Field Notes - Nov 13, '25

Executive Signals

Funnel is the product: compound step lifts beat chasing last‑click transactions
Certified first, faster after: one trusted dashboard unlocks modeling and experimentation velocity
Agents are interns, not execs : tool whitelists, audits, and human takeovers preserve trust
Logs are a bill: budget, sample, and retain only actionable classes
Pods beat projects : single ownership outperforms multitasking when speed really matters
Tokens down, trust is dear: aggressive pilots, measured outcomes, renegotiations as prices fall

One focused owner working a single problem eight hours daily will beat a multitasking team. Scale with short, planned expert injections to unblock specialized tasks without breaking flow. Publishing a one‑pager with objectives and exit criteria keeps governance tight and prevents scope creep.

Assign a single owner for 4–6 weeks; block interrupts
Schedule specialist support windows; measure cycle‑time deltas
Publish objective, metrics, and exit criteria before kickoff

Tokens Favor Early Pilots

Inference costs per conversation are falling, but trust and outcomes remain scarce. Spend to learn now with strict budgets and telemetry, then push vendors on ceilings and data terms as prices decline.

Cap daily pilot spend; publish cost per resolved intent and per booking
Require CAC payback unchanged or better within 60 days of launch
Negotiate usage ceilings and data‑egress terms up front

Marketing

External Signals Help; Funnel Math Wins

Weather, competitor promotions, ratings, and OEM ad cycles explain volume noise, but your lever is step‑to‑step conversion on owned surfaces. Focus on the step rates you control and isolate effects with clear regime tags.

Build a daily signals table and tag sessions by regime
Run diff‑in‑diff on step conversion; prioritize ≥3 pp lift
Separate ownership: acquisition vs progression

CTAs Are Product Surfaces, Not Decoration

Calls‑to‑action materially change lead flow. Treat them like features with full metadata so models can learn which variants deepen progression, not just click.

Track CTR and downstream lift; declare winners at 95% confidence
Maintain a weekly ranked top‑10; archive losers to prevent regression
Expose copy/color/size/placement in analytics for modeling

Sales

Conversational Agents With Guardrails

Reputation risk is asymmetric. Start narrow with an intent router, answer library, and whitelisted tools, then widen as confidence grows. Keep human‑in‑the‑loop for outbound and create instant takeover paths for sensitive moments.

Auto‑escalate on price, contracts, and promotions; forbid free‑form quoting
Log and audit 100% of threads; target <30% takeover and CSAT >4.3
Store conversation state in CRM; feed a next‑best‑action policy

Build vs Buy for AI Follow‑Up

Speed favors vendors; defensibility lives in workflow integration and context. Run a time‑boxed bake‑off, then decide based on time‑to‑viable and long‑run unit economics.

30‑day head‑to‑head on identical cohorts; track response, conversion, compliance
If in‑house viable ≤ 8 weeks and 30% cheaper per lead, build
Require vendor APIs for event ingestion and supervised sending before pilots

Product

Optimize Layered Outcomes, Not Just Transactions

When only a fraction of activity ties to sales, design progressive outcomes—session → lead → call → financing action → transaction. Use AutoML on event streams to discover trigger combinations that move people one step deeper; compounding lifts win.

Define 5–7 progressive outcomes with timestamps and session/customer keys
Target +3–5 pp lift at each step; roll up by brand/geo
Escape thin local data by aggregating across locations

Ship the First Certified Dashboard, Then Accelerate

The slowest work is wiring and trust. Make “First Usable Dashboard” the milestone: unified events with reconciled outcomes, certified by the business within 1–2% of source. Once trusted, modeling and experimentation accelerate.

Land analytics in the warehouse; keep ops in existing OLTP
Dedicate a no‑interrupt pod for 4–6 weeks
Version metrics; promote only after reconciliation

Engineering

Ship Observability in This Order: Sentry → Logs → Sonar

Early hardening should maximize developer signal, control cost, then enforce quality. Instrument error monitoring first, enforce logging discipline second, and gate on static analysis once builds stabilize.

Wire Sentry on critical paths with performance and session replay
Default prod logs to error; sample debug/trace; standardize an /logs path
Start Sonar locally; add a CI quality gate on blocker/critical issues

Treat Log Ingestion as a Cost Center

Teams overpay for “just in case” logs. Budget logging with SLOs, sampling, and retention tiers. Keep what you’ll act on and drop the rest.

Keep logging ≤10–15% of infra spend; investigate if >25% or >5% WoW growth
Sample noisy classes to ≤1%; redact PII at source; 7–14 day hot retention
Alert on error/latency SLOs; page only for customer‑impact thresholds

Turn Monthly Toil into Seconds—with Observability

Agentic automation can collapse week‑long, multi‑person workflows into parallel jobs. Keep humans for exceptions and build health checks from day one to guard against silent drift.

Track per‑task runtime; alert on >2x baseline variance
Rollout stages: synthetic dry‑run → gated submit → full auto
Retain 10–20% staff as controllers; wire Slack/Webhook alerts

Local testing lies when state persists forever; prod redeploys reset state and resurrect popups. Make suppression deterministic and test like production.

Include a deployment_id in suppression; expire on release or after 7 days
Add e2e tests that clear storage and validate suppression behavior
Gate popups with a server flag keyed to user/session

Use Vendor Delays to Harden Reliability

If a CRM or iPaaS handoff is weeks out, spend the slack on guardrails that compound: observability, smoke tests, and CI gates.

Add cross‑env smoke tests and block deploys on failure
Escalate if error rate >1% or p95 breach persists 5 minutes
Route issues to Slack with owners; prune dead alerts weekly

Vendors tied to a single mailbox tempt OTP sharing—an audit and security risk. Solve structurally with managed accounts, hardware keys, and pressure for proper multi‑user or SSO.

Create an IT‑owned service account; enroll multiple hardware keys; audit all use
Push vendors for multi‑user or SSO with per‑user audit; negotiate delegated access
If forced, treat shared TOTP as break‑glass: strict logging, rotation ≤30 days