Saturday, September 12, 2026

7 signals

AI adoption without documented motion creates 6 private versions of the same process—the tool amplifies inconsistency rather than fixing it. Win rate becomes the average of all private interpretations.
Shared context files (4 core documents with named owners) beat prompt libraries because files age slowly while prompts age badly across model versions. Foundation > seat > model.
New reps should learn your motion from documented files in week one, not from reverse-engineering your calendar. This is the measure of whether AI adoption is team asset or private trick.
The real cost of AI tools is opaque: outbound agents score every row (including rows rules could eliminate), and invoices arrive as 4 opaque lines. Spend grows unchallenged or gets cut on feeling, neither is a decision.
Distributed/European teams cannot rely on corridor conversations to patch motion inconsistency—written files are the only version that survives distance and enables second-market expansion from something other than zero.

AI models (GPT-6 Astra, Fable 5.1) have crossed a capability threshold where they can autonomously handle end-to-end complex tasks including spawning sub-agents, managing context, and self-correcting—moving from 95% solution quality to near-complete task execution
Multi-agent orchestration is now practical: agents can spawn other agents, communicate asynchronously, evaluate codebase agent-friendliness, and perform black-box regression testing without explicit instruction on implementation details
Cost trajectory is exponential: Navier-Stokes solution cost $millions in compute (300B tokens), but o3→Astra cost dropped from $500K to $20 for superior performance, suggesting $50 solutions within 3 years—creating winner-take-all dynamics
Compute scarcity is the binding constraint: OpenAI paused $200 Pro subscriptions due to GPU/CPU shortage despite massive demand, indicating infrastructure bottleneck, not capability limitation
Open science is under threat: Terence Tao warns that AI-powered research teams racing to solve published problems before original researchers finish creates perverse incentives to hoard research directions, potentially reversing centuries of open science tradition

FDE role has been diluted across industry—same title describes fundamentally different jobs (sales engineers, quota-carrying reps, consultants) with different reporting lines and incentives; lack of clarity creates organizational confusion
True FDE function is product extension, not services: the role must both solve last-mile customer problems AND feed insights back to product team to inform generalizable platform improvements; without feedback loop, it's consulting with better branding
Operating model discovery is the core FDE skill: learning customer 'nouns' (how they define entities) and 'verbs' (how those entities move through workflows) reveals undocumented systems that live in spreadsheets and institutional knowledge—this is where real value lives and wher
Low-hanging fruit is exhausted: repeatable SaaS motion solved; remaining value migrates to customization and last-mile problem-solving that no product could anticipate; this structural shift explains why every company suddenly needs FDEs
Palantir's Project Frontline model (250 engineers rotated through FDE roles) created feedback loop that turned field insights into platform features; this rotation model differs from permanent embedded FDE structures and may explain why some FDE programs fail to generate product

AI is not just automating tasks—it's breaking the cultural proxies (legible achievements like puzzle-solving) that fields use to identify and reward expertise, creating a Goodhart's Law scenario where the measurement itself becomes gamed and meaningless
Mathematics distinguishes between 'puzzle-solving' (high-legibility, high-prestige work) and 'idea-generating' (the actual intellectual work); AI solving puzzles without generating new ideas undermines both the motivation system and the validation mechanism for real progress
Software engineering faces identical structural crisis: GitHub projects, rapid coding, and shipping speed were legible proxies for skill that AI now counterfeits; fields must either silo 'human work' from 'AI work' (like chess) or discover new, AI-resistant skill signals
Historical precedent from chess and speedrunning suggests human prestige can survive AI dominance through separate competitive spheres and improved human performance via AI insights, but this requires deliberate cultural reconstruction
The real risk isn't job displacement but motivation collapse—if the traditional paths to prestige become meaningless, talented people may exit fields entirely rather than compete in devalued human-only leagues

The ubiquitous '30% annual data decay' figure is likely a 24-month rate misquoted as annual; actual US rate is 12.25% annually (~1% monthly), validated through re-measurement
Contact details (email/phone) remain 93.9-100% accurate post-job-change; job titles and employer fields decay instead—teams verify the wrong attributes
Sales function experiences highest volume of movement (121,238 in 5 months) but marketing has highest rate (13.73% annually); IT changes carry highest deal risk per occurrence
CRO transitions create tightest evaluation window (30-60 days); C-suite departures are 1.6% of volume but represent largest relationship/contract exposure
Promotions (194,165 detected) are 'invisible decay'—emails remain valid so sequences don't bounce, but messaging becomes misaligned; 7.6x more common to detect departures than promotions

a16z GP Anish Acharya positioning 'loops' as foundational to future company building
Concept remains undefined in source material - requires full podcast episode for context
Likely refers to feedback loops, process automation, or iterative product cycles but unconfirmed