EO StudioEveryone Builds, Ships, and Sells. Winners Do It Differently. | Kimberly Tan, a16z Investing Partner
CHAPTERS
- 0:00 – 1:01
From “GPT wrapper” to real enterprise value: the hidden work in AI apps
Kimberly Tan reframes the “GPT wrapper” critique by explaining how much engineering and domain translation is required between a foundation model and a usable enterprise product. She emphasizes that critical business knowledge often isn’t written down or machine-ingestible, making customer discovery and mapping workflows essential.
- •“Wrapper” is misleading: production-grade AI requires significant work beyond calling a model
- •Enterprise value depends on aligning model outputs to real workflows, not just capabilities “out of the box”
- •Much domain knowledge lives in employees’ heads and must be elicited directly
- •Founders gain leverage by learning customer processes deeply and building specifically for them
- 1:01 – 2:31
Kimberly’s path into a16z and how she learns: founders as the trend radar
She introduces her background as an early-stage B2B investor at a16z and describes how she ramped quickly despite limited initial tech exposure. Her investing intuition is heavily informed by what founders on the ground report before trends become obvious to the broader market.
- •Joined a16z at 23; focused on early-stage B2B and applied AI
- •Self-driven learning loop: write down unknown terms, study nightly, read venture classics
- •Uses founder conversations as primary signal for emerging shifts
- •Noted early founder conviction (2020–2021) that “AI will be in everything”
- 2:31 – 3:32
Demo vs. production: why AI products are harder than deterministic software
Kimberly explains the core gap between impressive demos and reliable deployment. Because AI is inherently non-deterministic, building for enterprises requires new product muscles, rigorous validation in real settings, and clearer expectations about failure modes.
- •AI behaves probabilistically; enterprise software expectations are deterministic
- •A demo can be engineered; production must handle messy real-world variance
- •Non-determinism raises the bar for testing, monitoring, and user trust
- •Seeing real production performance matters more than polished demos
- 3:32 – 4:02
Forward-deployed AI: integrating context, guardrails, and the “last mile”
She highlights the rise of forward-deployed teams who work onsite to implement AI successfully. This motion ensures the system is integrated into the right tools, grounded in business context, and constrained by explicit do/don’t guardrails.
- •Forward-deploy helps bridge the last-mile gap from model to business value
- •Onsite setup supports integrations, context capture, and workflow alignment
- •Guardrails and business rules are essential for safe, useful outputs
- •Implementation and onboarding are becoming a major differentiator in AI startups
- 4:02 – 4:32
Vertical AI moats: encoding industry rules, culture, and compliance
Kimberly argues vertical/applied AI is especially promising because industries have complex rules that must be translated into code. Being industry-focused becomes an enduring moat: it shapes what to build, how to sell, and how to deliver outcomes in a way horizontal tools can’t easily replicate.
- •Vertical AI wins by encoding regulations, processes, and domain-specific logic
- •Out-of-box model capability isn’t enough for regulated or nuanced industries
- •Industry focus creates leverage: clearer requirements, tighter feedback loops
- •Moat comes from deep domain empathy and repeated workflow-specific learning
- 4:32 – 5:32
Routing and chaining models: cost, latency, and intent-aware orchestration
She explains that enterprise-grade agents typically use multiple models rather than always calling the “best” one. Strong systems route requests based on intent and optimize for tradeoffs like latency and cost, chaining models to reach the required output reliably.
- •Using the most advanced model for everything is often too slow/expensive
- •Intent detection enables routing to the best model per task
- •Production systems often require sophisticated multi-model chaining
- •Optimization across quality, cost, and latency is a key application-layer advantage
- 5:32 – 6:03
Customer-site learning as a moat: extracting tacit knowledge and workflows
Kimberly returns to the idea that prompts and documentation rarely capture what enterprises actually need. The highest-leverage work is founders spending time with customers to uncover tacit knowledge, map the real process, and translate it into product requirements.
