AI Moves the Starting Line
For teams using AI to make prototypes and drafts, widen who can start useful work before assuming you need fewer specialists. Measure the payoff in accepted results, not first versions.
A Number That Matters
20% — approximately the reduction in time Figma researchers report for successfully completed design tasks, averaged across three standardized exercises. The experiment recruited 50 product designers and 50 product managers. Product managers gained more. An average speedup was not an equal dividend.
The timing estimates exclude unsuccessful attempts. Participants edited specified reference designs; they did not take an uncertain customer problem through to a released service. Twenty percent is a useful result. It is not a staffing ratio.
This Week's Thesis
Our read: the opportunity is a change in who can begin. A product manager can attempt an interface change. A commercial leader can build a customer demonstration. The person who understands the problem can bring something testable to a specialist instead of joining a queue with a description.
Starting, judging and finishing remain different capabilities. Cheaper beginnings can help specialists—or send them more work. The first bet should be wider participation with a measured path to completion, not a smaller team justified by an average speedup.
The decision shifts from which tool makes an employee faster to which handoff should no longer require a wait. Does changing that handoff help the whole team deliver?
What Supports the Thesis
Figma researchers randomly assigned participants within each role to work with or without Figma Make on dark-mode conversion, menu modification and comment-flyout exercises. Product managers saw larger time gains; designers benefited on the more involved interaction task. That supports lowering a barrier to participation, not treating the professions as interchangeable.
Participants had at least two years of relevant experience and viewed or edited Figma files at least monthly. Moderators checked submissions against reference images; broader design-quality ratings were self-reported. The study did not establish independent expert approval of accessibility, originality or production readiness.
A September 25 OpenAI customer story describes a similar handoff. Proaction's non-technical COO says he builds customer-specific interactive demos with Codex that previously required engineering involvement. When a prospect becomes a customer, engineers receive the demo as a visual reference.
Avoiding engineering effort on a demo is one benefit. Clarifying customer needs before engineering starts is another. Proaction reports fewer questions and less back-and-forth, without a controlled downstream measurement. Our read: clearer problems could save work that faster implementation would merely execute sooner.
GitHub's September 25 review-metrics release offers a way to examine the rest of the journey. Repository reports separate ready-for-review to first review, first to final review, and final review to merge, with a median and 90th percentile for each. They count human reviews on qualifying human-authored pull requests, exclude bot reviews and have no historical backfill.
Outside software, the equivalents are proposal to specialist assessment, assessment to agreement, and agreement to something a customer can use. The sources suggest an operating hypothesis: change who prepares the first version, then follow what happens after the handoff. They do not establish a market-wide throughput gain.
What Challenges the Thesis
The strongest countercase is simpler: let people do familiar work faster and leave responsibilities intact. Redesign can be an expensive answer to a problem a useful tool already solved. Sometimes the right outcome is a calmer working day.
A 2023–2024 field experiment across 66 firms and 7,137 knowledge workers supports that case. Random access to Microsoft 365 Copilot reduced email time by about 1.4 hours per week across workers offered access; its instrumented estimate for adopters was about two hours. It detected no meaningful shifts in measured task quantity or composition. Some time returned as less out-of-hours work. These are historical tool results, not a verdict on September 2026 agents.
The current evidence also challenges our priority. Figma's designers benefited on a complex task, and Proaction reports engineering effort avoided. AI can improve or substitute for bounded specialist tasks. Where that effect dominates, helping existing specialists may matter more than widening participation.
More prototypes can also mean more commitments to unassessed ideas. A convincing demo may conceal integration work or be mistaken for a delivery promise. These are plausible risks, not incidents established by the study. Specialists could spend less time translating requests and more time rejecting polished but unsuitable ones.
We favor testing the handoff in prototype-led work, but the simpler route deserves equal treatment. If unchanged responsibilities deliver as much with less coordination, keep them. Organizational redesign is not a prize for using AI.
How Sure Are We?
Supported read: first versions and completed outcomes need separate measures. Our read, with moderate confidence: widening who can initiate useful work is a promising first experiment for prototype-led teams. Confidence is low that this generalizes across occupations or creates net capacity gains.
The evidence has material limits. Figma's September 22 preprint is vendor-authored. Its results describe a marginal aggregate designer time gain, while its discussion calls the aggregate reduction statistically insignificant. That inconsistency rules out a clean aggregate claim about designers; our basis is the larger product-manager gains and task-specific findings. Proaction is vendor-selected and customer-reported. GitHub's release is not causal evidence about AI. The Copilot study includes Microsoft researchers and measures application activity, not completed work's content or quality.
What would strengthen the read: independent field comparisons showing usable problem-owner prototypes, less specialist clarification, and more accepted work at unchanged quality without higher specialist effort per result. Count abandoned and rejected work too.
What would weaken it: longer specialist queues, more rework, lower acceptance quality or no end-to-end time saving. If existing specialists using AI outperform wider participation on those measures—or gains disappear when failed attempts count—we would change the recommendation.
What to Do With This
Choose one low-stakes workflow where a problem owner waits for a specialist to make an idea visible: a customer demo, interface proposal or draft service change. Keep specialist capacity, production authority and delivery commitments unchanged.
Compare similar work through two routes: a specialist using AI within the current handoff, and a problem owner preparing an AI-assisted prototype for the same specialist to assess. Set acceptance criteria first and include failed attempts and rejected ideas. A contemporaneous comparison better controls for changing workload and experience than before-and-after results.
Track preparation effort, assessment waiting time, clarification and rework, and elapsed time to acceptance. Record who initiated the work and specialist effort per accepted result. Apply the same quality check to both routes. Use existing timestamps where possible; prototype, prompt and licence counts are not outcomes.
Continue only for better accepted results, shorter delivery or less specialist effort without lower quality. If participation merely enlarges the queue, narrow the trial or restore the handoff. If AI simply returns time, value it rather than manufacturing demand to consume it. Do not turn a bounded trial into a permanent staffing assumption.
Change one handoff and follow the work to its end. Faster beginnings are easy to demonstrate. Deciding which deserve to become finished work is the management task.
One Question to Take With You
If more people can bring a working first version, does your organization get better problems to solve—or just more work to finish?