Quick Navigation

I've been in AI investing for over a decade, and one pattern keeps showing up: when a startup claims its AI is better than the competition, the real question is how much better. That's where the 30% rule comes in. It's a quick litmus test I use to separate genuine breakthroughs from incremental tweaks.

Understanding the 30% Rule in AI Investment

The 30% rule states that if an AI model's performance improvement over the existing baseline is less than 30%, the investment is likely not worth the risk. I'm talking about relative improvement—not absolute. Say the current system solves a problem with 70% accuracy. A new AI that hits 84% accuracy gives a 20% improvement ( (84-70)/70 = 20% ), which falls short of the 30% threshold. In my experience, below 30% improvements are rarely enough to shift user behavior, justify switching costs, or create a durable moat.

This rule isn't written in stone, but it emerged from analyzing hundreds of AI deals. I first heard it from a partner at a top VC firm, and after applying it myself, I saw why. Startups that showed >30% improvement tended to raise follow-on rounds; those below often struggled to gain traction.

Why 30%? The Science Behind the Threshold

Why 30% and not 20% or 40%? It's partly psychological and partly empirical. From a behavioral economics lens, a 30% improvement crosses the just noticeable difference threshold in many domains. For example, if an AI call center agent resolves 30% more issues than a human, managers notice the difference and consider replacing staff.

I once backed a startup that improved document processing speed by 28%—just under the magic number. The customers said, "It's faster, but not faster enough to justify retraining our team." They churned. In contrast, another company that delivered a 35% speed gain saw rapid adoption. The 30% number isn't arbitrary; it correlates with the point where switching costs are outweighed by benefits. Several internal studies at large enterprises I've consulted for show that a 30% improvement in accuracy or speed triggers a re-evaluation of existing workflows.

How to Apply the 30% Rule to Evaluate AI Startups

Here's my step‑by‑step process for using the 30% rule in real due diligence:

Step 1: Identify the baseline. What's the current best solution? It could be a human benchmark (e.g., radiologist detection rate) or a previous AI system. Always get numbers from an independent source, not just the startup's white paper.

Step 2: Calculate the relative improvement. Use the formula: (New Performance - Baseline) / Baseline × 100%. Be careful—a jump from 90% to 95% is only about 5.5% relative improvement, not 5 percentage points. Many founders try to fool investors with absolute percentage differences.

Step 3: Compare against the 30% threshold. If the improvement is 30% or higher, the AI may have a real edge. If it's below, I dig deeper into other factors like cost reduction, latency, or user experience that might compensate.

Step 4: Contextualize. The rule isn't a binary pass/fail. In high‑stakes fields like autonomous driving, even a 10% reduction in accidents is massive. But for general use cases, 30% remains a solid bar.

Let me give you a concrete example. I evaluated a startup that claimed their AI could predict machine failure 48 hours in advance, compared to the current 36‑hour warning. That's a 33% improvement ( (48-36)/36 = 33% ). Despite some skepticism, I greenlit the investment because it crossed the 30% line. They later got acquired.

Common Mistakes When Using the 30% Rule

I've seen investors and founders make three recurring errors:

Mistake 1: Focusing only on accuracy. An AI might improve accuracy by 30%, but if it also increases latency by 2 seconds, users hate it. The 30% rule should consider the net value, not just one metric. I once passed on a startup that had a 32% better F1 score but required 10× the compute—killing its economic viability.

Mistake 2: Using the wrong baseline. A founder might compare their model to a random baseline (e.g., 50% accuracy) instead of the actual deployed solution. Always challenge the baseline. If the current system already achieves 85%, a new model hitting 92% is only 8% relative improvement—not enough.

Mistake 3: Ignoring the cost of adoption. Even a 30% improvement won't sell if integrating the AI costs more than the savings. I learned this the hard way after investing in a company that improved warehouse picking accuracy by 35%, but the hardware retrofit took 18 months and wiped out the gains.

Case Study: A Real AI Investment Decision

About three years ago, I personally visited a medical imaging startup in Boston. Their AI detected lung nodules with 92% sensitivity, versus the average radiologist's 88%. Sounds great, right? But let's run the numbers: (92-88)/88 = 4.5% relative improvement. Way below 30%. The team argued that even small improvements save lives, but I asked the hard question: would hospitals pay a premium for a 4.5% gain? Their own pilot showed that radiologists ignored the AI because it didn't change their workflow meaningfully. I declined.

Fast forward two years, the startup pivoted to a different approach and eventually shut down. The 30% rule saved me from a bad bet. On the flip side, I later backed a company that improved fraud detection accuracy from 80% to 96%—a 20% absolute jump, but relative improvement of (96-80)/80 = 20%. Wait, that's still below 30%! But here's the nuance: they also reduced false positives by 60%, which translated to huge cost savings. So the rule isn't blind; I made an exception because the value wasn't just accuracy. The 30% rule is a filter, not a jail.

Alternatives and Criticisms of the 30% Rule

No rule is perfect. Critics say 30% is arbitrary—and they're partly right. In some verticals, like AI for drug discovery, a 10% improvement in hit rate can mean billions in revenue. The rule also ignores the potential for future improvements: an early‑stage model might start at 15% improvement but have a clear path to 50%.

Other frameworks exist: the "10x rule" (the AI should be 10× better or cheaper), or the "minimum viable improvement" concept. I personally use the 30% rule as a conversation starter, not a final verdict. It helps me ask better questions: Why is this improvement enough? What's the switching cost? If the founder can't articulate a clear reason why 20% improvement matters, I walk.

Frequently Asked Questions

Does the 30% rule apply to all AI sectors, like healthcare vs. finance?
Not equally. In heavily regulated industries, switching costs are high, so the bar is often higher than 30%. In consumer apps, even a 15% improvement in personalization can drive engagement. I adjust the threshold based on domain and existing solution stickiness.
How do I calculate performance improvement when the metric isn't a percentage?
Convert to a percentage where possible. For example, if the AI reduces error rate from 10% to 7%, that's a 30% relative reduction. But if the metric is latency in milliseconds, compute the percentage improvement in speed. Be consistent with relative change.
What if the AI reduces cost instead of improving accuracy?
Great point. The rule still applies: compare the cost savings as a percentage of the current cost. If the AI cuts customer support costs by 25%, that's under 30%—but if the savings are recurring and scalable, I might still invest. The rule is a heuristic, not a law.
Can the 30% rule be used for non‑investment decisions, like choosing an AI tool for my business?
Absolutely. If you're evaluating a software vendor, ask for their performance relative to your current solution. If the improvement is under 30%, question whether the hassle of migration is worth it. I've helped dozens of companies avoid bad purchases this way.
Is there any research supporting the 30% threshold?
Several academic papers on technology adoption show that a 30-40% improvement is often the tipping point for users to switch. I'd recommend looking at the work of Everett Rogers on diffusion of innovations, and specific studies from McKinsey on AI ROI. The 30% number comes from practical experience more than a single study.