From Prototype to Product: A Twelve-Month Field Study of Generative Design Adoption Across Forty Product Teams

1.Introduction

Generative tooling is almost always justified on iteration speed: more concepts, produced faster, at lower marginal cost. The premise is well supported at the level of the individual designer. What remains poorly evidenced is the organisational consequence — whether faster concept production shortens the path to a shipped product, or simply moves the bottleneck to a stage that was already constrained.

This distinction matters because the two outcomes imply opposite investment decisions. If generation shortens end-to-end cycle time, the rational response is to buy more generation capacity. If generation relocates work to review and validation, the rational response is to staff and fund review before expanding generation, and the failure to do so will present as unexplained senior-engineer overload rather than as a tooling problem.

We therefore ask three questions. RQ1: How does generative design adoption change effort distribution across concept, review and validation stages? RQ2: Does it change end-to-end cycle time from concept commitment to release? RQ3: Which organisational conditions distinguish teams that retain the iteration gain from those that do not?

2.Related Work

Studies of computer-aided design and, later, automated code generation repeatedly describe a pattern in which local productivity gains fail to propagate to delivery outcomes because a downstream verification stage absorbs them. The literature on this phenomenon is largely retrospective and cross-sectional, which makes it difficult to separate absorption from selection effects among adopting teams.

Adjacent work on iteration economics argues that the value of additional concepts declines sharply once review capacity is saturated, predicting exactly the rejection-rate increase we observe. Our contribution is longitudinal field evidence with pre-adoption baselines, and identification of a specific, actionable organisational moderator — pre-adoption assignment of validation ownership — rather than a general claim that culture matters.

3.Methods

Setting and sampling. Six organisations agreed to twelve months of observation spanning their generative design rollouts: two enterprise software firms, two consumer hardware manufacturers, one medical-device developer and one industrial equipment supplier. Within them, forty teams were purposively sampled for variation in product risk class, team size (4–19 members) and prior tooling maturity. No team was selected on the basis of expected outcome.

Data collection. Each team was interviewed at months 0, 3, 6, 9 and 12 using a structured protocol covering effort allocation, review practice, artefact flow and perceived quality, yielding 214 interviews. For 31 teams we obtained workflow telemetry from issue trackers and design systems, giving objective concept counts, review durations and rework events. We additionally collected design-review minutes and rejection records for all forty teams.

Analysis. Interviews were coded by two researchers using a codebook developed on two pilot teams and refined until agreement reached κ = 0.79. Telemetry was analysed as an interrupted time series against each team's own twelve-month pre-adoption baseline, which controls for team-level differences in product complexity. Cross-case pattern matching identified organisational conditions co-occurring with gain retention; candidate moderators were required to be observable before adoption to be admitted, in order to reduce reverse-causal readings.

4.Findings

Effort relocation. Concept iteration time fell in 36 of 40 teams, with a median reduction of 38% (IQR 24–49%). Review and validation effort rose in 33 teams, median 21% (IQR 9–34%). Aggregate end-to-end cycle time changed by −4%, which our confidence interval cannot distinguish from no change.

Who absorbed the load. In the 22 teams without a named validation owner, the additional review work concentrated on a median of 1.5 senior engineers per team, who reported it as informal, unscheduled and largely invisible to planning. Three of these teams later reported attrition or reassignment of the individual carrying that load. In the 18 teams that named an owner before adoption, review work was scheduled, distributed across a median of 4 reviewers, and retained 71% of the iteration gain in end-to-end cycle time against 12% for the remainder.

Quality gate behaviour. Rejection at design review rose from a baseline of 0.14 to 0.29. Interview data consistently attributed this to a presentation effect: generated candidates arrived visually and structurally complete, so reviewers could no longer use surface incompleteness as a cheap early filter and instead carried more candidates into detailed evaluation before rejecting them.

Risk class interaction. The medical-device and industrial teams, which operate mandatory validation gates, showed the smallest iteration gains but also the smallest informal-absorption effect, because their review capacity was already explicit and funded. Regulated process, in this reading, functioned as an accidental protection against the failure mode we document.

5.Discussion and Implications

The dominant practical implication is that generative design should be procured as a change to the whole development system, not as a tool for the concept stage. Budgeting the licence without budgeting review capacity produced, in our sample, no delivery improvement and a concentrated increase in senior-engineer workload — an outcome that is difficult to detect in tooling metrics and easy to misdiagnose as individual performance.

The rejection-rate finding also has a measurement consequence. Organisations that track review throughput will see it deteriorate after adoption even where the change is beneficial, because each review now examines a more plausible artefact. We recommend tracking rejections per shipped feature alongside rejections per review to avoid penalising the gate for doing more work.

Finally, the moderator we identify is inexpensive. Naming a validation owner is an organisational act, not a capital expenditure, and in our data it separated teams that converted iteration speed into delivery speed from teams that did not.

6.Limitations

Forty teams in six organisations cannot support statistical generalisation, and our design is observational: organisations that named validation owners may differ in unmeasured ways from those that did not, despite our restriction to pre-adoption observables. Telemetry was available for 31 of 40 teams, so effort estimates for the remaining nine rest on self-report and review records.

Twelve months is short relative to hardware and device development cycles; three teams had not completed a full release cycle at study close and contribute to stage-level but not end-to-end results. Tooling in this category changed materially during the observation window, so absolute magnitudes should be read as period-specific.

7.Conclusion

Generative design reliably accelerated concept iteration across forty product teams and reliably failed to accelerate delivery unless downstream validation was owned and resourced in advance. Effort moved from concept to review; it did not disappear. Organisations planning adoption should fund review capacity in the same decision that funds generation, name a validation owner before rollout, and expect design-review rejection rates to rise for reasons that indicate a working gate rather than a failing one.

8.References

1. [1] Ng, A. (2025). Iteration economics in digital product teams. International Journal of Product Innovation 9(2), 88–104.
2. [2] Ferreira, L. & Vaz, T. (2026). Validation ownership as a moderator of tooling gains. JIPCET 1(1), 55–72.
3. [3] Vaz, T., Ng, A. & Ferreira, L. (2025). Longitudinal designs for design-tool evaluation. Design Studies 96, 101–124.
4. [4] Hale, E. (2024). Verification bottlenecks in engineering workflows. Research in Engineering Design 35(3), 271–290.
5. [5] Paxton, D. (2026). Measuring automation at the level of delivered work. JIPCET 3(1), 5–24.
6. [6] Beaumont, H. & Whitmore, O. (2025). Surface completeness and reviewer judgement. Human–Computer Interaction 40(4), 388–412.