The Product Desk

The Displacement Evidence Is Now Concrete

Start with what a user actually did this week, not what a vendor pitched. An Atlanta real estate manager, non-technical, opened Replit and Claude Code and rebuilt what Salesforce sold him, saving $100K/year. Five firms confirmed ending Salesforce and HubSpot contracts in the past six months. This is the part the incumbent decks miss: these customers are not swapping vendors with capability. The competitor is not a company you can study on G2. It's the user with AI coding tools open.

PMF is now 'ephemeral.' Companies see bookings spikes when they launch, but within a quarter or two, foundation models catch up. — Brandon Gleklen, Battery Ventures

HubSpot's Reversal Completes the Pattern

HubSpot shipped opt-out customer data pooling for an AI sales leads feature on July 1, hit an immediate revolt, and reversed within 4 days. Watch the sequence, because it repeats: opt-out default, social media discovery, outrage, apology. Zoom ran it in 2023. Slack ran it in 2024. HubSpot ran it in 2026. Shares are down 75% since early 2025, which is the context for the data grab. Here is the strategic trap. Legacy SaaS needs multi-tenant data to make AI features work. B2B customers treat CRM data as competitive intelligence they are paying you to protect. Those two facts do not reconcile with an opt-out toggle.

The Market Is Repricing Around Infrastructure, Not Features

The DataRobot write-down separates the pitch from the thing being done. On paper they were an AI/ML company, positioned for exactly this era. What they actually built was tooling to make ML accessible to non-experts, and foundation models now do that job cheaper. Proximity to AI is not a moat. Selling AI tools is not a moat.

The Viable Data Strategy Playbook

HubSpot's failure leaves three paths for AI features that need multi-tenant data. (1) First-party only: each customer's AI uses only their data. Quality drops, trust holds. (2) Opt-in with explicit value exchange: a clear bargain, default off. (3) Privacy-preserving architectures: federated learning, differential privacy, synthetic data. The first CRM to credibly ship option 3 owns the narrative. Attio is already taking HubSpot churn on cost. Add the trust story and HubSpot's mid-market is genuinely exposed.

The durable positions are narrow. Proprietary data aggregation that compounds with usage. Network effects between users. Compliance certifications AI can't self-issue. Integration ecosystems expensive to replicate. Everything outside that set has a measurable half-life, and the DataRobot number is what the half-life looks like when it runs out.

The Budget Burn Is Real — And Happening to Sophisticated Teams

Uber rolled out Claude Code across engineering in 2026 and burned their entire annual AI budget in four months. This isn't a startup scaling problem. Gartner now projects 40% of AI agent projects will be cancelled by 2027 due to cost overruns alone. The root cause isn't that tokens are expensive — per-token costs dropped 80% from 2023-2026. It's that agentic workflows consume 5-30x more tokens per task, and Stanford found 62% of what gets sent to agents is redundant repeated content (system prompts, tool definitions, documents).

The math: if your agent uses 20x more tokens but 62% are repeated, your effective cost multiplier is still 7.6x even at 80% lower unit prices.

The Reliability Ceiling Compounds the Problem

AutomationBench-AA — the first credible independent agentic benchmark — tested 657 tasks across 40 simulated SaaS applications. Results:

• Fable 5 leads at 48.6% task completion — meaning the best model fails more than half the time
• Every single model still breaks business rules in testing
• Open-weight models top out at 27.8% — a 20.8pp gap that makes self-hosted agents non-viable today
• Gemini 3.5 Flash at 42.6% was notably strong on cost efficiency and guardrail compliance — for high-volume automation, Google may be the better bet

François Chollet's critique is now validated: benchmark scores without cost-per-task are meaningless. Your evaluation framework needs cost-per-successful-task as the primary metric.

The Cost Architecture That Works

Three proven optimizations change the economics materially:

1. Small-model pre-filtering for RAG: A cheap model discarding irrelevant context before your expensive reasoning model achieves 68% context pruning with 96% recall retention — roughly 1/3 the inference cost with negligible quality loss
2. KV cache optimization: LMCache (10K+ GitHub stars, EuroSys 2025 Best Paper) delivers 14x faster time-to-first-token. Alibaba's production data shows 10% of cache blocks serve 77% of hits — prompt structure is a product decision
3. Prompt architecture for cacheability: Anthropic's prompt caching delivers 90% cost reduction on cached tokens, but only if your prompts are structured correctly: stable system prompts first, tool definitions in consistent order, user-specific content last

The Timing Window

OpenAI cut inference costs 50% (compounding to ~90% reduction in 12-15 months). Fable 5 subsidy ends July 8. GPT-5.6 Sol Ultra is imminent. Nvidia is offering neoclouds guaranteed buyback on unused GPU capacity — a demand-softening signal. The next 30 days have maximum pricing uncertainty. The architectural move is model abstraction: teams that can swap providers with a config change will exploit the chaos while locked-in teams eat the margin compression.

Separate the tool from the person

A PM sat through her second annual review this year still telling herself the anxiety on her team is about tooling. Lenny's second annual workforce survey (n=5,920, with 46.9% PMs) says otherwise. The strongest predictor of career optimism, field recommendation, and burnout trajectory is not role, level, comp, or company size. It's how each person relates to AI: amplified (49%), redefined (27.4%), destabilized (13.9%), or diminished (5%).

AI identity stance is the #1 predictor of career optimism (β = +0.39) and field recommendation (β = +0.60). That beats role, level, and company size combined.

The squeeze is the fear, not replacement

Only 22% fear losing their job to AI. Teams tell themselves people fear being replaced. What people actually fear is structural exploitation.

• 51% fear more output expected for same pay
• 46% fear unsustainable pace
• 41% fear declining quality
• 77% hold both positive AND negative emotions at once. Smiling exhaustion.

The team isn't afraid of AI. They're afraid of you using AI gains to ratchet up expectations without sharing the value. A Q3 plan that assumed 'AI makes the team 30% more productive, so add 30% more scope' is the reason retention craters in Q4.

Design and research are the roles under water

Designers report 63% overwhelmed, 61% tired. Researchers report 51% job-security anxious, 36% fear AI replacement. These are the roles where management rates worst and attrition risk runs highest. These are also the roles that determine whether your product is good. Product quality depends on human judgment in design and user insight. It does. So these partners need proactive protection now, not after they leave.

The hiring market compounds it

The hiring market has bifurcated. AI/ML engineers get 2-3 inbound messages daily without looking. Generalist SWEs get zero callbacks from 40 applications. Staff+ and EMs are nearly unhirable even at 90th percentile comp. DuckBill Group receives 1,000 applications/day with only 2 relevant. The roadmap is constrained by what can actually be staffed. The product engineer with taste and judgment is harder to find than someone who can stand up complex agent architectures.

Manager effectiveness is the highest-leverage move

Manager effectiveness is the #1 driver of burnout in the whole dataset. Workers with excellent managers report 65% higher job enjoyment. Only 25.5% rate their manager as highly effective, unchanged year over year. That is the cheapest evidence-backed intervention on the table. Invest in management capability and the AI identity distribution drifts toward amplified on its own.