The Product Desk

The most important enterprise AI pricing signal of 2026 dropped this week: Microsoft's M365 E7 at $99/user/month, launching May 2026, bundles Office 365, Copilot, Teams, Outlook, security software, and the new Agent 365 governance platform into a single SKU. This is the first new enterprise tier in a decade. But the strategy behind it matters more than the product itself.

Satya Nadella publicly admitted in January that only 15 million people pay for 365 Copilot — roughly 3% of the total Office 365 base. Microsoft's response isn't to make Copilot better. It's to eliminate the adoption question entirely.

EVP Rajesh Jha confirmed to UBS that while seat growth remains 'healthy,' the bigger growth driver is now ARPU expansion through E5 upsells and Copilot add-ons. Microsoft is explicitly choosing revenue-per-user over user-count as its primary lever. The E7 pricing is aggressive: it costs more than buying E5 plus Copilot add-on separately, offering only a ~$15 discount vs. full a la carte. Microsoft is betting enterprises will pay a premium for consolidated procurement. Adam Mansfield of UpperEdge, who negotiates Microsoft deals for large enterprises, was blunt: 'Bundles can be problematic because, by design, you might be buying employees tools they don't need.'

The Low-Adoption Problem Microsoft Is Pricing Around

Low Copilot usage among employees has been a 'key point of concern' for Jha. Flat per-seat subscription pricing deliberately insulates Microsoft from this adoption problem — revenue stays consistent whether employees use Copilot daily or never touch it. This is financially smart but strategically dangerous. Meanwhile, a separate Atlassian survey of 500+ IT professionals found 98% of organizations use AI in service workflows (up 10pts YoY) but still cannot measure ROI — and bills are growing. Only 7% of enterprises have AI-ready data; 73% struggle with data preparation.

These two data points converge into a single message: enterprise AI adoption is near-universal in theory and shallow in practice. Microsoft's bundle is the logical corporate response — make the revenue consistent regardless of actual usage. But for your product, this creates a specific competitive threat: your enterprise champion now has to justify paying for your AI features and an E7 license that includes similar capabilities bundled for free.

What Survives the Bundle

Microsoft chose Anthropic's Claude — not OpenAI — to power Copilot Cowork's autonomous task execution. This confirms enterprise AI is going model-agnostic; the differentiation has permanently moved up the stack. Three categories survive the E7 gravity well:

1. Domain-specific depth: A generic Copilot can't understand your customer's compliance rules, proprietary data model, or industry workflow. Bessemer's portfolio validates this — EvenUp (injury law) and Abridge (clinical notes) don't compete with Copilot because they embed workflow expertise Microsoft can't replicate.
2. Cross-platform orchestration: Copilot Cowork is cloud-native and M365-exclusive. Workflows spanning M365 + Google Workspace + Slack + vertical tools remain open territory.
3. Measurable ROI: When 98% adopt AI but can't measure returns, the products that surface quantifiable outcomes (time saved, errors prevented, revenue gained) in-product will win renewal conversations. If 'AI ROI measurement' isn't on your feature roadmap, it should be — unmeasurable tools get cut first when CIOs consolidate vendors to fund AI.

Bessemer's Byron Deeter adds a critical nuance: legacy SaaS stock prices have cratered on AI disruption fears, but no major enterprise has actually cut SaaS vendor seats yet. The market has priced in disruption that hasn't materialized. This sentiment-reality gap is your strategic window — enterprises are psychologically ready to explore alternatives but haven't pulled the trigger.

This week marks the moment AI agents crossed from 'impressive demo' to 'production system with unit economics.' Three data points make this case unambiguously:

LangChain's internal GTM agent achieved 250% higher conversion rates and 86% weekly active usage across the sales team — the first credible dogfooding case with both outcome AND adoption metrics.

That 86% WAU is the number that should change your thinking. Most AI feature launches see initial enthusiasm followed by steep dropoff. LangChain cracked stickiness through two design decisions: per-rep memory systems and Slack-native delivery (meeting reps where they already work). These are replicable patterns, not magic.

