Field Notes - Oct 21, '25

Executive Signals

• RACI is the new roadmap: clarify AI decision rights before sprints
• SSO is the new VPN: enterprise AI needs audit trails and model defaults
• Spec is a prompt: acceptance tests encoded, shipped, and regressable like code
• Volume x delay is gold: automate queues before approvals to unlock ROI
• Add-ons aren't confetti: price integrity lives outside core bundles and sales comp

CEO

CTO-Led AI Governance

Mixed-seniority teams revert to legacy habits without explicit ownership. Centralize AI process changes through the CTO to protect engineering authority while welcoming cross-functional input. Champions influence through artifacts and metrics, not pronouncements. Decisions land only after a written cutover plan is signed, and friction is handled privately before it becomes public process.

• Publish a one-pager RACI for AI projects and cutover authority
• Hold weekly 30-minute AI office hours to capture ideas without derailing sprints
• Require written RFCs and timeboxed experiments before any rollout

Sales

Protect Price Integrity by Gating Add-Ons

When the core SKU is high-ticket, tossing in AI add-ons during late-stage deals erodes ARPU and confuses value. Move add-ons to Client Services or Launch, where education is deeper and discount pressure is lower. Treat expansion as a post-sale motion with clear SKUs and proof-focused demos.

• No comp for bundling add-ons into core; full credit for post-sale expansion
• Separate SKUs and decks; require a demo checklist before quoting
• Review discount rate by channel quarterly; clamp leakage where it creeps

Product

Prompting Is the New Spec

Durable leverage is in precise intent, constraints, and acceptance tests, not more keystrokes. Treat prompts as living PRDs that engineers and agents can execute and regress. Codify role, goal, schema, and evals so product intent survives implementation and iteration.

• Adopt “spec as prompt” templates with role, goal, constraints, schema, evals
• Coach for prompt hygiene with canonical examples, counter-examples, edge cases
• Add an automated eval harness to regression-test prompts like code

Automate Where Volume and Wait Times Compound ROI

Repetitive workflows that stall in compliance or approvals are first-win territory. If a process repeats often and idles for days, even modest automation moves material throughput. Start with prep, QA, and queue management, keeping final approvals human-in-the-loop, and measure latency reductions alongside hours saved.

• Select 2–3 high frequency x delay flows and baseline cycle time now
• Automate prep, QA, and queueing first; keep humans for final approval
• Track hours saved per month and latency reduction for payback

Engineering

Standardize the AI Tool Stack Before Training

Personal LLM licenses and ad hoc copilots create drift and risk. Freeze sprawl, select one or two primary models with documented fallbacks, and move to an enterprise plan with SSO, data controls, and auditable logging. Keep partner discounts on a need-to-know basis until procurement closes.

• Issue company accounts; sunset personal logins within 14 days
• Decide default model plus fallbacks and document when to use each
• Gate access via SSO; log prompts and outputs to a secure audit store

Build Taste With Hands-On Demos, Not Memos

Teams absorb new platform surfaces by feeling “what good looks like.” Run small novelty-app sprints across chat-native apps, agents, and retrieval to calibrate UX and latency bars. Promote what sticks, kill what does not, and turn findings into reusable patterns that speed future builds.

• Host a weekly 60-minute App Hour: one tiny build, one teardown
• Enforce a two-week shelf: promote winners, retire the rest
• Capture reusable patterns for batching, caching, and schema prompts