The Product Desk

The Market Has Already Priced In the Death of Per-Seat SaaS

The SaaStr.ai Index confirmed a 50.5% market capitalization collapse across top public software companies in six months. This isn't a cyclical correction — it's the market explicitly pricing in a paradigm shift: if AI agents do the work of 3-5 human seats, per-seat pricing becomes a melting ice cube. If your product charges $X per user per month and an AI agent can do what 3 of those users do, your addressable revenue per customer just dropped 66% under your current model.

ServiceNow's Triple Move Sets the New Standard

ServiceNow made three moves simultaneously that every enterprise PM should study. First, they eliminated separate AI licensing entirely, making their whole portfolio AI-native by default. This reframes AI from premium upsell to expected capability — like mobile responsiveness was a decade ago. Second, their Context Engine leverages 85 billion workflow records to feed LLMs with real-time business context — a data moat that validates Karpathy's insight that structured, persistent context is what makes AI useful in enterprise, not impressive in demos. Third, they opened agent deployment to external IDEs like Cursor and Claude Code, making themselves the deployment target for enterprise AI agents.

If AI is still a separate SKU in your product, you're charging extra for something ServiceNow just made free. Every quarter you wait, the gap widens.

a16z's Field Data Reveals How Buyers Actually Decide

a16z's field research upends several assumptions. Enterprise buyers deliberately deploy 2-3 AI tools per use case — not because they can't choose, but as hedging policy. A head of AI at a top financial institution runs redundant tools because performance fluctuates and hallucinations happen. This means you're not fighting for a single winner-take-all deal — you're fighting for the premium slot in a multi-tool portfolio. That premium slot sustains 10-20% price headroom without material churn, and it's won through reliability and onboarding quality, not discounts.

The most dangerous signal: enterprises are converging on a build-vs-buy framework where non-core AI gets purchased and core AI gets built in-house. A B2C logistics company plans to move off third-party AI entirely. Your real competitor in 12-18 months isn't another vendor — it's your customer's engineering team.

OpenAI's Codex Tiers Are Your Monetization Template

OpenAI's $20/$100/$200 Codex pricing (Plus/Pro/Max with 1x/5x/20x usage) demonstrates the emerging pattern: usage-based tiers with non-linear value scaling. Going from 5x to 20x costs only 2x more — meaning heaviest users get volume discounts that lock them in. The gap between $20 and $100 with no $50 tier signals a clean segmentation break between casual and professional use. If you can capture the $30-$70 range, there's an underserved segment.

The Dual Pricing Imperative

The clearest action from a16z's data: offer customers a choice between predictable spend and outcome-based pricing for the same product. Per-outcome pricing makes apples-to-apples vendor comparison structurally harder, shifting the conversation from 'what's your per-seat cost' to 'what results do you deliver per dollar' — a conversation premium products win. Building this capability requires product instrumentation that tracks value at the action level, not the user level — telemetry infrastructure that takes 2-3 quarters to build properly. Start scoping now.

The 100:1 Ratio That Rewrites Your PRD

Enterprise agents now outnumber humans up to 100:1. If that ratio is even directionally correct, it rewrites every assumption in your PRD about user personas, interaction patterns, and interface design. SaaS companies are already rebuilding around APIs, CLIs, and structured outputs, encoding domain expertise into 'skill files' and exposing functionality via MCP servers. Google expanding Skills across Gemini confirms this isn't a niche pattern — it's becoming the expected interface contract.

If your product's primary interface is a GUI that only humans can use, you're building for a shrinking percentage of your total addressable interactions.

Three Infrastructure Layers Shipped This Week

Payments: Visa launched Intelligent Commerce Connect — a protocol-agnostic on-ramp for AI agents to discover products and initiate transactions. This is the Stripe-for-agents moment. When the dominant payments network builds specifically for agent-driven transactions, the 'agent as buyer' use case has crossed from demo to deployment.

Discovery: Agentic Engine Optimization is emerging as a must-have discipline. Misconfigured robots.txt and token-heavy pages directly cause AI agents to hallucinate solutions attributed to your product. The five AEO pillars — discoverability, parsability, token efficiency, capability signaling, and access control — are the information architecture patterns you already know, applied to non-human consumers.

Governance: HushSpec defines an open policy format for AI agent runtime constraints — filesystem access, network egress, tool usage, secrets. This is your 'RBAC moment' for agents. Within 2-3 quarters, enterprise security questionnaires will have an agent containment section. First-movers define the constraints; late arrivals inherit them.

DeepMind's 6-Vector Attack Taxonomy Is Your Threat Model Update

Google DeepMind identified 6 distinct attack genres against AI agents: Content Injection, Semantic Manipulation, Cognitive State, Behavioural Control, Systemic, and Human-in-the-Loop. The 'Systemic' category is alarming for multi-agent architectures: jigsaw attacks split harmful instructions across agents, and fabricated agent identities corrupt collective decision-making. An attacker can embed adversarial instructions in CSS metadata that no human would ever see. These attack surfaces don't exist in traditional software security.