- •Good prompting is hard; business context can’t be captured in a single prompt
- •Tacit knowledge must be discovered through deep customer interaction
- •Founders/employees who go onsite learn what models can’t infer
- •Workflow mapping is a core competency for successful enterprise AI
- 6:03 – 7:24
Case study — Prepared: vertical AI for 911 centers and outcome-driven value
She describes Prepared as an archetypal vertical AI success: deep market knowledge, strong relationships, and empathy enabled a purpose-built product for emergency response. The value proposition is intuitive and outcome-based—helping people get assistance faster—creating strong market pull even in a historically slow-buying segment.
- •Prepared serves 911 centers with an AI assistant platform
- •Founder’s industry trust and relationships accelerated adoption
- •AI delivers differentiated outcomes (triage, faster response) customers understand
- •Vertical focus provides a durable moat and clearer product direction
- 7:24 – 8:04
EO Magazine interlude: the show’s stories in written form
A brief channel segment promotes EO Magazine as a written companion to the video content. It highlights features like saving, quoting, sharing lines, and receiving additional deep dives via email.
- •EO Magazine republishes channel stories in writing
- •Readers can save/quote/share notable lines
- •Newsletter offers additional deep dives beyond the videos
- •Call to subscribe at eomag.io
- 8:04 – 9:35
Case study — Decagon: finding the wedge by selling ROI, not “AI”
Kimberly recounts Decagon’s early journey from exploring ideas to landing on customer support automation by directly asking enterprises about their biggest pain points and ROI. The story emphasizes strong commercial instincts, customer-driven problem selection, and rapid attainment of product-market fit once the wedge was clear.
- •Decagon started with a thesis: enterprise wants AI but struggles to realize value
- •Founders did direct discovery: repeated pain-point and ROI questioning
- •Customer support emerged as the highest-ROI, most obvious wedge
- •Clear problem selection + strong execution led to fast PMF
- 9:35 – 11:06
Designing pilots that convert: sponsorship, production readiness, and measurable ROI
She outlines why enterprise AI companies must run pilots engineered to reach production quickly and prove value with explicit metrics. She contrasts categories with intuitive/quantifiable ROI (coding, support) versus partial-automation or augmentation tools where value is harder to communicate.
- •Great tech isn’t enough; adoption depends on explainable, provable value
- •Pilot design should prioritize fast path to production and clear success criteria
- •Secure internal sponsorship to avoid “pilot purgatory”
- •Support automation stands out as a uniquely quantifiable ROI use case today
- 11:06 – 12:06
Back-office automation with Sola: recording processes to build agentic bots
Kimberly discusses why mundane enterprise back-office work is ripe for automation and how Sola approaches it. By letting business users record their processes, Sola captures otherwise tacit context and generates dynamic automation that understands steps like login, extraction, and data entry.
- •Large enterprises have abundant manual back-office workflows (data entry, claims)
- •Sola captures process context by having users record how work is done
- •Agentic framework identifies workflow steps and builds adaptable bots
- •Goal: shift humans from repetitive tasks to higher-leverage strategic work
- 12:06 – 13:37
Where automation should stop: oversight, escalation paths, and regulation
She draws a line between tasks that can be automated end-to-end and professions where liability, trust, and regulation necessitate humans in the loop. Even in “automatable” domains like support, she expects long-term human oversight, escalation, and management of AI agents.
- •Automation feasibility depends on both technical complexity and domain constraints
- •Regulatory/liability dynamics often require human sign-off (e.g., legal, medical)
- •Trust is built via oversight and human-in-the-loop review
- •Even support needs escalation paths and human managers monitoring AI agents
- 13:37 – 17:44
Being “on your side”: venture risk, founder support, and moving fast in AI
Kimberly reflects on the responsibility and uncertainty inherent in early-stage venture, emphasizing steady partnership through both highs and rough patches. She closes with a call for founders to press the gas in a fast-moving AI landscape—winning through empathy, technical excellence, and speed.
- •Venture outcomes are uncertain; investors must accept risk and manage responsibility
- •Founders need both tactical help (hiring, customers, runway) and patient support
- •In crises, avoid rash irreversible decisions; be present and strategic
- •AI is intensely competitive—winners stay focused, move fast, and stay close to customers
