Supporting material
Value is not one number
Working support material for a provisional manuscript; not a verified claim about any reader’s organization.
Research snapshot: November 2026. This is a synthesis of published research and engineering accounts found through Exa and checked against the cited publications. It is not a firsthand interview, an AI vendor benchmark, or a measured result for this book's readers. The research notes and desk-researched interviews provide the wider engineering context.
The four perspectives
An engineering initiative creates value when it changes an outcome that matters to someone, after the costs and harms of producing that outcome are considered. A faster tool is an input; a merged pull request is an output; neither is the outcome by itself. The four perspectives are related but should not be collapsed into a single productivity score.
| Perspective | What value means | Evidence that could establish it | What does not establish it alone |
|---|---|---|---|
| Customer | People can accomplish an important job with less effort, delay, or risk; the product meets a need they actually have. Include users, buyers, creators, and other affected people, whose interests may differ. | Task success, time to a useful result, appropriate adoption and retention, complaint or support trends, accessibility, trust, and customer-impacting reliability, segmented by customer group. | Features shipped, release frequency, or a satisfaction quote from a few enthusiastic users. |
| Company | The organization can sustain its mission: profitable growth or mission outcomes, lower realized cost, managed risk, and useful future capacity. | Incremental contribution margin, actual expense reduction, cost per validated outcome, avoided-loss estimates with stated probabilities, and the use made of freed capacity. | Multiplying self-reported hours saved by salaries; treating more PRs as revenue; calling unused capacity cash savings. |
| Team | A group can repeatedly deliver and operate valuable changes, learn, and support one another without passing unpriced work or risk to another group. | End-to-end lead and waiting time, review/rework load, reliability, cross-team handoff delays, onboarding success, and the rate at which experiments yield decisions. | Local coding speed, one team's throughput, or a dashboard that hides reviewer and on-call work. |
| Individual | A person has useful agency, meaningful work, fair recognition, sustainable workload, and growing competence. This is a benefit in its own right, not just a retention lever. | Voluntary, protected feedback about autonomy and workload; observed opportunities to practice, explain, and develop skills; mentoring access and fair advancement over time. | Seats activated, prompt counts, purported “hours saved,” or an individual's lines of code. |
These are proposed definitions and measures, not claims that any single source proved all four. The engineering competency guide frames organizational goals as effective, efficient, innovative, and engaging; the research notes make reliability explicit so speed cannot conceal harm. Those five outcomes describe how an engineering organization performs. The four perspectives here describe who receives the value and who bears the cost.
What the evidence says
1. Customer value: the outcome arrives and works
Spotify's June 24, 2026 podcast-publishing incident is an unusually concrete example of lost value: video episodes normally available within minutes were delayed for hours; creators reported missing episodes and some re-uploaded them. A full transcoding queue, competing batch work, increased per-item processing costs, and a scheduling bug contributed. The report documents customer delay and operational failure, not a revenue figure or an AI-caused incident. It also does not give a count of affected creators. For this workflow, a sensible outcome measure would be time from a creator's upload to a playable episode, with queue age and creator notification as diagnostic and trust measures.
Speed to learning matters too. The Thoughtworks/Etsy account reports reducing a pilot team's time to validate an idea with customers from about 50 days to five. That is a reported change in the discovery process, not an AI result or proof that every idea produced value. Test what happens after the release: did the intended customers succeed, continue to use the result, or say it solved the problem? Do not mistake faster exposure to a bad product for customer value.
2. Company value: benefits survive a full-cost calculation
Company value can be realized revenue or margin, actual expense savings, avoided costs, reduced expected losses, or capacity redeployed to worthwhile work. These are not interchangeable. IBM's AI-assisted development ROI account explicitly separates capacity creation from cash savings, calls for counting review, rework, integration, governance, and operating costs, and reports an approximately tenfold modeled return for its IBM Bob use in a RevTech program. Its baseline uses matched pre/post twelve-week windows rather than a randomized control. The article is published by a company selling the tool, and its monetary valuation depends on effort estimates and business-loss assumptions: read the result as a vendor-published model, not an independently verified cash return or a general AI ROI multiplier.
The Superhuman/AWS account is a useful check on financial overreach. For 410 engineers active in both comparison periods, it reports merged PRs per engineer per week rising from about 2.5 to 4.15 (H1 2025 to H1 2026), while median PR cycle time stayed near 20 hours. That measures more merged work by the same cohort, not its incremental profit or customer effect. Enablement, platform changes, and workflow redesign occurred alongside AI adoption. The account is coauthored by the customer and AWS, which has a commercial interest in the platform. Count the additional validation and operating work before crediting net value.
3. Team value: remove the constraint, not just move it
The Microsoft-led analysis of three field experiments randomized access to a coding assistant across 4,867 developers and found 26.08% more completed tasks in the pooled estimate (standard error 10.3%). This is stronger causal evidence for the task outcome in those settings than a usage survey. It did not measure an equivalent increase in customer benefit, company revenue, or team well-being.
