a16zDecagon’s Playbook for Building Enterprise AI Applications
CHAPTERS
- 0:00 – 1:54
Decagon’s vision: AI as the “front door” for every customer interaction
The conversation opens with Decagon’s north star: AI agents should handle both reactive support and proactive customer engagement. The hosts set up the core themes—enterprise deployment reality, open vs. closed models, and what enduring moats look like as models improve.
- •AI agents as the primary interface between a business and its customers
- •Enterprise AI adoption framed as more than “thin UI wrappers”
- •Tension between frontier labs’ progress and application-company moats
- 1:54 – 4:59
From frontier APIs to 90% open source: latency, control, and fine-tuning reality
Jesse walks through Decagon’s progression from using OpenAI/Anthropic APIs to primarily running open-source models. The inflection point was enterprise scale and voice latency requirements—plus the need to control and adapt smaller models through fine-tuning.
- •Early-stage: ship fast with frontier APIs; later-stage: optimize for production constraints
- •Latency (especially for voice) drives the move to smaller models
- •Fine-tuning is required to make small models strong enough for specific tasks
- •Today’s stack: ~90% open source, ~10% frontier for new/experimental work
- 4:59 – 6:57
The “false trade-off”: small fine-tuned models can beat big models on specific tasks
Ashwin argues it’s not simply ‘smarter but expensive’ versus ‘dumber but cheap.’ With task decomposition and strong fine-tuning, smaller models can deliver better task performance while also being cheaper and faster.
- •General-purpose intelligence isn’t required for many agent sub-tasks
- •Fine-tuned smaller models can outperform SOTA large models on narrow objectives
- •You can sometimes win on all three dimensions: performance, cost, and latency
- •Frontier models still matter for broad, exploratory tasks
- 6:57 – 12:26
Decagon Labs as a “model factory”: continuous training, evals, and build-vs-buy decisions
They explain why Decagon built an internal research/labs function: model capabilities and open-source baselines shift constantly. The key bottleneck is not training per se, but assembling data and—especially—building end-to-end evals tied to customer outcomes.
- •Enterprise adoption of post-training will happen, but slower than expected
- •Evals must be bespoke and outcome-driven, not generic benchmarks
- •Rapid model churn requires constant retraining and deprecating old models
- •Framework: build tightly coupled tooling in-house; buy commoditized components (e.g., labeling)
- 12:26 – 15:07
Tokenomics in practice: why Decagon optimizes for accuracy and latency first
The group reframes the token-cost debate: Decagon focuses on quality and speed because customers buy outcomes, not token efficiency. As a growth-stage company, they’re willing to spend more tokens per conversation to raise reliability and UX.
- •Customers care about conversation quality, not internal token costs
- •Tokens per conversation can rise as systems add checks and parallel model calls
- •Cost pressure is highest for companies still running mostly on frontier APIs
- •Once you can decompose problems and deploy open-source quickly, cost becomes less central
- 15:07 – 21:21
Are frontier labs the last startups? Why apps still matter in an AGI world
Sarah challenges the ‘labs will take everything’ narrative. Jesse and Ashwin argue enterprise-ready AI requires substantial software around models—business logic capture, integrations, governance, QA, monitoring—and that apps and labs will increasingly converge rather than replace each other.
- •Most fine-tuning is for Decagon’s use case (customer ops), not per-customer customization
- •Enterprise value is in platform + business-process encoding, not just model access
- •Labs are building apps; app companies are building models—convergence is inevitable
- •Even with AGI, software systems of record (databases/CRMs) remain necessary
- 21:21 – 28:14
The forward-deployed trap: avoiding “consulting with APIs” by productizing workflows
They distinguish necessary early forward-deployed work (to learn new workflows) from an unhealthy dependency that prevents scale. The Decagon philosophy: FDEs should ‘eat pain and excrete product’—customer learnings must become reusable core product, not one-off implementations.
- •AI workflows are new, so early embedding with customers is unavoidable
- •Long-term goal: productize learned workflows to regain SaaS-like scaling
- •If you can’t productize, you risk becoming a modern consulting firm
- •Decagon positions itself as product-led, with sales informing product direction
- 28:14 – 34:45
Duet: the agent that builds the agent (AOPs, tools, tests, and QA automation)
Jesse describes the “oh wow” product leap: Duet, a slower ‘meta-agent’ that automates the heavy internal work of deploying and improving customer-facing agents. Duet can generate procedures (AOPs), write integrations and tests, monitor production conversations, and propose improvements.
- •Duet automates authoring AOPs (Agent Operating Procedures) and tool/integration setup
- •Generates test suites and simulations to validate behavior
- •Monitors large volumes of real conversations and flags weak topics
- •Became possible as reasoning models improved; uses frontier models where needed
- 34:45 – 37:02
Decagon’s moat under AGI: enterprise deployability, controls, collaboration, and compliance
They respond to the ‘what’s your right to exist?’ question by grounding moats in enterprise reality. Even with very capable models, enterprises need robust permissioning, testing, auditability, multi-stakeholder collaboration, legacy-system integration, and insight extraction at scale.
