Browse the book & resources

Supporting material

Book research: engineering bottlenecks and AI adoption myths

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

Research snapshot: November 2026. Working notes for Trail Markers and the book outline.

Research question and working thesis

What keeps engineering organizations from becoming effective, efficient, reliable, innovative, and engaging, and which mistaken assumptions lead managers to adopt AI without improving business or employee outcomes?

Working thesis: AI can increase the pace at which work is produced, but organizational value depends on whether teams choose the right work, move it through the entire delivery system, verify it, and learn from the result. The book's “trail markers” should be explicit decision points—outcome definition, context and permissions, review and evaluation, deployment, and learning—not instructions controlling every action an engineer or agent takes. This interpretation follows the README and the nine areas in the engineering competency guide.

This is a research synthesis, not a claim that every organization has every problem or that the “myths” below have been measured as the beliefs of all managers. Surveys describe respondents' experiences; experiments estimate effects in particular settings; the diagnostic questions and interventions below are proposals for the book. Do not convert correlations or self-reported hours saved into a universal AI ROI estimate.

1. Five outcomes: engineering team and hiking perspectives

The competency guide names four organizational goals: effective, efficient, innovative, engaging. For this book's measurement lens, add reliable as a fifth, explicit outcome. This does not replace the guide's nine competencies: Engineering Quality & Reliability and Security & Compliance already contribute to its “effective” goal. Separating reliability makes the cost of moving faster unsafely visible.

OutcomeEngineering team: what good looks likeOutdoor hike: what good looks likeCandidate team-level measures
EffectiveSolve a real customer or business problem, not merely close tickets.Choose a worthwhile destination and reach it.Share of initiatives that meet a pre-agreed customer or business outcome; feature adoption or customer retention where relevant; outcomes reviewed after launch.
EfficientMove valuable work through the whole system with less waiting and rework.Take a sound route without needless detours or exhausting the group.Idea-to-production or change lead time; time waiting for review, CI, access, or another team; rework and unplanned-work share.
ReliableDeliver and operate safely, detect failure, and recover promptly.Navigate hazards, adjust to the weather, and get everyone home.Change failure rate; failed-deployment recovery time; customer-impacting incidents and SLO attainment; material security events.
InnovativeTest promising ideas and spread practices that prove valuable.Explore a better route when the familiar trail no longer reaches the goal.Time from idea to validated learning; share of experiments with a keep/kill/scale decision; uptake and measured benefit of successful changes.
EngagingGive people meaningful ownership, support, sustainable work, and room to grow.Let every hiker contribute, learn, and want to hike again.Team engagement and participation in surveys; mentoring and skill-development opportunities; sustainable on-call load; internal mobility or regretted attrition.

These are candidate indicators, not universal targets. Agree on the outcome and its baseline before a change; track trend and task mix, not just a single month's number. Pair each improvement with a counter-metric: faster lead time with reliability and reviewer load; fewer incidents with release delays; faster experiments with customer value; higher engagement with actual learning and sustainable workload. Use DORA-style delivery measures and SPACE's broader view of satisfaction, performance, collaboration, and flow together rather than judging individuals by activity. 114 In the metaphor, AI is equipment; the trail markers are the destination, route decisions, safety checks, learning checkpoints, and responsibility for the group.

2. Persistent pain points and bottlenecks

The recurring pattern is a mismatch between local production speed and end-to-end capacity. DORA's 2025 survey found AI adoption associated with higher delivery throughput and product performance and greater delivery instability; its 2024 report had found a negative relationship with both throughput and stability. These are observational findings from different years, not proof that AI caused either change. 12 Atlassian's 2026 developer-experience survey reports that 87% of developer respondents say AI helps them ship faster, while 72% report the same or more friction than before AI. Its top reported friction points include finding information, meetings, context switching, technical debt, and unclear requirements. 3

“Outcomes affected” means that a weakness in the competency can harm those outcomes, not that the outcomes constrain the competency. The first is the primary outcome at risk; subsequent outcomes are also affected. The justification gives the causal reasoning for each mapping. These are diagnostic hypotheses, not a claim that the nine competencies or five outcomes are independent.

