Bezos Wants AI to Approve Building Permits in 10 Seconds. Here's What the Data Says.

Last week, Jeff Bezos made the case for AI in city government with characteristic directness:

“Miami should have an AI application that reads your building permit and it should give you a yes or a no in 10 seconds. Why does it take months and months and months to get a building permit? It doesn't make any sense.”

The clip went viral. 1.8 million views, 11,000 likes. And for anyone who has watched a development project bleed cash while waiting on a permit decision, the frustration is visceral.

He's right that the problem is real and that data can help solve it. But the 10-second yes/no framing — while satisfying — misses what actually breaks project economics.

The real problem isn't slow decisions. It's unpredictable ones.

We've analyzed 1.8 million building permits across seven major U.S. cities. The data reveals something counterintuitive: it's not the median timeline that destroys projects. It's the spread.

In Austin — the fastest market in our data — the median permit takes 22 days. Straightforward. But the 90th percentile takes 179 days. The same type of permit, the same city, an 8× difference in outcome depending on factors a developer can't easily observe in advance: which examiner reviews the application, how loaded the queue is that month, whether a neighboring project has triggered a zoning review.

In NYC, the P50 is 79 days. The P90 is 344 days. A developer underwriting a construction loan at 8% on a $20M project is staring at a difference between $350K and $1.5M in interest reserve requirements — for the same project, in the same city, filed the same way.

A 10-second yes/no doesn't solve that problem. Even if AI could instantly approve every permit, developers would still face the question that actually drives capital allocation decisions: what's my timeline risk before I close on this site?

What AI can actually do — and what we're building

The path to better permit outcomes runs through data before it runs through AI-automated approvals. Cities don't have consistent, machine-readable permit data. Zoning codes aren't standardized across jurisdictions. Most permit offices can't query their own historical approval rates by permit type, examiner, or filing month.

That data problem is what we've been solving at Prevesta. Our database covers:

• 1.8 million building permits across Los Angeles, San Francisco, NYC, Seattle, Austin, Chicago, and Miami
• 1.5 million duration records — from filing date to final issuance, mapped to permit type and district
• District-level congestion data — how loaded each permit office is in a given quarter
• Contractor performance profiles — which GCs and architects have shorter review cycles
• Examiner-level data in NYC — 1 million+ decisions tracked to individual reviewers

What we're building isn't an approval engine. It's a risk intelligence layer that tells developers — before they file, before they close, before they draw down on a construction loan — what their realistic timeline distribution looks like and what's driving it.

The carrying cost math changes everything

Here's what the numbers look like in practice. A developer in San Francisco with a 100-unit multifamily project, $25M in construction financing at 8%:

That $2.3M gap between a median outcome and a tail outcome is the number that determines whether a project gets financed, whether LP returns hold, whether the deal pencils at all. No investor is modeling that spread today — because the data to model it hasn't existed in a usable form.

Bezos is pointing at the right problem

The frustration behind the 10-second-approval idea is legitimate. Permit processes in most U.S. cities are opaque, inconsistent, and paper-dependent in ways that serve nobody — not developers, not cities, not the people waiting for housing.

The deeper fix requires exactly what Bezos is gesturing at: getting permit data into machine-readable form, building systems that can learn from historical decisions, and using that intelligence to reduce variance. That's a multi-year project for any city government.

In the meantime, developers are making $10M+ capital allocation decisions using data that was available in 1995. The permit database exists — it's just siloed in city systems, inconsistently formatted, and almost entirely unanalyzed.

We've spent the last year building the infrastructure to change that. If you're underwriting a development in any of our seven markets and want to know what your actual timeline risk looks like — not the anecdote from your last project, but the distribution from 250,000 comparable permits — that's exactly what Prevesta is built to show you.