Browse the book & resources

Supporting material

Book Outline: *Trail Markers*

Working support material for a provisional manuscript; not a verified claim about any reader’s organization.

Subtitle (author-confirmed 2026-11-13): Choosing Where AI Fits in Engineering

Purpose: Help engineering leaders decide where AI is worth using, where it is not, and how to tell the difference. The reader should leave able to frame a real problem, trace a proposed AI intervention to an outcome, run a bounded evaluation, and keep, change, scale, or stop it without hiding costs in another part of the system.

Core argument: More generated work is not necessarily more value. A sound decision asks what improved for customers, the company, teams, and individuals, who absorbed new work or risk, and what evidence supports each claim. AI can help; it can also amplify weak product choices, review queues, or poor operating practices. The "trail markers" are the consequential decisions—destination and owner, permitted context, verification, release, and learning—not instructions that dictate every step an engineer or agent takes. See the value synthesis and research notes.

Primary reader: CTOs, VPs and directors of engineering, and engineering managers making investment, staffing, risk, and operating decisions. Companion readers: product and finance partners, platform/DevEx leaders, staff engineers, and team leads who participate in those decisions. Speak to leaders without treating practitioners as interchangeable capacity.

Scope: AI used in software engineering and its surrounding delivery, product-discovery, and operating workflows. AI-native product features enter only when they illuminate the same value decisions; this is not a model-building manual, a tool catalog, or a promise that every team should become "AI-native." The nine engineering competencies remain the diagnostic map, not nine mandates to adopt AI. The guide's four goals (effective, efficient, innovative, engaging) remain intact; reliability is made explicit as an additional safety/outcome lens for this book, not a replacement for the framework.

Reader's recurring question: What became better for whom, compared with what, and at whose expense?

---

Structure Overview

PartThemeChapters
IDefine value and find the real constraint1–4
IIApply the nine competencies to valuable work5–13
IIIMake and test consequential decisions14–18
IVScale what works; leave room to change course19–20
—Field tools and referenceA–F

---

Front Matter

  • Preface: why "more AI" and "more productivity" are inadequate destinations; what this book will and will not claim.
  • How to read: a leader's path from problem to decision, with product/finance and practitioner prompts; cases are published accounts unless explicitly identified as firsthand.
  • Optional foreword.

---

Part I — Define value and find the real constraint

Chapter 1: Faster Is Not the Destination

  • Open with a customer outcome that failed, not a tool demo: Spotify's documented podcast-publication delay shows why a queue and reliability problem matter to creators; it is not an AI-caused incident.
  • Distinguish inputs (access, usage), outputs (code, PRs, releases), outcomes (customer success, financial results, sustainable capability), and harms.
  • Ask why more AI-generated work can increase review and verification queues rather than shorten time to customer value. Introduce the trail-marker decision cycle and the book's evidence standard.

Chapter 2: Four Perspectives on Value

  • Customer: useful, accessible, trustworthy results; company: realized returns, risk, and funded capacity; team: end-to-end flow, safe operation, learning; individual: agency, growth, fair recognition, sustainable work.
  • Separate capacity from cash savings; keep customer use, financial return, delivery activity, and lived experience in different units.
  • Show explicit trade-offs and non-negotiable safety, privacy, and fairness constraints. Use the value research, not a composite "value score."

Chapter 3: Diagnose the Engineering System

  • Use the nine competencies and five maturity levels in the guide to locate constraints, without assuming a higher maturity score alone creates value.
  • Make effective, efficient, innovative, engaging the guide's four goals; track reliable explicitly so faster delivery cannot mask customer harm.
  • Trace a named customer outcome through discovery, design, build, review, approval, release, and operation. Distinguish active work, necessary safeguards, external waits, and avoidable queues; find the person or team receiving the extra work.

Chapter 4: Where AI Fits—and Where It Does Not

  • Briefly distinguish assistants, agents, retrieval, pipeline automation, and product-facing AI by task and risk; avoid a version-specific product survey.
  • Decide to use AI, improve the surrounding workflow without AI, defer, or decline. Evaluate context availability, reversibility, verification cost, permissions, and who owns the result.
  • Match autonomy and human checkpoints to consequence; preserve a justified right not to use AI and protected time for independent practice.

