We thought the problem was AI. We were wrong.
The real bottleneck isn't intelligence—it's ambiguity. That's why we pivoted Bounti from horizontal go-to-market tools to real estate. And what we discovered shattered our assumptions about "horizontal" AI and revealed a blueprint for building AI-first companies.
AI doesn't struggle with thinking. It struggles with vague instructions.
Watch what happens when you plug AI into most business workflows.
It doesn't fail because it's not smart enough. ChatGPT can write compelling copy. Claude can analyze complex datasets. These models are genuinely capable.
But when you ask them to "improve our messaging" or "make this campaign better," they flail. Because no one has clearly defined what "correct" means.
AI systems execute specifications brilliantly. They do not execute loosely defined processes.
If your goal is unclear, your constraints are fuzzy, and "good enough" is subjective, you end up supervising output instead of automating work. That's expensive human time spent on clarification, not creation.
The Difference Between Procedural and Declarative Systems
Most companies run on what I call procedural coordination.
Take customer onboarding. In most companies, it's a comedy of errors. Sales lobs the account over the wall to Customer Success with a breathless, "They're gonna be HUGE!" CS then "reviews the account" (translation: skims the notes while brewing coffee), someone "reaches out to schedule a call" (three emails and a LinkedIn message later...), and finally, you get on a Zoom to "walk through next steps"—which are different for every customer, naturally.
Every step requires human interpretation. What does "review" mean exactly? How long should the call be? Which next steps matter most? It works because experienced people navigate the gray areas.
But it doesn't compound with AI.
A declarative system is different. Instead of describing how work usually happens, you define:
- What must be true at the end
- What constraints cannot be violated
- What "done" actually means
- How to validate success
For onboarding, that might be: "Customer has completed setup checklist, accessed core features, and can demonstrate primary use case within 5 business days."
Now AI can iterate toward that outcome. Test implementations. Flag blockers. Escalate exceptions. The system gets better as models improve.
This is how we operate internally at Bounti now. We define validation criteria upfront. Agents generate implementations. They test against those criteria. They iterate until constraints are satisfied.
The result? Less back-and-forth interpretation. Faster cycles. And as models improve, our throughput compounds automatically.
Why Horizontal AI Felt Like Pushing Rope
Originally, we built horizontal go-to-market workflows. Generate landing pages, outbound campaigns, research briefs, content at scale.
Technically, it worked. But practically, it was exhausting.
We launched a campaign for a SaaS company that was initially a home run. Clean messaging, strong CTAs, solid performance metrics. Then the CEO saw it and declared it "too Millennial." Suddenly, we had to swap out all the emojis for stock photos of people in suits shaking hands. The casual tone became "professional and authoritative." The bright colors became navy and gray.
What's worse, these preference shifts happened constantly. A campaign that tested well last quarter suddenly needed to be "more aggressive" or "less salesy" based on new feedback from leadership.
We found ourselves building around subjective preferences instead of objective constraints. No matter how good the AI got, we were still in the interpretation business.
That's when we realized we needed a domain with more structure.
Real Estate: Beyond the Handshakes Lies AI's Hidden Potential
At first glance, real estate feels relationship-driven. All networking events and coffee meetings and "people buy from people they trust."
And yes, relationships matter. But underneath that surface layer, real estate is actually a series of highly structured problems:
- Pricing under time and risk constraints
- Matching buyers with weighted preferences and budgets
- Evaluating offers across contingencies and timelines
- Coordinating escrow milestones across multiple parties
These are coordination and optimization problems that repeat thousands of times per day across the industry.
The economics matter too. Even after the 2024 NAR settlement shook up commission structures, buyer-agent commissions have held around 2.5% according to Redfin's recent data. When you're working with median home prices, small improvements in efficiency create meaningful economic impact.
And here's the kicker: NAR's 2024 research shows that 76% of buyers found the home they ultimately purchased online first. This means that perfecting the digital representation of a property—the listing description, the photos, the virtual tour—is now a critical, spec-driven process that directly impacts sales velocity.
That combination—economic pressure plus digital dependency—creates a perfect opening for structured AI systems.
Converting Real Estate Into Executable Specifications
Once you stop thinking about real estate as phone calls and start thinking about it as structured constraints, everything shifts.
A seller has: minimum acceptable price, target timeline, risk tolerance for repairs/concessions, flexibility on terms.
A buyer has: budget range with financing certainty, non-negotiable requirements, weighted preferences (location vs. size vs. condition), timeline constraints.
A transaction has: disclosure requirements, inspection windows with defined contingencies, financing deadlines, title and escrow milestones.
Once you express these as specifications, workflows become executable: pricing becomes simulation under constraints; offer evaluation becomes utility comparison across weighted criteria; escrow becomes state transitions with automated validation.
This is the difference between automating steps and restructuring the system.
Why We Started With Marketing and Staging
We entered through virtual staging and content generation for a strategic reason.
It's one of the few areas in real estate that already behaves like a clean input-output system: Inputs: Property photos, walkthrough video, basic attributes. Outputs: Listing descriptions, social content, staged visuals, marketing video.
But more importantly, we could define concrete constraints: spatial consistency, style coherence, lighting realism, transformation integrity. The system doesn't just generate—it converges toward specifications.
NAR's staging research consistently shows that staged homes sell faster and for higher prices. But traditional staging costs $2,000-5,000 per property and takes weeks to coordinate. Virtual staging that actually works—maintaining spatial accuracy while creating compelling visuals—solves a real economic problem.
Staging wasn't just a service; it was our Trojan Horse for introducing declarative workflows into an industry drowning in subjectivity. The same specification-based approach that makes our staging spatially accurate and style-consistent extends seamlessly to documents, transaction workflows, and client coordination.
Why This Creates Unfair Leverage
Most incumbents will add AI features to existing systems. Copilots for their current workflows. AI assistants for their procedural processes.
That helps. Productivity goes up. But if your underlying coordination model is still built around human interpretation and manual handoffs, AI gives you linear gains.
If your system is built around structured specifications from day one, AI compounds. As models improve, your execution speed improves automatically—without rebuilding core workflows.
In a commission-based industry under economic pressure, structural speed advantages matter enormously. Imagine a brokerage that cuts transaction time by 15% and reduces marketing costs by 20% using AI-powered specifications. That's hundreds of thousands of dollars saved per year, per agent—a game-changing advantage in a tight market.
The companies that will win the next decade aren't the ones adding AI features. They're the ones rebuilding their entire coordination model around what AI actually excels at: executing specifications at scale.
Building the Declarative Future
We didn't pivot to real estate because it was trendy or because the market was hot.
We pivoted because it's one of the few industries where declarative architecture actually works today. Repetitive enough problems. Structured enough constraints. Large enough economics to justify the rebuild.
But the bigger insight extends beyond real estate. We're in the middle of a fundamental shift from procedural coordination to declarative execution. The companies that recognize this early and rebuild accordingly will compound advantages as AI capabilities improve.
The companies that bolt AI onto existing procedural workflows will get incremental benefits and plateau.
At Bounti, we're betting everything on being on the right side of that transition. Building systems that get exponentially better as models improve, rather than systems that just get marginally more efficient.
If you're building with AI and finding yourself constantly clarifying requirements and supervising output, you might be solving the wrong problem. The question isn't whether your AI is smart enough.
The question is whether your system is structured enough.