The Real Cost Problem in Your AI Stack Isn’t the Model — It’s the Waste
Every cross-border operator I know is running the same experiment right now. They’ve got an AI agent drafting Amazon listing copy, another one summarizing competitor reviews from Helium 10 exports, and a third one triaging customer service tickets in three languages. And every single one of them is getting a bill that makes no sense relative to the output. You’re paying GPT-class reasoning prices for the step that just reformats a CSV column. You’re paying for the same semantic answer twice because your agent doesn’t remember it already solved that problem ten minutes ago. The model isn’t the bottleneck — the routing is. That’s the problem that IQ Routing is trying to solve, and it’s the problem every seller with a serious AI toolchain should be watching.
The framing that matters for us isn’t “another AI wrapper.” It’s the shift from treating every API call as an isolated transaction to treating an entire agent run as one economic unit. For anyone running a DTC operation where margins are already thin after ad spend and fulfillment fees, that distinction is the difference between AI being a cost center and AI being a lever.
Why Your Agent Runs Are Bleeding Money (And Why Request-Level Optimization Fails)
The core insight from the maker, George Avila, is that most routing tools look at a single request in isolation. They ask “how do I make this one call cheaper?” But an agent run isn’t a pile of independent requests — it’s a flow. Some steps are busywork. Some steps are context-building. A few steps are the actual decisions that determine whether the final output is worth anything at all.
Think about what that means for a real workflow you’re running this week. Let’s say you’re using an AI agent to draft a product launch email sequence in Klaviyo. The run involves pulling your past campaign data, analyzing which subject lines performed, drafting the new copy, and then reformatting it into HTML. The model that’s smart enough to decide “this subject line underperformed because it’s too long” is wildly overqualified for the step that converts your draft into a table. But if your tooling pays the same rate for every step, you’re subsidizing the dumb work with the smart work.
The existing tools in this space — the ones that sit between your agent and the model providers — are mostly doing what Avila describes: they’re looking at each request on its own and trying to make that single request cheaper or faster. That’s the approach you’d take if you were optimizing a batch job. But agents don’t work like batch jobs. They work like workflows. And workflows have a trajectory.
The problem with optimizing at the request level is that you lose the context of the run. A request that reads identically can sit at a very different point in the run. The first time your agent asks “what’s our return rate for the UK market?” it’s doing discovery. The fifth time it asks the same question, it’s verifying a number it already computed. Paying the same price for both is the waste.
What IQ Routing Actually Does Differently
IQ Routing sits between your agent tools and the model providers, looks at each step in the run, and picks the cheapest model that can still do that step well. It also reuses work it has seen before, so you’re not paying repeatedly for answers the system already knows. The setup is meant to be boring in a good way: point your existing tool at IQ, keep your code the same, and keep working. It works today with Claude Code, Codex, and other popular apps.
The “boring” part is actually the most important feature for a cross-border operator. You’re not rebuilding your stack. You’re not migrating your data. You’re pointing your existing tool at a different endpoint and the economics change. That’s the kind of adoption curve that actually works in a business where the person making the decision is also the person who has to ship product today.
The model selection logic is the first piece of the value proposition. Instead of your agent defaulting to the strongest model for every step, IQ Routing evaluates each step and routes it to the cheapest model that can still do that step well. The context-building step that just needs to summarize a product description gets a cheaper model. The decision step that determines whether your ad copy converts gets the expensive model. You’re spending more only where the work actually needs it.
The second piece is the caching. And this is where the conversation gets interesting, because the comment thread on the launch page shows exactly how the maker thinks about it.
Where the Cache Math Gets Tricky
A commenter named Dmitriy Semenkevich asked the sharp question: does the cache reuse cross runs? If a step gets handed an answer produced under a different set of earlier steps, then a request that reads identically can sit at a very different point in the run. That’s a real concern for anyone who’s seen a cached answer poison a downstream decision.
Avila’s answer is the right one. The cache key for any request with prior conversation turns is built from the full history up to that point, not just the current message. A cache hit only occurs when the entire trajectory leading into that step matches as well. So a change anywhere earlier in the run changes the key and produces a miss instead of a hit. The one exception is the first call of a run, before any history exists to hash — that request is matched on content alone.
That’s the mechanism that rules out a step being served an answer computed under a different set of earlier steps. It’s the right design. But it also means the cache is only as good as your run’s consistency. If your agent runs are highly deterministic — same inputs, same order, same steps — the cache will crush your costs. If your runs are chaotic and every run has a different trajectory, the cache will mostly miss, and you’re back to paying for the model selection savings alone.
Why Amazon Sellers Should Care More Than Shopify Ones
Here’s where I’m going to be opinionated. The Shopify crowd talks a lot about AI tooling, but the Amazon Seller Central crowd is the one that’s actually feeling the cost pressure. Amazon sellers are running agents that pull PPC data, analyze review sentiment, generate listing variations for A/B tests, and draft dispute responses for suspended listings. These are long, multi-step runs with a lot of context-building. And Amazon sellers are also the ones with the thinnest margins — the fee structure takes a cut on every dollar, and the ad platform takes another cut. AI spend that runs 30% over what it should be is a real problem, not a theoretical one.