What This Means for Your Product

The strategic question isn't whether to build an agent-accessible interface — it's which features agents would use most and exposing those first. Treat agent-consumability as a product metric alongside time-to-value and NPS. Your documentation, API references, and developer content are now a product surface for agents, and content optimized for human readability may be actively hostile to agent consumption. The companies that get AEO right in the next 6 months will build the same compounding advantage that early SEO adopters built in 2008-2012.

The Platform Bundling Playbook Has Arrived

Anthropic leaked a vibe-coding app builder inside Claude that directly competes with Lovable ($6.6B valuation, raised $330M just four months ago). Simultaneously, Anthropic shipped Claude for Word — embedding directly inside Microsoft Word with native Track Changes integration and formatting fidelity, targeting the exact enterprise pain points where Copilot has underdelivered. This is the Microsoft Office playbook applied to AI: identify the most popular things people do with your platform, then make those things native features.

If you wrap an LLM API with a specialized UI, your supplier is your biggest competitive threat. Elena Verna's warning that 'Big Tech companies are more threatening than rival startups' is now confirmed.

Microsoft Proved Platform Risk Isn't Theoretical

Microsoft deliberately starved Azure customers of GPU capacity to serve M365 Copilot and GitHub Copilot. Amy Hood confirmed that Azure growth would have exceeded 40% had GPUs been allocated externally. Nadella justified this because internal workloads have higher gross margins. Let that sink in: the company you pay for cloud AI compute decided its own products were more valuable than serving you. This is the rational behavior of every hyperscaler when compute is scarce.

Silent Degradation Is the New Vendor Risk

Two data points expose a pattern that should alarm every PM building on LLM APIs:

• Anthropic silently cut Claude Code's cache TTL from 1 hour to 5 minutes on March 6 — no changelog, no announcement. A user discovered it via a GitHub issue. This can 12x your caching costs overnight.
• Leaked analysis of thousands of Claude Code sessions shows Opus 4.6 thinking depth dropped ~67%. Users report lazier code edits, shallower reasoning. Developers are migrating to Codex and GPT-5.4 in real time.

These aren't isolated incidents — they're the predictable behavior of providers managing scarce compute across growing user bases. Anthropic is likely conserving compute and will need to 'pay dearly' for more capacity from hyperscalers, costs that flow downstream to you via API pricing.

The Multi-Provider Imperative Is No Longer Optional

Sources disagree on which provider has the advantage — OpenAI tells investors its compute warchest gives it an edge; Anthropic's 233% revenue growth ($9B→$30B annualized) suggests otherwise; open-source models like MiniMax M2.7 hit 97% instruction compliance on tool-use tasks. But all three directions point to the same PM conclusion: build the abstraction layer now. OpenClaw has already formalized a GPT-5.4 vs Opus 4.6 agentic parity gate in its release pipeline. Your product architecture should be at least that model-portable.

Google Can't Get Its Own Employees to Use AI — Neither Can You

Steve Yegge leaked Google's internal AI adoption data: 20% power users, 60% casual chat tool users, 20% outright refusers. If Google — a company that builds the models, employs thousands of ML engineers, and has AI in its culture DNA — can only get 20% power adoption, stop pretending your customers will be different. This is the adoption ceiling for intent-driven AI features, the kind requiring users to consciously choose AI.

The unlock is the 60% — people who'll use AI if you put it in their path, embedded in workflows requiring zero additional cognitive load. Stanford HAI confirms per-user AI value tripled year-over-year, but it's concentrating in the 20% who know how to extract it. Your job is to democratize that value to the 60%.

Every AI feature in your backlog should be evaluated: 'Does this require the user to be a power user, or does it activate for the casual 60%?'

Building Is Cheap. Carrying Is Not.

The sharpest insight from TLDR Founders is deceptively simple: building features has never been cheaper, but carrying one is just as expensive as ever. AI can ship a feature in 2 days instead of 2 weeks. Stakeholders will push for more because 'it's easy now.' Your job is to resist. Products rarely fail from feature scarcity — they fail from feature bloat nobody had discipline to kill.

IDC confirms: 88% of AI proof-of-concepts never reach production. PE firms have shifted from asking 'what's your AI strategy?' to 'how far along are you in rebuilding the company around AI?' The companies winning (Robinhood doubled alert triage capacity with multi-agent systems) are picking 2-3 high-conviction use cases and investing in the unsexy work: evals, observability, guardrails, cost metering.

Gen Z Just Broke Your Adoption Assumptions

Gallup data shows Gen Z excitement about AI dropped from 36% to 22% (a 39% relative decline), with hopefulness dropping from 27% to 18%. This isn't marginal — it's a demographic you assumed would evangelize your AI features becoming actively skeptical. The implication: 'AI-powered' is no longer a selling point for younger users and may trigger friction. Products that embed AI invisibly (Google auto-complete, not ChatGPT's conversation UI) may have a structural advantage as this cohort enters the workforce in larger numbers.

The Human-Weight Boundary

Cate Huston's experiment using Claude for coaching found AI excels at structured tasks (validation, frameworks, action plans) but fails when users need 'human weight' — someone else's confidence in you. This isn't about AI capability — it's a category distinction. If your AI features touch moments where users need to feel seen or emotionally supported, design explicit handoff points: AI handles the structure so the human moment is higher-value and more focused. The winning UX pattern isn't 'AI does everything' — it's 'AI handles the scaffolding so the human interaction matters more.'