a16zHow AI Agents Will Transform in 2026 (a16z Big Ideas)
CHAPTERS
Big Ideas for 2026: AI interfaces, agent-first design, and voice agents
The host frames three major shifts expected by 2026: how we interact with AI beyond chat prompts, how products will be designed for AI agents as the primary users, and why voice agents are becoming real, deployable “AI employees.” The episode sets up these as near-term, builder-driven insights rather than distant predictions.
- •Three featured themes: post-prompt interfaces, designing for agents, and the rise of voice agents
- •Focus on practical insights from investors working with builders
- •Core thesis: AI moves from tool to intermediary/worker across workflows
The “death of the prompt box”: from asking to observing and acting
Marc Andrusko argues the chat-style prompt box won’t be the main interface for AI apps much longer. Instead, apps will observe user context and proactively propose actions, requiring far less explicit instruction.
- •Prompting becomes secondary as apps infer intent from behavior and context
- •Proactive interventions replace manual querying (AI suggests, user approves)
- •UI shifts from “type a request” to “review and accept actions”
AI as the ultimate employee: agency ladder and human-level competence
Marc frames the ideal AI app as a high-agency employee: it identifies problems, investigates causes, evaluates solutions, implements fixes, and only escalates for approval at the end. This model becomes the benchmark for AI products that truly “do work.”
- •High-agency employee model: diagnose → research → propose options → implement → inform/seek final approval
- •Goal: AI performs at least as competently as a strong human worker
- •Workflow design centers on delegation and oversight, not step-by-step instructions
TAM shift: from software spend to labor spend
The opportunity expands from replacing software subscriptions to replacing or augmenting human labor. Marc highlights labor spend as the much larger prize, reframing AI apps as labor automation/augmentation rather than just better SaaS.
- •Traditional software spend is dwarfed by labor spend
- •AI apps target “work performed,” not just “software purchased”
- •Market size expands dramatically when AI is treated as a worker
Proactive AI in CRM: the AI-native workflow example
Using CRM as a concrete case, Marc describes an agent that continuously manages pipeline actions: surfacing neglected leads, drafting outreach, and mining long-term history for next-best actions. The salesperson shifts from doing the work to approving and steering it.
- •Agent scans pipeline, calendar, emails, notes, and call history continuously
- •Revives dormant leads and drafts context-aware outreach
- •Turns CRM from a database UI into an autonomous deal-advancement engine
Human-in-the-loop vs. full autonomy: power users and trust boundaries
Marc expects most users to keep a final approval step, especially in high-stakes situations. Over time, power users may train agents with deeper context and memory until large portions of work happen with little to no oversight.
- •Default mode: “AI proposes, human approves,” especially for risk/liability
- •Power users invest effort to give agents more context and training
- •Long context windows + persistent memory increase autonomy and trust
Creating for agents, not humans: the shift in product and content priorities
Stephanie Zhang argues that as agents become the interface between people and software/web content, what mattered for human attention will matter less. Optimization shifts away from visual UI and hooks toward making information legible and useful to machines.
- •Agents can read everything; humans skim—content structure priorities change
- •Design focus moves from clicks and visual hierarchy to machine-readable clarity
- •Agent intermediaries reshape both software UX and content strategy
Machine legibility in practice: dashboards to Slack, Salesforce to summaries
Stephanie gives examples where humans no longer navigate complex UIs; agents ingest raw data and deliver synthesized insights where humans already work (e.g., Slack). The “product” becomes agent outputs and explanations, not the underlying interface.
- •AI SREs analyze telemetry and post hypotheses/insights directly to Slack
- •Sales agents summarize CRM data rather than requiring UI navigation
- •Interfaces become less about exploration and more about consumable conclusions
The new SEO: GEO tools and optimizing for agent discovery
As consumers ask ChatGPT-like systems for recommendations, companies increasingly optimize for being surfaced by AI agents. This creates a new tooling category and an open question: what signals agents reward and how to shape content for them.
- •Emergence of tools to improve visibility in AI-generated recommendations (GEO)
- •Organizations experiment to show up for “best X” prompts
- •Uncertainty remains about the ranking factors agents will prioritize
When humans exit the loop—and when they won’t
Stephanie notes some customer support contexts already run autonomously, while higher-liability domains keep humans involved longer. The dividing line is typically risk, complexity, and required accuracy.
- •Autonomous handling is already happening in some customer support use cases
- •Security ops/incident response often retains human oversight due to liability
- •Human involvement decreases as model accuracy and reliability improve
Content explosion risk: zero-cost creation and agent-attention spam
Agent-first consumption could incentivize mass production of low-quality, hyper-targeted content designed to capture agent “attention,” similar to keyword stuffing. Stephanie flags this as a key downside of cheap generation and unclear agent preferences.
- •Cost of content creation trends toward zero, increasing volume dramatically
- •Risk of low-quality content optimized for agents rather than humans
- •Potential replay of SEO spam dynamics in an agent-mediated web
Voice agents take up space: from sci‑fi to enterprise deployments
Olivia Moore argues 2025 marked the breakout of voice agents into scaled enterprise usage, and 2026 will expand platforms that can complete full tasks across modalities. Voice becomes a practical interface for the “AI employee” vision.
- •Voice agents shift from experiments to real enterprise rollouts
- •Platforms expand across channels/modalities to execute end-to-end tasks
- •Voice becomes a front door for delegating work, not just Q&A
Where voice is winning: healthcare, finance/compliance, and recruiting
Olivia highlights verticals with immediate ROI: healthcare staffing constraints, finance where compliance consistency matters, and recruiting where instant interviews reduce friction. These use cases are moving beyond simple scheduling into more sensitive conversations.
- •Healthcare: insurer/pharmacy calls, patient scheduling, follow-ups, even psychiatry intake
- •Finance: voice AI can enforce compliance more reliably than humans and is measurable over time
- •Recruiting: on-demand screening interviews across role types, then handoff to humans
Voice AI realities: latency, “human-ness,” multilingual strengths, and call-center disruption
Model improvements have reduced latency and improved accuracy, sometimes to the point that vendors intentionally slow agents down to feel human. Olivia also discusses the uneven disruption for BPOs/call centers and voice AI’s strong performance with accents and multiple languages.
- •Accuracy/latency gains; some teams add pauses/noise to avoid sounding too robotic
- •Call centers/BPOs face pressure; near-term winners may be those who adopt AI to cut costs
- •Voice AI handles accents/multilingual speech well; ASR quality is now very high
Next frontiers and how to build: government services, consumer wellness, and the voice stack
Olivia expects growth in government call workflows (e.g., 911-adjacent, DMV) and more consumer voice experiences, especially wellness and companionship. She closes with a builder-oriented view: voice is an industry with winners across the stack, and developers should experiment with emerging platforms.
- •Government services as a major next use case (less friction for citizens and workers)
- •Consumer voice AI: wellness tracking and companionship in assisted living/nursing homes
- •Voice is a multi-layer stack; builders should test platforms/models (e.g., ElevenLabs)