Competency from the bookOutcomes affectedJustificationEnduring pain point and likely constraintWhat to look for locallyBusiness and employee consequence
1. Delivery & ExecutionEffective; efficient, engagingUnclear priorities can deliver the wrong result (effective); replanning and dependencies waste effort (efficient); repeated deadline pressure drains the team (engaging). Reliability can suffer too if deadlines cause skipped safeguards, but that is not automatic.Too many commitments, shifting priorities, vague acceptance criteria, and large work items. Faster generation fills a queue whose prioritization has not improved.Work in progress, carryover, blocked-item age, planned versus achieved customer outcomes; interview product and engineering together.Unpredictable delivery, wasted effort, stressful deadline cycles. DORA's 2024 report links unstable priorities to lower productivity and more burnout. 2
2. Engineering Quality & ReliabilityReliable; effective, engagingWeak tests and slow recovery make failures more likely or costly (reliable); incidents disrupt customer outcomes (effective); repeated firefighting burdens reviewers and on-call engineers (engaging).Testing, review, release safety, and incident response cannot keep pace with changes. “Almost right” output adds verification and debugging work.Review wait and rework, failed deployments, recovery time, escaped defects, SLO breaches; segment by change risk rather than counting all changes alike.Outages and expensive rework; reviewers and on-call engineers absorb the extra load. In Stack Overflow's 2025 survey, 66% cited nearly-right AI solutions as a frustration and 45% cited more time-consuming debugging. 4
3. Flow & Developer ExperienceEfficient; reliable, engagingWaiting and context switches slow end-to-end work (efficient); rushed handoffs or weak feedback can let defects through (reliable); persistent friction makes work less satisfying (engaging).Waiting for reviews, builds, environments, approvals, and access; frequent context switching. Coding is only one stage.Map idea-to-production elapsed time by working time versus waiting time; examine CI duration, review turnaround, and self-service success.Local time savings fail to become shorter customer lead time. Atlassian's 2025 survey found 50% of respondents reported losing 10+ hours a week to organizational inefficiencies; these are self-reports, not measured recoverable hours. 5
4. Architecture & Technical StrategyEfficient; reliable, innovativeCoupling forces changes through multiple teams (efficient); unclear boundaries make failures harder to isolate or roll back (reliable); legacy constraints raise the cost of trying new approaches (innovative).Tightly coupled services, unclear boundaries, legacy dependencies, and unmanaged debt make changes hard to reason about or deploy independently.Cross-team dependencies per change, time to understand a system, ownership gaps, rollback difficulty, age of critical debt.Growing coordination cost and risk; AI may produce plausible changes without understanding implicit constraints. DORA emphasizes loosely coupled architecture and fast feedback as conditions for realizing gains. 1
5. Product & Customer FocusEffective; innovativeWithout a customer problem and success measure, teams may ship the wrong thing (effective); without feedback, promising ideas cannot be tested and refined (innovative).Teams optimize tickets, features, or code volume without a tested customer problem or post-launch measurement.Share of initiatives with a hypothesis, baseline, target outcome, and post-launch review; adoption, retention, customer effort, or revenue as appropriate.Shipping the wrong thing faster. DORA's AI capabilities work finds user focus is an important condition for improved team performance. 6
6. Innovation & LearningInnovative; engaging, effectiveDemos without decisions or shared lessons do not create repeatable change (innovative); no room to learn limits growth (engaging); unvalidated ideas consume capacity without helping customers (effective).Experiments are either unfunded or celebrated as demos without a keep/kill/scale decision; lessons stay inside a pilot team.Protected learning time; experiment-to-decision time; number of tested hypotheses and repeatable practices adopted by other teams.Novelty without durable advantage; employees have little room to develop judgment. The wider AI survey evidence distinguishes experimentation from scaled business impact. 7
7. People & CultureEngaging; innovative, reliableAnxiety, overload, and weak mentoring harm retention and growth (engaging); inadequate learning slows adoption of better practices (innovative); exhausted or undertrained reviewers may miss risks (reliable).Role anxiety, tool pressure, uneven access to training, weak mentoring, and insufficient time to learn to verify AI work.Training by role and seniority, mentoring time, confidence in reviewing generated changes, engagement, on-call burden, internal mobility.Skill atrophy and disengagement can offset short-term speed. The World Economic Forum's 2025 cross-industry employer survey found skills gaps were the most cited barrier to transformation (63%); this is not an engineering-only or AI-only estimate. 8
8. Collaboration & CommunicationEfficient; engaging, effectiveMissing context and unclear ownership create search and handoff delays (efficient); repeated blockers frustrate colleagues (engaging); lost customer or decision context can lead to the wrong solution (effective).Decisions, service knowledge, and customer context are scattered; ownership and escalation paths are unclear.Search success, onboarding time, stale docs, cross-team blocker age, repeated questions, handoff count.Engineers and agents both work from incomplete context. Atlassian reports information-finding and cross-team collaboration as persistent friction. 35
9. Security & ComplianceReliable; effective, efficientLeaks, excessive permissions, or unsafe agent actions can compromise trustworthy operation (reliable); customer harm undermines the intended outcome (effective); late-stage reviews cause rework and delay (efficient).Permissions, data handling, supply-chain review, and AI usage rules are unclear or imposed only at the end. Agents can magnify the consequences of over-broad access.Approved-tool coverage, sensitive-data exposure tests, privilege scope, audit trail, security review delay, remediation time.Breach, regulatory or IP exposure, and shadow use. OWASP identifies prompt injection, sensitive-information disclosure, and excessive agency among LLM application risks; NIST recommends risk-based governance and evaluation. 910