---

Part II — Apply the nine competencies to valuable work

> Common chapter pattern (Chapters 5–13): Start with a named outcome and the local constraint; show an AI option and a non-AI option; specify the decision owner, context and safety checkpoint; trace customer, company, team, and individual effects; define a primary outcome, baseline, comparison, cost, and counter-metrics; end with a keep/change/stop test. Use published cases only with their limits stated. Maturity levels describe organizational capability, not a prescribed AI usage ladder.

Chapter 5: Delivery & Execution — Are We Working on the Right Promise?

  • Diagnose shifting priorities, large batches, and dependencies before using AI to decompose plans or summarize status.
  • Trace a promised result through all handoffs; distinguish a safer release gate from avoidable waiting. Evaluate delivered customer outcomes, lead time, and team workload—not tickets closed.

Chapter 6: Engineering Quality & Reliability — Who Verifies and Recovers?

  • Compare AI-assisted tests, code review, triage, and incident follow-ups with better test design, review capacity, and operational discipline.
  • Define "substantial correction" before counting it; include rejected changes, reviewer effort, escaped defects, recovery, and on-call load. Never relabel a subsequent commit as a substantial AI failure.

Chapter 7: Flow & Developer Experience — Where Did the Queue Move?

  • Map coding, CI, review, approvals, and deployment; ask whether AI saved time end to end or shifted it to reviewers and partner teams.
  • Assess coding assistants and agents alongside small batches, self-service, and removal of needless handoffs. Pair lead time with rework, safety, and sustainable focus.

Chapter 8: Architecture & Technical Strategy — Can We Change and Undo Safely?

  • Use AI for codebase exploration and bounded modernization where an owner can explain and verify the design; compare with documentation and dependency cleanup.
  • Test whether a newcomer can identify an owner, run tests, and explain rollback limits in a safe exercise. Account for migrations and external side effects that rollback does not undo.

Chapter 9: Product & Customer Focus — Did Anyone Benefit?

  • Tie feedback analysis and prototyping to an actual customer job; distinguish time to validated learning from feature adoption, task success, and revenue.
  • Segment affected users; include trust, accessibility, and complaints. Product-facing AI is an option, not the default answer to every unmet need.

Chapter 10: Innovation & Learning — What Did We Learn That Changed a Decision?

  • Compare AI-assisted experiments with non-AI discovery and prototyping; make kill decisions as legitimate as scale decisions.
  • Measure idea-to-evidence time, decision quality, and transfer of proven practices, not hackathon or demo counts.

Chapter 11: People & Culture — Who Gets to Grow and Say No?

  • Protect mentoring, deliberate practice, and the ability to evaluate AI output. Examine whether junior engineers can explain changes and whether seniors have time to give "why" feedback.
  • Make justified non-use and risk detection legitimate performance evidence; assess autonomy, workload, skill development, and fair advancement without individual usage surveillance.

Chapter 12: Collaboration & Communication — Is the Context Usable?

  • Investigate handoffs, ownership, stale knowledge, and onboarding before adding AI search, summaries, or agent access.
  • Test an ordinary newcomer's ability to find the current owner and run the relevant tests; verify source freshness, access boundaries, and reductions in repeat questions or blocker time.

Chapter 13: Security & Compliance — What Is the Cost of a Mistake?

  • Scope data access, permissions, human approvals, auditability, and reversibility by risk; consider AI for vulnerability work only with independent validation.
  • Track exposure and remediation alongside review delay and customer trust. Do not equate absence of an incident in a small pilot with proof of safety.

---

Part III — Make and test consequential decisions

Chapter 14: Choose a Bet Worth Testing

  • Start with a problem, a recipient, and a named owner; compare AI, non-AI workflow improvement, and doing nothing for now.
  • Write the causal chain from intervention to outcome; nominate one primary value claim, checks for the other three perspectives, full costs, and conditions for stopping.

