What We Learned Building 7 AI Products in 18 Months
Between September 2024 and March 2026, SV Labs built and shipped seven AI products: Voice AI Agents, AI Marketing Engine, AI Outbound, AI Dev Command, WealthTech AI, Communication Intelligence, and Document Intelligence.
Some of these are working well. Some required significant pivots. One nearly failed entirely before we found the right approach.
This is the honest version. Not the investor deck. Not the marketing version.
Why Seven at Once
The answer is rooted in our investor-builder model. Salem Ventures invests in companies that need AI capabilities. Rather than have each portfolio company build or buy AI independently, we centralized development in SV Labs and built products that serve multiple portfolio companies simultaneously.
The logic was sound. The execution was harder than we anticipated.
What We Got Right
Shared infrastructure across all seven products reduced per-product costs dramatically compared to building each independently. Common authentication, monitoring, deployment pipelines, cost management -- the boring stuff that compounds.
Cross-product learning was real. Our work on speech-to-text for Voice AI Agents improved transcription in Communication Intelligence. Document parsing models improved contract analysis in WealthTech AI. Knowledge transfers compounded over time.
Internal customers from day one. We were deploying into our own portfolio companies, so we had real users before we had a finished product. This eliminated the typical startup challenge of building in isolation and discovering the market wants something different.
What We Got Wrong
We underestimated context-switching costs. We tried to have engineers work on two or three products simultaneously. This was a mistake. Context-switching between AI products is especially expensive because each has its own data pipelines, model configurations, and evaluation criteria. We eventually moved to dedicated teams, but the first six months of split attention produced slower progress and more bugs than we should have tolerated.
We shipped too early on some products and too late on others. Voice AI Agents and AI Marketing Engine shipped early and iterated fast -- that worked. Document Intelligence shipped too early with accuracy that was not high enough, and the first users had a poor experience that damaged credibility. AI Dev Command shipped too late -- we spent months perfecting features nobody asked for while delaying the core functionality they needed.
We hired for AI experience over domain expertise. Our first wave prioritized ML talent. Brilliant NLP models that do not integrate into actual work processes are worthless. We corrected by pairing every ML engineer with a domain expert who understood the end user's daily reality.
Product by Product: The Honest Version
Voice AI Agents -- our flagship. Starting with a narrow use case (appointment scheduling) and expanding was the right call. Our first attempt at sales conversations was too aggressive -- the AI tried to handle objections and close deals in the same call. Conversion rates were terrible. We pivoted to "qualify and schedule" and conversion rates tripled.
AI Marketing Engine -- the workhorse. Content production pipeline works brilliantly. Our first version of campaign orchestration tried to automate strategy decisions. The AI made reasonable calls in familiar scenarios but costly mistakes in novel ones. We pulled back to AI-assisted orchestration: the system recommends, a human approves strategy.
AI Outbound -- the grind. Personalization quality is our edge. The system researches prospects and generates messages that are genuinely relevant. Response rates are multiple times the industry average. But our LinkedIn automation violated terms of service at scale. LinkedIn's detection improved faster than our evasion. We pivoted to LinkedIn as a research source (compliant) and redirected outbound to email and phone.
AI Dev Command -- the late bloomer. Code review automation is the killer feature. It catches bugs and security issues that human reviewers miss, particularly in large pull requests. Our initial vision of a full AI pair programmer was too ambitious. The technology is not there yet for complex production code. We scaled back to augmentation and found genuine value.
WealthTech AI -- the specialist. Risk assessment models work well. Automated trading recommendations did not. Models performed well in backtesting but underperformed in live markets during volatility. The gap between backtesting and live performance is larger than we assumed. We now position AI recommendations as inputs to human advisors, not client-facing advice.
Communication Intelligence -- the surprise. Found its killer use case where we did not expect: compliance monitoring for financial services. We built for broad business communication analytics. Financial services companies loved it for monitoring advisor-client communications for compliance violations. Privacy concerns were larger than anticipated -- employees were uncomfortable knowing AI was analyzing their communications. We had to add extensive transparency features that were not in the original design.
Document Intelligence -- the near failure. Shipped the first version with accuracy that was good for general documents but not good enough for financial ones, where a single extraction error has regulatory consequences. Two early clients reported errors that destroyed their confidence. We pulled the product, spent three months improving accuracy substantially, and relaunched. Those three months were expensive.
The lesson: in financial services, "pretty good" accuracy and "reliable" accuracy are not close together. They are the difference between a product nobody trusts and one people rely on.
The Lessons That Stuck
Internal deployment is the best product development. If you can be your own customer, do it.
AI accuracy thresholds are step functions, not curves. Below a certain threshold, a product is useless. Above it, valuable. Finding that threshold early saves months.
The AI model is a small fraction of the work. The rest is integration, UX, and workflow design. Our best products are not the ones with the most sophisticated models. They are the ones that fit most seamlessly into existing workflows.
Hire for the workflow, not the algorithm. The most impactful hires understood the daily work of marketers, sales reps, financial advisors, and developers. They shaped products more than any ML researcher.
The journey continues. But eighteen months in, we are confident in the model: build AI products where you are simultaneously the builder, the customer, and the investor.
Building AI products or considering an investor-builder approach? We share these lessons openly. Reach out at info@salem.ventures.