Cross-cutting bottleneck: the scarce resource may shift from writing code to specifying, integrating, reviewing, and operating it. The question is not “How much more can we generate?” but “Where does a valuable change wait, fail, or lose its connection to customer need?” DORA's seven AI-enabling capabilities—clear policy, healthy data, AI-accessible context, version control, small batches, user focus, and quality internal platforms—are useful candidate trail markers, not a claim that buying a platform alone solves all nine competencies. 6

3. Management myths about AI adoption, role, and value

These are testable management assumptions, not diagnoses of individual leaders. Each should be presented with the strongest evidence and its boundary conditions.

Myth / assumption to challengeBetter working model and evidenceLeadership question or practice
“Buying licenses is an AI strategy.”Adoption is an input. In McKinsey's 2025 cross-industry survey, nearly two-thirds said their organizations had not started scaling AI across the enterprise and only 39% attributed any enterprise-level EBIT impact to it; respondents at organizations reporting more value also reported more workflow redesign. These are associations and self-attributions, not causal ROI proof. 7What customer, financial, risk, or employee outcome will this use case change, and who owns the workflow around the tool?
“More AI usage means more productivity.”Usage can be performative. In Atlassian's 2026 survey, 60% of developer respondents said AI usage is treated as a productivity signal and 41% said they sometimes use it more than necessary because they feel expected to. 3Measure outcomes and guardrails, not prompts, token spend, or percentage of AI-authored code. Allow a reasoned “do not use AI here.”
“Faster coding means faster delivery and better business results.”Delivery has queues, dependencies, and verification stages. DORA 2025 reports higher throughput alongside instability; Atlassian 2026 reports faster shipping alongside persistent friction. 13Trace one change from customer problem to observed outcome; address the longest wait or highest-risk handoff first.
“AI will fix weak engineering practices for us.”DORA describes AI as an amplifier: platform quality, small batches, version control, usable internal data, and customer focus condition its benefits. This does not mean every team must finish a giant transformation before trying AI. 16Pair a bounded AI pilot with one improvement to its surrounding workflow and one safety counter-metric.
“A reported task speedup is a universal headcount or ROI multiplier.”Results depend on task, worker, tool generation, and metric. Three field experiments involving 4,867 developers found an average 26.08% increase in completed tasks with a coding-completion assistant; this does not imply a 26% increase in revenue or a 26% reduction in staffing. 11 A separate early-2025 trial of 16 experienced open-source developers working in familiar repositories found 19% longer task times, but METR explicitly labels this result historical, not current. Its late-2025 follow-up suggests possible speedups but says selection effects make their magnitude unreliable. 1213Run small, task-specific comparisons with review, rework, cost, and quality included; revisit estimates as tools change.
“AI can replace engineering judgment, ownership, and mentoring.”Generated output still needs acceptance criteria, domain context, security judgment, testing, and an accountable owner. Stack Overflow's 2025 survey found more respondents distrusted AI-output accuracy (46%) than trusted it (33%); the survey measures sentiment, not actual defect rates. 4Name a human decision owner for high-impact actions; teach engineers, especially newer ones, to explain, test, and challenge outputs.
“More automation requires less investment in people.”The role can move toward problem framing, orchestration, evaluation, architecture, and customer learning. Whether time saved becomes growth or job loss is a leadership choice, not a mechanical result of the tool. WEF reports anticipated reskilling needs across industries, not an engineering headcount forecast. 8Budget learning time, mentoring, and updated career expectations alongside tooling; track internal mobility and engagement.
“Either prohibit AI to be safe or give agents unrestricted autonomy to move fast.”Both extremes ignore the difference between a low-risk draft and a production write operation. NIST recommends risk management aligned with context; OWASP documents risks from prompt injection, information disclosure, and excessive agency. 910Publish a usable policy; scope permissions, log actions, test on representative cases, and require approval for irreversible or high-risk actions.
“One metric tells us whether developers are productive.”The SPACE framework treats satisfaction, performance, activity, collaboration, and efficiency/flow as distinct dimensions; activity alone is not productivity. 14Use team-level outcome, delivery, reliability, and experience measures together; avoid individual ranking by code or AI use.
“Success means cutting cost, not creating new value.”Efficiency matters, but the business case can also include customer experience, new product capability, risk reduction, and employee growth. McKinsey's survey reports that organizations seeing the most AI value often pursue growth or innovation as well as efficiency; this is descriptive, not a guarantee. 7For each initiative, state both a company benefit and an employee-development benefit, plus the cost of verification and operation.

