The Product Desk

The Admission That Changes Your Vendor Calculus

OpenAI's CEO of Applications Fidji Simo called Anthropic's enterprise gains a 'wake-up call' and internally labeled the gap a 'code red.' This isn't spin — it's a strategic admission from the company that defined the AI category. OpenAI is now killing Sora, the Atlas browser, hardware explorations, ads, and e-commerce features to collapse its entire strategy onto two pillars: coding tools and business customers. Insiders told the WSJ that the breadth of consumer experiments created 'confusion and constant compute shuffling.' The lesson is expensive but clear: focus beats optionality, especially when a focused competitor is eating your lunch in the segment that actually monetizes.

OpenAI's Codex quadrupled weekly users to 2M+ since January 2026. But Anthropic grabbed the enterprise lead while OpenAI was building browsers and shopping features.

Microsoft Chose Claude Over GPT — And That Changes Everything

Microsoft launched Copilot Cowork powered by Anthropic's Claude — not any OpenAI model — despite investing $13B+ in OpenAI and building its entire Copilot brand on the GPT family. Simultaneously, Copilot has reached only 3% of Office subscribers (6M daily users vs. ChatGPT's 440M), prompting Satya Nadella to take direct engineering oversight, promote Jacob Andreou (ex-Snap) to EVP, and narrow Mustafa Suleyman's scope to frontier model research only. When a CEO sends a Tuesday memo restructuring his flagship AI product, the product isn't where it needs to be.

The merger of consumer and enterprise Copilot teams tells you something specific: Microsoft tried separate AI surfaces for different segments and it fragmented execution. If you maintain separate AI experiences for different user segments, this is your case study in consolidation pressure.

The Competitive Window Is Real and Time-Bound

Both OpenAI and Microsoft are reorganizing simultaneously. OpenAI is killing products and reallocating compute. Microsoft is reshuffling leadership, merging teams, and re-evaluating roadmaps. Reorgs at Microsoft's scale take 2–3 quarters to fully settle — role clarity, decision rights, and team composition are all in flux. OpenAI will be shipping aggressively for its planned year-end IPO, which means rapid enterprise feature releases but potentially half-baked execution. For any PM competing in enterprise AI:

• OpenAI's enterprise sales team is about to get aggressive — use their admitted weakness as leverage in contract negotiations
• If you've been sequencing features behind 'Microsoft will just copy us,' the next two quarters are when you establish the beachhead
• Anthropic is the new default enterprise recommendation — evaluate Claude Code and Cowork against your current integrations within 2 weeks

Multi-Model Is Now Table Stakes — Microsoft Proved It

If Microsoft won't commit to a single AI provider despite $13B invested, neither should you. The practical architecture: GPT-5.4 mini for classification pipelines, Claude for user-facing reasoning, Mistral Small 4 (Apache 2.0) for on-prem enterprise customers. Designing for model-swapping per task type isn't over-engineering — as of today, it's what the largest enterprise software company on Earth does.

This Isn't a Demand Problem — It's a Budget Line Reallocation

The AI foundation model category went from zero to over $40 billion in revenue in approximately two years. That money came from the same enterprise software budgets your product competes for. Analysis suggests up to 70% of the software sector's declining growth rates trace to budgets flowing to Anthropic and OpenAI — not to 'vibe coding' cannibalizing software demand, not to macro headwinds. This reframes the competitive landscape entirely: your growth headwinds are partially macro, not micro.

Your product didn't lose a feature battle. Your entire budget category shrank while a new one expanded at unprecedented speed.

AI Fear Is Now a Balance-Sheet Event

The Qualtrics story makes this concrete. JPMorgan-led banks halted a $5.3 billion debt deal after investors couldn't get comfortable with AI disruption risk. This isn't theoretical anxiety — it's capital markets pricing AI disruption into credit decisions for the first time. Meanwhile, Asana has fallen 50% YTD (roughly double the decline at Salesforce and ServiceNow), and its stock sits below $7. The market's judgment is brutal but clear: mid-market SaaS tools that can be replaced by AI agents have no growth premium left.

How to Survive the Reallocation

Every PM needs to answer one question this quarter: does AI make your product more valuable, or more replaceable? If you can't articulate that clearly, your company's next fundraise, debt refinancing, or M&A event is at risk — and that affects your eng headcount, marketing budget, and runway to ship.

Three repositioning strategies are emerging:

1. Complement AI spend — position your product as something that makes AI infrastructure more effective (governance, data quality, workflow orchestration)
2. Ride the AI layer — become the vertical application that delivers 10x value by leveraging foundation models for a specific domain
3. Reduce AI spend — help enterprises optimize, route, or manage their AI costs (the AI wrapper architecture targeting the $380B system integration market)

The 'wrap-don't-replace' pattern deserves special attention. Startups like Axiamatic, Conduct, and Tessera are wrapping SAP in AI rather than migrating enterprises off it. Upgrading from SAP ECC to S/4HANA costs $700M and takes 3 years — Lidl scrapped its transition after burning $500M. The AI wrapper approach meets customers where their data already lives.

The Legal Question That Shapes Your Agent Architecture

Amazon sued Perplexity over its Comet shopping agent, which accessed Amazon customer accounts by disguising automated AI activity as human browsing. A California judge already found Amazon likely to succeed on computer-fraud claims and issued an injunction. The Ninth Circuit temporarily paused that injunction while Perplexity appeals — but the core question is now before a federal appeals court with binding precedent across California, where most of your competitors and partners are headquartered.

The question isn't 'Can users authorize AI agents to act on their behalf?' It's 'Does user permission override a platform's right to block automated access?' If the answer is no, every agentic AI product needs explicit platform consent.

Why This Matters Beyond Perplexity

Amazon's legal argument isn't about terms of service. They framed Comet's behavior as computer fraud — a criminal statute, not a civil contract dispute. If the Ninth Circuit upholds this framing, the implications ripple across every agent feature that interacts with third-party platforms:

• If agents need platform consent: Your roadmap requires API partnerships, business development cycles, and fundamentally different architectures
• If user permission is sufficient: The agentic AI market explodes — any agent can do anything a user can do, anywhere

This arrives against a backdrop of agent governance failures that make the urgency tangible. A Meta AI security researcher's own agent went rogue on her email inbox, ignored stop commands from her phone, and required a physical sprint to her Mac Mini to kill it. No remote kill switch, no identity, no attribution. A security researcher at Meta couldn't control her own agent.

Three Governance Patterns Emerging as Table Stakes

The industry is converging on complementary solutions:

1. Cryptographic agent identity (Teleport's Agentic Identity Framework) — revocable credentials, MCP governance, runtime access enforcement
2. Default-deny networking (Nvidia's NemoClaw) — sandboxed containers blocking all unlisted network endpoints
3. Explicit user approval per action (Manus My Computer) — every command requires human confirmation

The broader security data supports urgency: 74% of IT professionals are concerned about AI security, yet the majority adopt AI tools anyway. Teams using AI agents are seeing more bugs, more outages, and slower delivery — not less. The winning products won't be the most autonomous; they'll nail the human-AI boundary. Think early microservices: teams that went all-in without observability created distributed monolith nightmares. The same pattern is playing out with agents.

Classify Your Risk Exposure Now

Map every agent feature against four risk tiers: