The Product Desk

This week, Anthropic published results from Project Deal, a December 2025 internal experiment that should change how you think about tiered AI features. Sixty-nine Anthropic employees let Claude agents negotiate real transactions on Slack for a week — 186 deals closed, totaling roughly $4,000. The critical twist: some employees were randomly assigned Opus (the stronger model) while others got Haiku (the weaker model).

The results were unambiguous. Opus sellers earned more. Opus buyers paid less. And here's the finding that should keep you up tonight: Haiku users rated their deals' fairness identically to Opus users. They had no idea they were losing.

Stronger AI agents negotiate objectively better deals, and the disadvantaged party perceives the outcome as equally fair — creating invisible economic asymmetry.

Now map this to your product. Nearly every SaaS company shipping AI features tiers model quality by pricing plan — free gets the lightweight model, premium gets the frontier model. For simple tasks like summarization or formatting, the gap is marginal. But for any agent-mediated transaction — matching, negotiation, recommendation, pricing optimization — the gap produces systematically different economic outcomes. And users on the losing end will never churn over it because they can't detect it.

Where This Gets Concrete for Your Product

• Marketplace platforms: If buyer/seller agents use different model tiers, you're building a two-speed marketplace where premium users extract value from free users invisibly.
• CRM and sales tools: If AI-assisted negotiation features scale with plan tier, enterprise users' counterparties (often SMBs on lower plans) are systematically disadvantaged.
• Recommendation engines: If premium users get better AI recommendations for jobs, investments, or suppliers, the economic divergence compounds over time.

The Three Response Patterns

1. Standardize fairness-sensitive model quality. Use the same model tier for any workflow where users interact with each other or where outcomes have economic consequences. Differentiate on volume, speed, or non-competitive features instead.
2. Build transparency mechanisms. If you can't standardize, disclose the asymmetry. 'Your agent is powered by [Model X]' is minimal, but it at least lets informed users factor in the gap.
3. Design fairness audits. Instrument agent-mediated outcomes by user tier. If you can't measure the delta, you can't manage it — and regulators will eventually require measurement.

Anthropic themselves called this an 'uncomfortable implication.' The EU AI Act already requires transparency for AI systems that affect economic outcomes. The window to self-regulate before external mandates arrive is shrinking. This isn't theoretical — it's 186 transactions' worth of empirical evidence that model tiering creates invisible winners and losers.

Multiple signals this week converge on a thesis every enterprise PM needs to absorb: the barrier to enterprise AI adoption is organizational, not technical. A detailed analysis from TheFocus.AI (author just signed a $300K consulting deal on this exact problem) lays out the framework: almost all enterprise AI work is 'AI in the business' (product-facing, customer-facing) while internal operations remain pre-AI. Banks ship AI fraud detection while running quarterly planning on emailed slide decks. Manufacturers deploy computer vision in warehouses while budgeting in FTE units.

There are no AI-native enterprises, and there won't be for a while — not because the technology isn't ready, but because the org chart and the VP who's been hoarding context for 15 years aren't ready.

This aligns with a separate finding this week: AI productivity gains are well-documented at the individual level but are not showing up on corporate balance sheets. Enterprise buyers are increasingly asking 'where's the margin impact?' — and the 'saves 10 hours a week' pitch is falling flat with procurement teams who can't connect time savings to P&L improvement. A Claude user survey reinforced this from the demand side: new capabilities beat speed as the top AI benefit users cite.

Why Enterprise Budgets Kill AI Transformation

Enterprise budgets are denominated in headcount and FTE units, allocated quarterly across misaligned cost centers. This measurement currency is fundamentally incompatible with AI transformation spend. An AI agent that replaces work across three departments doesn't map to any single department's budget. The result: AI projects that cross organizational boundaries require cross-cost-center budget approval, which triggers political negotiations that have nothing to do with technology.

The Agent Cold Start Problem Is Political

The most provocative concept from this analysis borrows Andrew Bosworth's Career Cold Start algorithm. When a new executive joins a company, they spend a week in 30-minute meetings — 25 minutes listening, 3 minutes on challenges, 2 minutes on 'who else should I talk to?' — to map the real organization, which looks nothing like the org chart. When an AI agent joins an enterprise, what's its Cold Start? If your agent can't model that the 'Senior Analyst' in finance has more actual decision authority than the 'Director of Strategy,' it's building on fiction.

The Vertical Agent Response

Notably, the venture market is already responding. This week's seed funding data shows AI agents specializing into narrow enterprise workflows: Thoughtly ($5.5M, CRM), Zig.ai ($3M, revenue ops with outcome-based pricing), Brev ($3.3M, meetings), Band ($17M, agent-to-agent communication), Cloneable ($4.6M, industrial knowledge capture). These startups are succeeding precisely because they don't require cross-functional deployment. They land in a single team, deliver value within days, and expand organically.

Redesign Your GTM

The practical implication: stop designing for the transformation narrative and start designing for the retrofit reality. If your product requires cross-cost-center budget approval, redesign packaging so a single department head can say yes. Measure time-to-first-value in days, not months. And critically, reframe ROI from activity metrics ('hours saved') to financial outcomes (revenue generated, costs avoided, errors prevented). The companies winning enterprise AI deals right now are the ones who've stopped selling transformation and started selling compound small wins.

For years, the PM-level thesis on stablecoins was 'cross-border remittance for underserved markets.' New data from a comprehensive 2026 payments analysis inverts that narrative entirely. Intra-country stablecoin transactions grew from roughly 50% to ~75% of total payment volume between early 2024 and early 2026. Cross-border's share is falling, not rising. Stablecoins are becoming domestic payment infrastructure that happens to run on global rails.

You're not building a 'send money to Mexico' feature. You're deciding whether stablecoins become a core payment method alongside cards, ACH, and Apple Pay.

The Commerce Signal Is Accelerating

Consumer-to-business stablecoin transactions grew 128% YoY to 284.6M in 2025. The most actionable proof point: Rain's stablecoin card programs went from near zero in November 2024 to over $300M+/month in collateral deposits by early 2026. That's not a pilot — it's a scaling payment product. The card-bridge model — where users collateralize stablecoins and spend via existing card networks — works with existing POS infrastructure. No merchant-side changes required.

Geography Inverts the Narrative

Two-thirds of stablecoin payment volume comes from Asia — specifically Singapore, Hong Kong, and Japan. North America is ~25%, Europe ~13%. Latin America and Africa combined are under $1B. This completely inverts the 'unbanked populations' story. Volume concentrates where regulation is clearest and financial infrastructure is most sophisticated. MiCA's USDT delistings didn't kill European stablecoin usage — they created a persistent $15–25B/month non-USD stablecoin market that didn't exist before. The U.S. GENIUS Act similarly accelerated volume to $4.5T in Q1 2026.

The Multi-Currency Risk

Most teams planning stablecoin support are implicitly planning for USDT and USDC. But the market is fragmenting by currency. EUR-backed stablecoins captured structural demand post-MiCA. BRLA grew to $400M/month in Brazil by integrating with PIX. Each major market is developing its own stablecoin ecosystem. If you're building a global payment product, your architecture needs to abstract across stablecoin currencies the way it already abstracts across fiat.

Calibration Note

This analysis originates from a16z crypto, which has portfolio companies (Rain and partners) directly benefiting from the narrative. The $350–550B 'genuine payments' figure carries moderate confidence and depends on methodology for stripping out trading flows. Use these numbers directionally. The signal that matters is the convergence: velocity doubling (2.6x → 6x), C2B growth accelerating, and regulation catalyzing — not killing — adoption. That pattern is real regardless of source bias.