The Sniff Test for When to Pivot Your AI Startup

Three to four months into building an AI-powered patent infringement detection tool, Jesse Xu and his co-founders had a realization that killed their original product idea. They couldn’t tell if their own AI was right.

Jesse is CTO and co-founder of Podqi, an AI platform that now automates the detection and takedown of counterfeit goods and trademark infringements across online marketplaces. But Podqi didn’t start there. The company started in patents — and the story of why they left is one of the cleanest pivot frameworks I’ve heard from a technical founder.

The problem with building what you can’t judge

The original idea made sense on paper: use AI to detect patent infringement across products and listings. Patent disputes are expensive, slow, and manual. Automating detection seemed like a clear win.

Then reality hit.

“The sniff test for us was us three together, we’re not patent lawyers. If we were to train an AI on how to recognize patent infringement, we wouldn’t be able to tell you what is right or wrong,” Jesse explains.

This is the kind of insight that separates founders who pivot early from founders who burn eighteen months before admitting a problem. If the people building the AI can’t evaluate its output — even roughly — the feedback loop is broken. You can’t iterate on something you can’t judge. Every training run, every model improvement, every edge case requires bringing in an expensive outside expert just to know if you’re making progress.

The sniff test as a framework

Jesse’s pivot decision wasn’t based on market size or competitive analysis. It was based on a single question: Can the founding team validate the AI’s output themselves?

This is worth unpacking because it applies to any AI startup, not just IP enforcement:

• Can you look at the AI’s output and know if it’s roughly right? Not perfectly — roughly. If you need a domain expert for every evaluation, your iteration speed drops to whatever pace that expert is available.
• Can you build your own training data? If you can’t label examples accurately, you can’t supervise the model. You’re outsourcing your core competence.
• Can you explain to a customer why the AI flagged something? If you can’t articulate the reasoning, you can’t debug false positives, handle complaints, or build trust.

Patent infringement failed all three. Trademark infringement — matching logos, brand names, product images against known assets — passed all three. The founding team could look at a flagged listing and immediately tell whether it was a real counterfeit or a false positive. They could build and label training data. They could explain every detection to a customer.

Speed as the real variable

The pivot happened in three to four months. That’s fast — most startups take six to twelve months to abandon a bad direction, if they abandon it at all. Jesse attributes the speed to the sniff test being binary. There was no ambiguity. They couldn’t validate patent infringement output, period. No amount of persistence or iteration would change that without hiring patent lawyers as full-time team members.

Contrast this with the slow pivots that drain startups: the product is kind of working, customers are kind of interested, the metrics are kind of trending up. Those “kind of” situations are where companies waste years. Jesse’s framework cuts through the ambiguity by asking a question that has a clear yes or no answer.

What this means for AI builders

The output validation sniff test is especially relevant now because the barrier to building AI products has dropped to near zero. Anyone can wrap an API call and ship something in a weekend. The hard part isn’t building — it’s knowing whether what you built actually works.

If you’re in a domain where you can’t evaluate your own model’s output, you have three options: hire domain experts to the founding team (expensive and slow), partner with domain experts who will provide continuous feedback (fragile), or pivot to a domain where you can validate the output yourselves.

Jesse chose door three. Podqi now processes takedowns across Amazon, Walmart, Meta, Google, and the open web. “We have a number of classifiers, we match against any assets that we detect. We try to link to any existing entities that we’ve seen,” he says. Every detection is something his team can verify without calling a lawyer.

The lesson is portable. Before you commit to a problem space, run the sniff test. If you can’t tell whether your AI is right, you’re not building a product — you’re building a dependency.

FAQ

How do you know when to pivot your AI startup?

Apply the output validation sniff test: can the founding team evaluate the AI’s output without relying on outside domain experts? If no, your feedback loop is broken — you can’t iterate, label training data, or explain results to customers. This was the framework that drove Podqi’s pivot from patents to trademarks in under four months.

What is the sniff test for AI product-market fit?

The sniff test asks three questions: Can you tell if the AI’s output is roughly correct? Can you build and label your own training data? Can you explain a flagged result to a customer? If the founding team fails any of these, the problem domain may be wrong for the team — regardless of market opportunity.

How fast should a startup pivot?

The strongest pivots happen in three to four months, when the signal is binary — either the team can validate the AI’s output or it can’t. Slow pivots (six to twelve months) typically happen when metrics are ambiguous — “kind of working” situations where founders delay the decision because there’s no clear failure point.

Why did Podqi pivot from patents to trademarks?

The founding team discovered they couldn’t evaluate whether their patent infringement AI was producing correct results without consulting patent lawyers. Trademark infringement — matching logos, brand names, and product images against known assets — was a domain where the team could visually verify every detection. The same AI capabilities applied, but the validation loop was immediate.

What makes some AI problems harder to validate than others?

Problems with clear visual or factual ground truth (counterfeit detection, image matching, data extraction) are easier to validate because non-experts can judge output accuracy. Problems requiring specialized legal, medical, or scientific interpretation (patent law, diagnosis, regulatory compliance) create validation dependencies that slow every iteration cycle.

How do AI startups build domain expertise for their products?

Three approaches: hire domain experts to the founding team, partner with external experts for continuous feedback, or choose a domain where the founding team can already validate output. The third option preserves iteration speed because feedback loops don’t depend on external availability.

What is the most common mistake AI startups make when choosing a problem?

Choosing a problem with a large market but no internal capability to validate AI output. Market size means nothing if the founding team can’t iterate independently. The strongest AI startups build in domains where the team can evaluate, label, and explain every model output without outside help.

How does Podqi’s AI detect counterfeit products?

Podqi uses multiple classifiers that match product listings against a brand’s assets — logos, images, trademarks, and known counterfeiter patterns. The system prioritizes recall (catching everything possible), then refines precision over time. Detections are linked to known entities and counterfeiter networks across platforms.

Should you pivot if your AI product is partially working?

Partial results are the most dangerous signal for an AI startup. “Kind of working” delays hard decisions by months. Apply the sniff test: if the founding team still can’t independently validate output quality after three to four months of iteration, the problem-team fit is wrong regardless of early traction.