
When Snowflake wrapped its Expedition 2026 event on November 3, the headline items were CoWork (a rebrand of Snowflake Intelligence), fresh Cortex Agents capabilities, MCP connectors, a Business Ontology feature, and a Cortex AI Gateway. Behind the scenes, an engineering team at EveryInfra decided to stress-test the new connectors against a genuinely messy business problem: figuring out why customers return products by joining returns data with public complaints from Amazon reviews and TikTok. Their writeup, published on dev.to, is one of the more honest pieces of AI engineering documentation you'll read. It's also, perhaps unintentionally, one of the clearest arguments for why building agents that actually work is harder than the tooling brochures suggest.
Key points
- Snowflake launched CoWork, new Cortex Agents features, MCP connectors, Business Ontology, and Cortex AI Gateway at Expedition 2026 on November 3.
- An engineering team testing the new MCP connectors found OAuth requirements blocked direct connection, forcing a scheduled ingestion workaround instead.
- A keyword-matching rule flagged 42 of 200 TikTok comments as complaints, but only 4 were real, roughly 9.5% precision, showing why language understanding beats simple rules.
- The test's total API cost was about half a cent across 12 calls, with monthly estimates around $1.26 in API fees for one brand.
- The core lesson for business owners: an agent is only as good as the data and wiring behind it, which is exactly why done-for-you builds exist.
Retail returns are a staggering cost. The National Retail Federation projected roughly $849.9 billion in US merchandise returns, about 15.8% of sales, and 19.3% for online sales. Most brands warehouse the numbers: what was sold, what came back, when. What they almost never store is the public commentary explaining why. That lives in Amazon reviews, Reddit threads, and TikTok comment sections, scattered across platforms that don't talk to each other. The EveryInfra team's idea was straightforward. Pull complaint data from those public sources, land it in tables alongside returns data inside Snowflake, then let a Cortex Agent look for correlations between complaint themes and return rates. If a lid that leaks shows up repeatedly in comments and returns spike for that SKU, you've found something a spreadsheet would never surface on its own.
The team ran 12 live API test calls on November 6, at a total cost of $0.005016. That figure deserves attention on its own: fractions of a cent per call. The economics of running agents against external data are not the barrier people assume they are. The results were mixed, and the team said so. All five Amazon review API calls came back empty (four 503 errors and one blank result), and none were charged. The TikTok side worked better. Across 273 comments on two videos, 17 described genuine product problems, with lid and spill issues the largest single theme at 6 mentions. Then came the most instructive result of the whole exercise. A keyword-matching rule applied to 200 comparison-video comments flagged 42 as complaints. Only 4 were actual product complaints. That's roughly 9.5% precision, which means the rule was wrong more than 90% of the time. An LLM-based classification approach, by contrast, is built to read context and distinguish 'this product leaked all over my bag' from 'I hope this doesn't leak, thinking of buying one.' If you've ever wondered why your spam filter or review alerting tool produces garbage, this is why. Keyword rules don't read. Models do.
Here's the finding with the biggest practical weight. Snowflake's MCP connectors currently accept OAuth only, requiring either dynamic client registration or a client registered against Snowflake's specific OAuth callback URL. EveryInfra's API-key-based server met neither requirement, so a direct MCP connection wasn't possible. The workaround was a Snowflake external access integration: a scheduled Python stored procedure pulls external data into tables, and the Cortex Agent queries those tables. That workaround isn't a hack. It's arguably the better architecture, and the team's reasoning is worth repeating. MCP connector results stay inside a single chat session and never land in tables. For durable, joinable data, scheduled ingestion is the correct pattern. Agents should read well-structured tables, not poke at live tools on every question. The cost model backs this up. Snowflake warehouse costs for the use case were estimated at about $40.54 per month running serially, dropping to roughly $4.05 with 10 parallel calls. Add the roughly $1.26 in monthly API fees, and you're looking at a few dollars to low tens of dollars per month for a brand-scale complaint analytics pipeline. The author was careful to note the limitations: no live Snowflake testing, no returns dataset available, no Amazon data retrieved. This is a build plan and a cost model, not a validated finding. That honesty is rare and useful.
You're probably not running Snowflake warehouses. But the underlying lesson applies to any business thinking about AI agents, and it's the lesson we build our practice around at Agents by AIQ. The lesson is this: the agent is the easy part. The value is in the wiring. The EveryInfra engineers, who do this for a living, hit OAuth registration requirements, network rules, secrets management, stored procedures, and scheduled tasks before a single question could be answered. If professionals with engineering backgrounds need that much configuration to connect one data source to one platform, the DIY path for a busy owner-operator is a dead end. The integration work is the product. There's also a governance angle that scales down. Snowflake's Business Ontology exists because if two agents define 'return rate' differently, you get two different answers to the same question. The same applies to a small business running an AI receptionist alongside a follow-up agent and a support agent. Someone has to agree on what counts as a qualified lead, a booked job, a resolved ticket. That definitional work is unglamorous and it's where multi-agent deployments either hold together or fall apart.
The returns use case has a direct small-business parallel. Large retailers lose enormous sums to returns they can't explain. Small businesses lose revenue to missed calls, slow lead follow-up, and inbox threads that go cold, and they usually can't see the pattern either. The fix in both cases is the same shape: connect the unstructured signals (a phone call, an email, a review) to the structured systems (the calendar, the CRM, the books), so an agent can surface what's actually happening. An AI receptionist that answers on a real phone number, a follow-up agent that reads the inquiry and responds in context, a support agent that can tell a genuine complaint from a casual mention: these are the same pattern as the returns analysis, just at the scale of one business instead of one retail category. And the keyword precision finding is the quiet warning underneath all of it. Cheap automation that pattern-matches without reading will misclassify most of what it touches, and your customers will feel it.
The Expedition 2026 announcements point to where the industry is heading: agents that take governed action across systems rather than just chatting with your data. The EveryInfra writeup points to what the industry keeps learning the hard way: getting there requires OAuth registrations, ingestion pipelines, shared definitions, and honest cost accounting. The model reads the comments. The plumbing makes the reading worth something. If you're weighing agents for your business, spend your evaluation time on the plumbing question, because that's where the outcome gets decided.