- •Enterprises can’t simply ‘give the model access to everything’
- •Need guardrails: what the agent can/can’t do and how to prevent catastrophic actions
- •Governance and compliance: testing, regulatory constraints, monitoring
- •Operational software around models is durable value—at least for the next few years
- 37:02 – 44:44
Winning enterprise deals: the “glass box” approach vs. Sierra’s “black box”
Sarah frames the market as a two-horse race; Jesse explains why enterprises choose Decagon. The key differentiator is transparency and self-serve iteration: customers can understand, control, and rapidly build new journeys without going through vendor FDEs for everything.
- •Customer story: switched off Sierra due to dependence on FDEs and slow iteration
- •Decagon’s ‘glass box’ enables customer teams (including non-technical) to iterate fast
- •Example outcome: ~7 journeys built in a month vs. ~3 in a year previously
- •Enterprises vary: some want ‘done-for-you,’ others demand control and insight
- 44:44 – 47:31
Founder-led sales and deployment playbooks: how big contracts close unusually fast
They unpack Decagon’s go-to-market intensity and why sales cycles compress in this category. Founders stay heavily involved, and the company productizes the enterprise deployment journey—model risk, testing, rollout, and issue remediation—so buyers can visualize a path to going live.
- •Jesse spends ~80% of time on sales to push speed and unblock deployment
- •Decagon emphasizes empathy for enterprise fears: governance, security, org navigation
- •Tactic: start with 1–2 high-impact use cases to secure early wins
- •A key ‘product’ is a repeatable process from first meeting to scaled go-live
- 47:31 – 51:57
From customer support to AI concierge: expanding to sales and operational workflows
Decagon’s scope broadens as models improve: from support to inbound sales, proactive outreach, and operations. Ashwin emphasizes they built an agent that follows business process well, and improved instruction-following enables more open-ended conversations like discovery and qualification.
- •Customer pull drives expansion: ‘you already know our product/brand—can you do inbound sales?’
- •Operational workflows: proactive outreach based on account signals/issues
- •Core abstraction: business-process execution, not a single ‘support bot’
- •Model improvements: better instruction-following enables flexible, natural conversations
- 51:57 – 55:06
Limits and bottlenecks: hiring, taste, and what agents still can’t decide
Asked about technical bottlenecks, they say the biggest constraint is talent and execution capacity, not tokens. Agents can assist with steps, but aren’t yet trusted to decide what to build or to judge ‘done-ness’ with product taste.
- •Primary constraint: hiring enough AI/product talent to build fast
- •Agents help with execution but not with prioritization/taste decisions (yet)
- •AI tooling increases ambition: companies build more rather than hiring less
- •They’re watching frontiers like voice-to-voice models and smarter small models
- 55:06 – 1:06:57
Culture, scaling offices, and founder communication as the company grows globally
They discuss ‘grind slop’ discourse and describe a team-sport culture where functions blur (engineers on sales calls, sales debugging product). As they expand internationally, they seed new offices with SF culture via onboarding immersion and rotating experienced staff.
- •In-office culture is framed as maximizing communication, not mandated hours
- •Cross-functional collaboration is core to how Decagon ships and closes deals
- •Scaling challenge: culture must be re-instituted as headcount grows
- •New hires spend weeks in SF; hubs seed new offices until local culture stabilizes
- 1:06:57 – 1:14:45
Founder brand and distribution: using X vs. LinkedIn, AI-assisted thinking, and narrative shaping
Jesse explains why X matters less for direct selling and more for influencing the shared ‘timeline’ of what people discuss, with downstream effects (media, hiring, executive awareness). They also describe using AI for brainstorming and for capturing business context, while keeping writing largely human-authored.
- •LinkedIn: enterprise-facing announcements; X: idea placement on the shared discourse timeline
- •Indirect distribution: influencers/media amplify to executives who may not scroll X
- •Ashwin’s personal agents: compile business context for better decisions
- •Modern models are more useful when prompted to disagree and be direct
- 1:14:45 – 1:20:15
Will AI kill jobs? Jevons Paradox in customer support and the shift from cost to growth
They argue automation doesn’t automatically mean mass layoffs because support demand is often constrained by cost and staffing. Cheaper, higher-quality support increases usage (Jevons Paradox), enabling better retention and customer experience—and shifts humans toward higher-leverage, less repetitive work.
- •Many companies have more support demand than supply; AI unlocks latent demand
- •Example: support made prominent across pages; even free users get immediate help
- •Framing: AI may kill specific jobs/tasks, but not necessarily careers
- •As systems mature, revenue-generating conversational workflows become the next frontier
