the risk compounds. A thousand agents making procurement decisions based on fake reviews is not a thousand isolated mistakes. It is a systemic misallocation of enterprise spending.
The Trust Layer Imperative
OpenAI’s governance guidance gets close to this issue without naming it directly. The third pillar, “govern advanced workflows before they scale,” recommends defining what context ChatGPT can use, which tools it can access, and what actions it can take. For commerce workflows, this governance must include a trust verification step.
The architectural pattern is straightforward. Before an agent makes a purchase recommendation, it should consult a trust layer that has already verified the underlying product data. The agent queries GoBuy’s MCP server at gobuy.ai/api/mcp, which has filtered fake reviews, detected manipulation patterns, and computed Smart Scores from 0 to 100 based on review quality rather than quantity. The agent reasons over verified data instead of corrupted data.
This is not a theoretical concern. GoBuy’s analysis of Amazon product data consistently finds significant gaps between apparent quality and verified quality. Products with 4.8-star averages from 12,000 reviews frequently show verified scores 20 to 30 points lower once fake reviews are filtered and authentic reviews are weighted appropriately. Sponsored listings that dominate search results often fail trust verification entirely.
An agent equipped with this intelligence makes different recommendations. Not because the model is smarter, but because the data is better. The trust layer does what the model cannot: it separates signal from noise before the model ever sees the data.
The MCP Connection
The Model Context Protocol makes this trust layer pattern accessible to every agent framework. Claude, ChatGPT, Cursor, and VS Code Copilot all support MCP natively. An agent does not need a bespoke integration to access GoBuy’s product intelligence. It needs an MCP client and a connection to gobuy.ai/api/mcp.
As OpenAI noted in their launch materials, plugins connect ChatGPT to external tools and systems. MCP provides the standardized protocol for those connections. The agents that produce the best commerce recommendations will not be the ones with the most powerful reasoning models. They will be the ones connected to the best intelligence layers.
This is the same insight OpenAI applied to coding. Codex works well not just because GPT-5.6 is a strong reasoning model, but because it connects to code repositories, test runners, and deployment systems that provide verified context. The same principle applies to commerce. A shopping agent works well when it connects to a trust layer that provides verified product data.
What Enterprises Should Do Now
For organizations adopting ChatGPT Work or building custom agentic workflows, commerce trust should be part of the governance framework from day one.
Audit which agent workflows involve purchasing decisions. Map every workflow where an agent recommends, compares, or selects products, vendors, or suppliers. These are your commerce-exposed workflows.
Insert a trust layer before the recommendation step. For Amazon product data, connect GoBuy’s MCP server. For other marketplaces, evaluate whether the data source has manipulation risks and whether a verification layer exists.
Track trust-adjusted outcomes. Measure not just whether the agent completed the task, but whether the recommendation was sound. Did the product actually meet the need? Did the vendor deliver as promised? Trust-adjusted outcome metrics are the commerce equivalent of OpenAI’s “cost per accepted outcome.”
Build trust verification into approval gates. For high-value purchases, require the agent to present both the raw marketplace data and the trust-verified data before a human approves the transaction. Discrepancies between the two are red flags.
Prioritize MCP-native trust layers. Bespoke integrations create maintenance burden and vendor lock-in. MCP-native trust layers work across agent frameworks and can be swapped or upgraded without rewriting agent logic.
The Market Is Moving
OpenAI’s launch of ChatGPT Work signals that the agentic era is no longer experimental. It is a product category with millions of users, enterprise governance controls, and a clear roadmap toward deeper workflow integration. Anthropic’s Claude has similar capabilities through MCP. Apple Intelligence just received regulatory approval in China. The agent infrastructure layer is being built rapidly.
The trust infrastructure layer is not being built at the same pace. Most organizations deploying agents have not yet encountered the commerce trust gap because most agent workflows today focus on internal tasks: coding, document creation, data analysis. But as agent capabilities expand to include external commerce decisions, and as token costs continue to fall, the volume of agent-led purchasing will explode.
The organizations that build trust verification into their agent architecture now will have a structural advantage. Their agents will make better recommendations. Their procurement will be more reliable. Their spending will be based on verified product quality, not manipulated ratings.
The agents are here. The reasoning is impressive. The missing piece is trust.
Build trust into your agent stack. GoBuy’s MCP server at gobuy.ai/api/mcp provides verified product intelligence for AI shopping agents. Filter fake reviews, compute trust scores, and give your agents the data they need to make sound recommendations. Full developer docs at gobuy.ai/agent-docs.