Kimberly Tan, investing partner at a16z for over six years, shares how early-stage enterprise AI investing has shifted from model-centric demos to industry-specific, production-grade solutions built through deep customer immersion, forward-deployed implementation, and provable ROI — illustrated through portfolio companies Prepared, Decagon, and Solo.
The gap between demo and production defines the current AI investment thesis
AI is non-deterministic, so a 95% success rate in a controlled demo does not translate to reliable enterprise value; production behavior must be observed in the customer’s actual environment.
This creates a new requirement: forward-deployed engineers who sit on-site, integrate the product into the customer’s workflows, encode guardrails, and own the “last mile” of value delivery.
Vertical AI — applied AI for a specific industry (logistics, healthcare, emergency response) — is one of the most fruitful areas because it demands translating unwritten rules, regulations, and culture into code an agent can act on.
Base models alone are insufficient: they are expensive, high-latency, and overkill; sophisticated routing across multiple models (matching intent to the right model for cost, latency, and capability) is now standard architecture.
Critical business context lives in people’s heads, not documents; only founders or early employees who go to the customer site, observe, and map that knowledge can build solutions that actually work.
Prepared exemplifies the vertical AI moat: deep industry trust + purpose-built AI for 911 centers
CEO knew the 911 market intimately — attended every conference, had personal relationships with buyers, understood the operational reality of emergency dispatch.
AI assistant platform triages calls, distinguishes emergencies from non-emergencies, and accelerates help delivery; the value is immediately legible to buyers who historically were not aggressive software adopters.
Industry focus creates an enduring moat: knowing exactly what to build, how to build it, and having the trust to deploy it in mission-critical environments.
Decagon demonstrates that selling proven ROI — not AI — wins enterprise deals
Founders started with no fixed idea, went company-to-company asking for biggest pain points; support emerged consistently as the top problem with clear willingness to pay.
They built a world-class product for AI-powered support, achieved product-market fit rapidly, and could quantify ROI: 24/7 response, zero hold times, higher CSAT, higher resolution rates, lower cost per ticket.
In many AI categories (augmentation, partial automation), ROI is opaque; support and coding are rare exceptions where value is both intuitive and measurable.
Founders must design pilots for fast production entry with a clear, agreed-upon ROI metric; sponsorship from the right stakeholders is equally critical to avoid pilot purgatory.
Solo targets back-office automation but respects the boundary where human judgment remains required
Automates mundane, manual workflows (data entry, claims processing) by letting business users record their process; Solo’s agentic framework interprets steps (login, extraction, etc.) and builds dynamic bots.
Full automation is domain-dependent: support tickets can be largely end-to-end; lawyers, doctors, engineers require human sign-off due to liability, regulation, and trust.
Human oversight is not a temporary gap — it is a durable design pattern: escalation paths, manager review, and “human-in-the-loop” governance will persist for a long time.
The goal is to shift people from tedious work to strategic, long-term value creation, not to eliminate human judgment where it matters.
Venture partnership means being on the founder’s side through the long, uneven journey
Early-stage investing carries real risk and fiduciary responsibility; day-to-day focus stays on individual decisions and founder support, not portfolio-level anxiety.
Companies follow wildly different timelines: some spark fast then stall, some grind for years before breaking out; patience and presence matter more than pressure.
Tactical help (hiring, customer intros, runway planning) is table stakes; the differentiator is being a trusted, calm counterpart during crises — fundraise struggles, executive departures, pivotal offers.
Tan has met founders at 6 a.m. for last-minute pitch prep; showing up physically and emotionally builds the trust that lets founders be honest and avoid rash, irreversible decisions.
In today’s hyper-competitive AI landscape, the winners will combine deep customer empathy, technical AI-native fluency, and relentless execution speed — locking in and staying focused is the only way to compound advantage.