Chapter 15: Set Boundaries Before Delegating

  • Publish usable policy for data, permissions, audit, and justified non-use; distinguish drafting from production-changing agency.
  • Define evaluation on representative tasks, human decision rights, rollback limits, and escalation for high-consequence work. Policies are guardrails, not a usage quota.

Chapter 16: Measure the Result, Not the Usage

  • Choose a baseline and comparable work or a phased/randomized rollout where feasible; record changing task mix and simultaneous process changes.
  • Trace tool use → validated change → customer outcome; reconcile financial returns without counting created capacity as cash savings or counting the same avoided loss twice.
  • Use DORA/SPACE/DevEx measures as diagnostics and guardrails, not a personal productivity score; distinguish instrumented outcomes, self-reports, modeled estimates, and unknowns.

Chapter 17: Lead the Human Transition

  • Include product, finance, security, reviewers, operators, and affected engineers in defining success; do not treat "resistance" as a defect to be overcome.
  • Protect practice and mentoring, recognize thoughtful refusal as well as effective use, and check whether benefits and review burdens are distributed fairly.

Chapter 18: Build Only the Enabling Platform You Need

  • Give teams usable context, approved access, observable workflows, and inexpensive ways to test and verify without building a grand platform first.
  • Fund reviewer capacity, evaluation, support, and ongoing model/compute costs; retire or change infrastructure when its measured benefits do not justify it.

---

Part IV — Scale what works; leave room to change course

Chapter 19: When to Scale, Change, or Stop

  • Scale a workflow only when its stated primary value claim withstands scrutiny, its full costs are understood, and the other perspectives' guardrails hold; change or stop when they do not.
  • Revisit ownership, capabilities, policy, and evidence as tools change. Avoid "Level 5 = maximum automation": maturity means better choices and safer adaptation.

Chapter 20: Your Next Ninety Days

  • Days 1–30: pick a customer problem and map its queues, people, baseline, and alternative solutions.
  • Days 31–60: run one bounded experiment with a comparison and pre-agreed pause/stop conditions.
  • Days 61–90: review all four perspectives with stakeholders; document a keep/change/scale/stop decision and the next unanswered question.

---

Appendices

  • A. Nine-competency reference and organizational self-assessment (cross-reference the guide).
  • B. Decision cards: AI, non-AI, defer, or decline; risk and verification prompts.
  • C. Four-perspective value ledger, pilot design, and keep/change/scale/stop template.
  • D. Example AI usage policy, including justified non-use, data boundaries, approval, and ownership.
  • E. Measures and evidence guide: definitions, units, counter-metrics, full-cost accounting, and causal caveats.
  • F. Case register, source notes, and glossary.

---

Evidence and writing plan

  • The seven files in interview/ are desk-researched question-and-answer explorations, not interviews conducted for this book. Use them for chapter-opening questions and gaps to investigate locally; do not quote their synthesis as participant testimony. Do not invent missing company policies, reviewer hours, correction rates, or pilot outcomes.
  • Label each case as a firsthand study, published organizational account, vendor/customer case, or non-AI comparison. Separate measured observations from self-reports, financial models, and this book's recommendations. Where a before/after account changed other processes too, do not attribute all improvement to AI.
  • Carry several linked investigations across chapters: Spotify's customer delay (non-AI) for outcomes and reliability; Salesforce's review constraint for shifted work; owner/test/rollback exercises for operational independence; the right to decline AI and junior learning for individual value; and a bounded pilot with a full-cost ledger for cross-perspective decisions. These are examples of different settings, not one continuous case or causal proof.
PartChaptersStatusMain-text targetWriting priority
I1–4Not started~16,000 wordsDefine the promise, lenses, constraint map, and decision rules.
II5–13Not started~40,500 words~4,500 per chapter; apply the common outcome-led pattern.
III14–18Not started~20,000 words~4,000 per chapter; make the evaluation usable.
IV19–20Not started~6,000 words~3,000 per chapter; close with a decision and plan.

Estimated main text: ~82,500 words, excluding front matter and appendices. Working targets, not commitments to pad chapters.