Engineering Output Was Never the Bottleneck

Engineering teams now ship more code per person than they did a year ago, but the release dates have not moved. Whether an AI coding assistant changes anything past the individual keyboard is set by one decision: whether the workflow and the measure that judges it were redesigned, or left exactly as they were.

More Output Was Never the Shortage

The tool works at the task. ANZ Bank ran a controlled experiment with GitHub Copilot across a cohort of about 100 engineers over several weeks [1]. The Copilot group completed programming tasks roughly 42% faster than the control group, and the most experienced engineers gained the most [1].

That number measures one engineer finishing one set task, in a controlled setting, not a programme reaching customers [1]. An organisation does not sell tasks. It sells working software, decisions made, and controls that hold.

AI Amplifies the System You Already Have

Adoption is now near-universal. DORA, the DevOps Research and Assessment programme (not the European Union's Digital Operational Resilience Act of the same initials), reports in 2025 that about 90% of technology professionals use AI in their work, and more than 80% report it has increased their productivity [2]. The individual gain is real and widespread.

That gain is a fraction of the delivery cycle. Bain & Company's 2025 technology report finds teams using AI assistants see roughly 10 to 15% productivity boosts, but writing and testing code is only about 25 to 35% of the time from initial idea to launch, so "speeding up these steps does little to reduce time to market if others remain bottlenecked" [3]. The companies that report about 25 to 30% boosts are the ones pairing generative AI with end-to-end process transformation, with Goldman Sachs fine-tuning models on its own codebase and Netflix moving testing and quality checks earlier in the process [3].

DORA reaches the same place from the delivery side. Its headline for 2025 is that "AI's primary role is as an amplifier, magnifying an organization's existing strengths and weaknesses" and that "the greatest returns come not from the tools themselves, but from a strategic focus on the underlying organizational system" [2]. Unlike the year before, the report now observes a positive relationship between AI adoption and delivery throughput, while the relationship with delivery stability remains negative [2].

The mechanism is the spine of the whole argument. In DORA's words, "AI accelerates software development, but that acceleration can expose weaknesses downstream. Without robust control systems, like strong automated testing, mature version control practices, and fast feedback loops, an increase in change volume leads to instability" [2]. Teams in loosely coupled architectures with fast feedback loops see gains; those in tightly coupled systems and slow processes see little or none [2].

A faster keyboard speeds the coding step, but it does nothing for the human-heavy stages around it: the back-and-forth of requirements gathering and the sign-off of user-acceptance testing (UAT) take the same time they always did. Delivery stays gated by those stages until they are redesigned, which is why writing and testing code being only a fraction of time to market is the binding point, not a footnote.

Individual output rises almost everywhere. Whether it converts to delivered value depends on the surrounding system, and AI magnifies the operating model you already have.

Nubank Rebuilt the Sequence, Not Just the Toolkit

Nubank has told its investors this in its own words. In its Q1 2026 results, reported on 14 May 2026, the Brazilian digital bank split its AI work into three phases: AI Assistance, meaning individual and company productivity; Workflow Reinvention, meaning customer journeys rebuilt end to end; and AI-Native [4]. Chief executive David Vélez put the intent plainly: "We are not adding AI to banking, we are rebuilding banking around AI." [4]

The sequencing is the lesson, not the headline number. Assistance is "now largely complete", and inside that phase Nubank reports engineering throughput up 50% year on year and testing cycles 90% faster [4]. While those figures are self-reported, they give a clear view of the direction and the benefits Nubank is seeing [4].

Workflow Reinvention, the phase that actually rebuilds the work, is described as still in motion, with new customer experiences only expected to reach customers during 2026 and no delivery metric attached [4]. Even the most AI-native bank has banked the cheap part first and is only now doing the hard part.

Output Becomes Value Only When It Has Somewhere to Go

A model output is activity until something downstream changes because of it. Code, a document, a dashboard, an alert, a summary: each is cost until a decision, a workflow, a control, a role, or a customer outcome moves, and the metric that judges the work moves at the same time.

This is why distributing the tool and leaving the measure on individual velocity optimises for output that has nowhere to go. DORA names the same mechanism from the delivery side: the returns show up where automated testing, mature version control, and fast feedback loops are already strong, and those are properties of the operating model, not the tool [2].

What Teams Should Change First

Stop measuring individual output and start measuring end-to-end delivery. Lines shipped, tickets closed, and personal velocity all rise the moment the tool arrives, and none of them tells you whether anything reaches a customer sooner or a decision gets made better.

Pick one workflow and redesign it end to end before scaling the tool wider. Change the metric that judges it in the same move, so the work is measured against what it now produces rather than the job it used to be.

Take a typical delivery pipeline. The coding assistant lifts code and pull-request volume, but the bottleneck sits downstream in code review and release approval, so the extra output piles up in the queue. The redesign is to keep batches small, automate the tests and the release gate so review is not the choke point, and change the team's headline metric from pull requests merged to lead time to production. Nothing reaches a customer faster until the constraint moves, not the keyboard.

Above all, do not tell people to use AI while judging them on the old job. That instruction guarantees the outcome nobody wanted: more output, the same delivery date, and a leadership team convinced the transformation already happened.

Key Takeaways

• Retire individual-output metrics for AI-assisted teams and measure end-to-end delivery instead; faster keystrokes without a shorter release date is cost, not value.
• Redesign one workflow completely, and change its metric at the same time, before buying more seats; the conversion from output to value is a property of the workflow, not the tool.
• Treat "use AI" issued against unchanged targets as a governance gap, not a productivity win, and inspect batch size, testing discipline, and feedback loops as the leading indicators of whether output is converting.
• The real gap is not tooling but the operating model around it. The full case for treating AI transformation as an operating-model redesign, not tool adoption, is set out in From Digital Transformation to Cognitive Transformation. [5]

Sources

[1] Chatterjee, S., Liu, Y., et al. "The Impact of AI Tool on Engineering at ANZ Bank: An Empirical Study on GitHub Copilot within Corporate Environment." arXiv:2402.05636, February 2024. https://arxiv.org/abs/2402.05636

[2] DORA / Google Cloud. "State of AI-assisted Software Development" (2025 DORA report). DORA / Google Cloud, 23 September 2025. https://dora.dev/dora-report-2025/

[3] Bain & Company. "From Pilots to Payoff: Generative AI in Software Development." Technology Report 2025, 23 September 2025. https://www.bain.com/insights/from-pilots-to-payoff-generative-ai-in-software-development-technology-report-2025/

[4] Nu Holdings Ltd. "Nu Holdings Ltd. Reports First Quarter 2026 Financial Results." Businesswire, 14 May 2026. https://www.businesswire.com/news/home/20260514553909/en/Nu-Holdings-Ltd.-Reports-First-Quarter-2026-Financial-Results

[5] Hughes, Christopher. "From Digital Transformation to Cognitive Transformation." cgh.dev, 2026. https://cgh.dev/thinking/cognitive-transformation/