Salesforce describes the mechanism that can stop a local gain from becoming a team gain: more generated code and larger PRs increased human review pressure. In Superhuman's account, PR volume rose while median PR cycle time remained flat; it also reports rework on 28% of PRs, without showing what that rate would have been absent AI. These are organizational accounts with different definitions, not comparable treatment effects. Follow work from problem to release and operation; find where the queue grows, who verifies the output, and whether safety gates still work. The SPACE framework is a reminder that performance, satisfaction, activity, collaboration, and flow are distinct dimensions; activity by itself is not developer productivity.
4. Individual value: gains need not erase agency or learning
Anthropic's study of its own workforce surveyed 132 engineers/researchers and interviewed 53. Respondents described reaching unfamiliar work and prototyping faster, while some worried about lost debugging practice, fewer human questions, and weaker mentoring. These are reported experiences at an AI company, not before/after skill tests or evidence that everyone is losing expertise. A junior–senior mixed-methods study examined junior AI-assisted debugging and senior review of code, prompts, and reflections; it identifies ways to examine reasoning but does not establish a universal rate of skill loss.
Being offered a tool is not the same as having a meaningful choice about its use. A reasoned decision to work without AI on a high-context task, a sensitive dataset, or a learning exercise may protect individual judgment and customer safety. The published accounts do not show that organizations consistently reward such decisions in performance reviews; see the desk-researched non-use interview. Protect time for mentoring and independent practice, and ask people about workload without turning their answers into individual surveillance.
Where the perspectives conflict
| Apparent gain | Possible transfer of cost or harm | Decision check |
|---|---|---|
| More code completed | Review, security, support, or on-call queues grow; the customer waits just as long. | Measure the entire path to a customer outcome and record downstream labor by role. |
| Lower unit cost | People spend freed time on unprioritized output, or costs recur as rework and incidents. | Distinguish booked savings from capacity and verify where that capacity went. |
| Faster release | Safeguards are bypassed, errors escape, or a subset of customers cannot use the change. | Pair lead time with task success, change failure, recovery, and affected segments. |
| Easier AI-assisted work | Junior engineers get fewer explanations and less independent practice; reviewers carry more cognitive load. | Pair perceived ease with demonstrated learning, mentoring access, and sustainable review load. |
| Short-term safety through extra approvals | Valuable, low-risk changes wait unnecessarily; teams lose autonomy. | Keep risk-proportionate gates and compare the delay with the harm prevented. |
These are failure modes to test, not statements that every AI rollout causes them. An improvement in one column is not permission to worsen another invisibly. Nor must every perspective improve on every change: an organization can choose a temporary trade-off, but it should name who pays, obtain the appropriate agreement, and set a review date. Customer safety, privacy, and fair treatment are constraints, not quantities to trade away for PR volume.
A practical value test for a pilot
- Name the recipient and the job. State which customer group and workflow will change, and whose time or risk may increase. Choose one primary value claim: customer outcome, financial return, team capability, or individual growth. Record the other three as consequential checks rather than forcing a composite score.
- Write the causal chain before buying or scaling. For example: an AI-assisted triage system → more accurately classified reports → less queue time → faster verified fixes → shorter customer resolution time. A faster classification step is not the final customer result. Name an accountable owner for each handoff.
- Record baselines, units, and comparison. Use the same eligible population, risk mix, and observation window before and after; where feasible, randomize eligible work or phase rollout to comparable teams. Log concurrent changes such as staffing, release-process changes, and seasonal load. Keep elapsed time, engineer-hours, and money in separate columns.
- Pair an outcome with a guardrail for each perspective. An illustrative triage pilot could use customer resolution time as its primary measure; verified resolution and customer complaints as customer checks; incremental contribution and full operating cost as company checks; review queue age, escaped defects, and on-call work as team checks; and voluntary workload/learning feedback as individual checks. Define what deterioration triggers a pause before the pilot begins.
- Reconcile the ledger. Count licenses, tokens/compute, implementation, training, human review, rework, integration, compliance, and ongoing operation. Finance owns the interpretation of booked savings and margin; product/customer representatives validate the customer result; engineering and security own reliability; the people doing the work describe its effects on them. Do not double-count freed hours as both payroll savings and new output.
- Decide and revisit. At a pre-agreed date, keep, change, expand, or stop the workflow based on the primary outcome and the guardrails. If the sample is small or the comparison weak, report directional evidence and uncertainty rather than claiming a causal return. A pilot that rules out a harmful or unprofitable use case has produced learning, even if no tool is scaled.
The central question is not “How much did AI help us produce?” It is “What became better for customers, the company, the team, and the people in it—and what did it cost elsewhere?”
Source and interpretation notes
The cases above mix a randomized task experiment (Microsoft), a survey/qualitative study (Anthropic), a research measurement framework (SPACE), an incident report (Spotify), a non-AI discovery case (Etsy), and vendor/customer-published before/after and financial accounts (Salesforce, Superhuman/AWS, IBM). They answer different questions; none alone demonstrates a causal, four-perspective return from AI-assisted engineering. DORA's 2025 report likewise describes AI as an amplifier of existing organizational strengths and weaknesses, not a substitute for local customer, financial, reliability, and people measures. Consult the linked primary publications before transferring a number to another workflow.