AI Code Review: A Category Born in One Week

Three vendors launched competing products simultaneously:

This is the first productized agent category to hit simultaneous multi-vendor launch. The speed of commoditization here is a leading indicator for other agent categories — expect this pattern to repeat in customer support, data analysis, and content workflows within 6 months. If you had 'build AI code review' on your roadmap, that initiative just became a buy decision.

Karpathy's Autoresearch: AI Self-Improvement Is Accessible Now

Andrej Karpathy's autoresearch loop — just 630 lines of open-source code — ran ~700 autonomous experiments in 2 days on 8xH100s, found ~20 additive improvements a world-class ML researcher missed, and improved LLM training speed by 11%. Shopify CEO Tobi Lütke adapted it overnight for a 19% validation improvement, with the agent-tuned smaller model outperforming a manually configured larger one. OpenAI's chief scientist Szymon Pachocki targets an 'Automated AI Research Intern' by September 2026.

The agent infrastructure stack crystallized this week to support these use cases: Vercel shipped sandboxed browsers, Terminal Use (YC W26) provides sandboxed compute, VS Code Agent Kanban offers persistent task memory, Slash gives agents credit cards with human-in-the-loop approval, and Paperclip enables multi-agent orchestration with org charts. The engineering barriers to building agent features just collapsed — your competitive advantage shifts from 'can we build it' to 'do we understand which agent features users actually need.'

The Pricing Signal You Can't Ignore

Anthropic charging $15-25 per code review — not per seat, not per month, but per deliverable — establishes an important market anchor. This is value-based pricing that aligns cost with output. The $15-25 range represents roughly 10-15 minutes of a senior engineer's time, making ROI positive if quality approaches human-level. For any PM building agent features, this per-task model deserves serious study as an alternative to subscription pricing.

A structural shift in how users find products just crossed an inflection point, and most PMs are ignoring it. AI assistants now account for 56% of global search engine volume at 45 billion monthly sessions, with 83% on mobile and ChatGPT commanding 89% of sessions. Combined search + AI usage grew 26% since 2023, meaning AI isn't replacing Google — it's building a parallel discovery layer on top.

You now need a dual-channel discovery strategy. Traditional SEO still works for the 44% of conventional search. But for the 56% that's AI-mediated, your product needs to be citable, recommendable, and accessible to AI assistants.

Google Is Eating Its Own Results

Google's AI Mode self-citations jumped from 5.7% to 17.42% in under a year — more than the next six domains combined. Across 1.32 million citations and 68,000 keywords, Google leads in 19 of 20 niches analyzed. Travel (53% self-citation) and Entertainment (49%) are hardest hit. 59% of self-citations point back to Google search results and 36.1% to Google Business Profiles — Google is literally citing its own products as the authoritative source.

The e-commerce picture is even starker. Google product grids now appear on 96% of SERPs, grew 82% in 9 months, can hit 58% CTR themselves, and cut organic result CTR in half. Critical nuance: organic rankings and product grid visibility operate independently. You can rank #1 organically and have zero product grid presence. These are two separate growth levers with different inputs and increasingly different ROI curves.

The LLM Discovery Channel Is Open Now

Framer published a comparison page against Claude Code that's already appearing in AI-generated answers for product selection queries. LLMs heavily favor comparison and 'versus' content formats when answering high-intent queries. This is 2026's equivalent of building SEO landing pages in 2015 — cheap to produce, high leverage, and first movers define the narrative. The window is especially valuable because LLM training data has inherent lag: comparison content that exists today shapes AI-generated recommendations for months.

The Consumer Trust Paradox

Two contradictory data points constrain how aggressively you can lean into AI positioning: 46% of consumers say AI customer service rarely or never succeeds, and only 26% of Americans view AI positively while 46% view it negatively — yet more than half use AI tools. This adoption-without-approval paradox means your AI features need to deliver outcomes without requiring users to trust or even think about the 'AI' label. Audit every surface where you use the word 'AI' and A/B test whether removing it improves conversion.