4. Proposed research-to-practice structure for the book

The book can make every competency chapter follow a common investigation: symptom → constraint → AI opportunity → trail marker → outcome and counter-metric → employee growth. For example, for a slow change: identify whether the constraint is missing product clarity, review capacity, flaky CI, or permissions; use AI only where it reduces that constraint; put a human checkpoint before risky deployment; observe lead time and failed-deployment recovery and reviewer load. This applies the outline's chapter template without assuming one maturity score or tool fits every team.

Questions for leader and practitioner interviews (to validate these hypotheses):

  1. Which customer outcome was last delayed despite engineers feeling busy? Where did the work actually wait?
  2. Which work became faster with AI, and which downstream team received more work as a result?
  3. What percentage of AI-assisted changes required substantial correction, and who performed that correction?
  4. Can a new engineer find a service owner, run its tests, and explain a safe rollback without asking a particular veteran?
  5. When is it acceptable not to use AI? Are people rewarded for critical evaluation rather than visible usage?
  6. Which skills are engineers gaining or losing? How are junior engineers getting feedback and practice?
  7. What would count as evidence that a pilot helped customers, improved financial performance, or reduced operational risk—not merely that employees liked a tool?

Suggested measures for a bounded pilot: choose one team, one recurring workflow, and a primary outcome from the five-outcome table; record a pre-pilot baseline and compare similar work over time (or use a credible comparison group where possible). For an AI-assisted release pilot, track whether the intended customer result occurred (effective), end-to-end lead and waiting time (efficient), failed changes and recovery (reliable), lessons that lead to a keep/kill/scale decision (innovative), and review load, mentoring, and engineer sentiment (engaging). Record tool and verification cost. The primary outcome matters most; the other four are checks for unintended effects, not five targets to maximize at once. Small samples, changing task mix, and concurrent process changes limit causal claims. Never use team diagnostics to score individuals.

Sources and evidence cautions

  1. DORA / Google Cloud, 2025 report announcement and full report — nearly 5,000 technology-professional survey responses and qualitative work; associations and respondent perceptions, not randomized causal estimates.
  2. DORA, 2024 State of DevOps report — earlier survey wave, useful for the change in findings and for priority instability.
  3. Atlassian, State of Developer Experience 2026 summary — developer/leader survey as described by its publisher; self-reported speed, friction, and perceptions of strategy, not instrumented end-to-end productivity.
  4. Stack Overflow, 2025 Developer Survey: AI — voluntary developer survey; use the actual respondent group and distinguish “using or planning to use” from current usage.
  5. Atlassian, State of Developer Experience 2025 summary — 3,500 developers and managers across six countries; estimated time saved/lost is self-reported, and Atlassian sells developer tools.
  6. DORA / Google Cloud, 2025 AI Capabilities Model — survey-based analysis of enabling practices, not proof that a particular purchase produces the same effect everywhere.
  7. McKinsey, The State of AI in 2025 — cross-industry survey of 1,993 participants; EBIT impact is respondent-attributed and workflow-redesign findings are correlational.
  8. World Economic Forum, Future of Jobs Report 2025 — employer expectations across industries about future skills and transformation; not a measured AI impact on engineering jobs.
  9. OWASP, Top 10 for LLM Applications (2025) — threat taxonomy and mitigation guidance, not incident prevalence at a specific organization.
  10. NIST, AI Risk Management Framework: Generative AI Profile (2024) — voluntary risk-management guidance, not a blanket rule to prohibit AI.
  11. Cui et al., field experiments with software developers — randomized access to a coding assistant; outcome is completed tasks, and individual experiments are noisy.
  12. METR, early-2025 experienced open-source developer trial — 16 developers, 246 tasks; narrow setting and tools from early 2025; authors say this historical result is out of date.
  13. METR, February 2026 experiment update — late-2025 tools show signs of speedup, but recruitment/task-selection effects make the current magnitude unreliable.
  14. Forsgren et al., “The SPACE of Developer Productivity,” ACM (2021) — conceptual measurement framework, not an AI-effect estimate.