Shopify operators have more SKU-level control and usually fewer marketplace constraints, so they can absorb AI waste more easily. Amazon sellers are operating in an environment where the marketplace already takes its pound of flesh. Every dollar wasted on an overqualified model for a busywork step is a dollar that could have gone toward a bid adjustment or a better listing image. The routing problem is existential for them.
What Cross-Border Sellers Can Steal From This Playbook
You don’t have to adopt IQ Routing to benefit from the thinking behind it. The principle — spend more only where the work actually needs it — applies to your entire tooling stack, not just your AI spend.
The first thing to audit is your current agent runs. If you’re using an AI tool for listing generation, for review analysis, for ad copy variations — map out the steps. Which steps are actually decisions that determine the final output quality? Which steps are just formatting, context-gathering, or busywork? If you can’t tell the difference, you’re overpaying. The tool doesn’t need to be smart to reformat a table. It needs to be smart to decide which table format will convert.
The second thing is to look at your repeat work. How many times does your agent re-answer the same question in a single run? How many times does it re-fetch the same data? If you’re running a weekly listing optimization workflow and the agent is re-pulling the same sales data every time it needs a number, you’re paying for redundant computation. A caching layer — whether it’s IQ Routing or just a smarter workflow design — eliminates that.
The third thing is to think about your model selection policy. If you’re defaulting to the strongest model for everything, you’re leaving money on the table. The strongest model is for the decisions that matter. The cheap model is for the steps that just need to be adequate. This is basic portfolio theory applied to compute spend.
Where the Math Breaks
Let’s be honest about the limits. The model selection savings are real, but they’re bounded by the mix of steps in your runs. If your agent runs are mostly decision-heavy — if every step genuinely needs the strongest model — then routing saves you almost nothing. The caching is where the real money is, but it only works if your runs are consistent.
There’s also the latency question that the launch page doesn’t address. If IQ Routing is evaluating each step and deciding which model to route to, that evaluation itself takes time. For a seller running a real-time chat agent for customer service, an extra 200 milliseconds per step might be unacceptable. For a batch workflow that runs overnight, it’s irrelevant. The tool is probably better suited to the latter than the former, and the launch page doesn’t make that distinction.
The pricing is also not disclosed. That’s fine for a launch — you’re trying to get people to test the routing quality first — but it makes the ROI math hard to run. If the tool charges a percentage of your model spend, it’s worth it if the savings exceed the fee. If it charges a flat subscription, you need to run enough volume to justify it. Without pricing, you’re flying blind.
The Tooling Stack You Should Actually Be Running
The broader lesson here is that your AI tooling stack has the same economics as your logistics stack. You don’t ship every order via expedited international courier. You route the cheap, slow stuff through standard shipping and reserve the expensive, fast stuff for orders that need it. Your AI spend should work the same way.
I’m seeing more operators build this routing logic into their own stacks. They’re using LangChain or similar frameworks to build their own model selection layers, and they’re using Redis or similar key-value stores to build their own caches. You don’t need a dedicated routing tool if you have the engineering capacity to build it yourself. But most cross-border sellers don’t have that capacity — they’re running a three-person team and they need a tool that works out of the box.
That’s the gap IQ Routing is filling. It’s not the most sophisticated solution to the problem — a custom-built routing layer with a semantic cache would be more tailored to your specific workflows. But it’s a solution that exists today, that works with the tools you’re already using, and that doesn’t require you to rebuild anything. For a seller who’s bleeding AI spend and doesn’t have a dedicated ML engineer on staff, that’s worth a test.
What I’d Watch / Test Next
Here’s what I’d do this week if I were running a cross-border operation with meaningful AI spend.
First, audit one workflow end-to-end. Pick your highest-volume agent run — the one you execute most frequently — and map every step. Tag each step as “decision” or “context” or “busywork.” If more than half of your steps are context or busywork, you’re a candidate for routing. If your runs are mostly decisions, skip the routing tool and focus on caching instead.
Second, test IQ Routing on a non-critical workflow. Point it at your listing draft generator or your review summarizer — not your customer-facing chat agent. Run it for a week and compare the output quality and the cost against your current setup. The launch page explicitly invites you to try it and break it, so the maker is looking for feedback on where the routing feels wrong. That’s a signal that the tool is early but that the maker is responsive.
Third, look at your repeat computation. If you’re running the same analysis on the same data more than once a week, build a simple cache. You don’t need a sophisticated tool — a spreadsheet that stores your known answers is a start. The point is to stop paying for answers you already have.
Fourth, watch the pricing announcement. The economics only work if the routing fee is less than the savings. If the fee is a percentage of spend, do the math on your current bill. If it’s a flat fee, calculate your break-even volume. Don’t adopt a tool that saves you 20% on model spend but charges you 25% for the privilege.
The bottom line is this: the AI cost problem in cross-border e-commerce isn’t going away. The models are getting more expensive, the agents are getting more complex, and the margins are getting thinner. The operators who figure out how to route their compute spend intelligently are the ones who’ll still be profitable when the next fee increase hits. IQ Routing is one tool in that fight. The thinking behind it — spend more only where the work actually needs it — is the part you should steal regardless of whether you adopt the tool.
