When Code Is Cheaper, Delivery Still Has a Cost
Faster code generation changes delivery economics, but requirements, verification and operational outcomes still need attention.
Written as at 31 October 2026.
ai-engineeringsoftware-architectureproductivityai-native-sdlc
A coding agent can produce a convincing change before a team has agreed what the change should achieve. It can also produce tests, a pull request and a clear explanation. None of those, by itself, tells a service owner whether the change is safe, useful or ready to operate.
That does not make coding agents a poor investment. It changes the question worth asking. If a team generates software faster, where does the work wait next, and can the organisation recognise a good outcome when it arrives?
Faster generation is not free delivery
Illustrative scenario: a fictional composite written to make the point concrete. It does not describe a real organisation, and its figures are not reported outcomes.
A service team asks an agent to add an export feature for customer records. The agent produces a working endpoint, tests and documentation. A demo shows a successful download. The pull request is ready in minutes.
Then the harder questions begin. Which roles may export which records? What happens when a customer withdraws consent? Does the export include fields that are useful to the model but unnecessary for the recipient? Who investigates if a file is sent to the wrong person?
The answers involve product owners, security specialists, engineers and the people operating the service. The generated code is part of the work, but it has not made those decisions disappear. Nor is its marginal cost zero: model usage, compute, review, testing, integration and operation all consume resources.
Find the constraint before declaring it
It is tempting to announce that the bottleneck has moved from coding to review. For one team that may be true. Another may wait on data access, a vendor dependency, uncertain requirements, deployment approval or an operational handover. The constraint can change with the service and over time.
Track a change from request to an accepted outcome. Record time spent waiting as well as time spent working. Look for queues, repeat defects and decisions that routinely return for clarification. Then improve the constraint the evidence shows, rather than optimising generated lines or pull-request counts in isolation.
Evidence about AI productivity is not uniform. In a randomised study of 16 experienced open-source developers working on 246 tasks in familiar, mature repositories during early 2025, METR found that allowing the then-current AI tools increased completion time by 19 per cent. This is evidence for that setting, not a prediction for every team or today’s tools. In a February 2026 update, METR said selection effects made its later experiment an unreliable measure of the current productivity effect and that its evidence was weak on the size of any speed-up. [1, 2]
DORA’s 2025 research describes AI as an amplifier of organisational strengths and weaknesses. That is a useful prompt to examine delivery practices alongside tool adoption, not a guarantee of a particular outcome. [3]
Make acceptance observable
Before asking an agent to change the export service, a team can write down the outcome and its boundaries: authorised users can request a defined set of fields; an unauthorised request is denied; access and transfer are recorded; and the service has an agreed response if delivery fails.
Those examples are not a complete security specification. They illustrate how a product requirement becomes something a team can test and review. A passing test provides evidence about the conditions it checks; it does not prove the requirements are complete or the system is safe in every environment.
Human review should follow the risk and the uncertainty of the change. A small formatting change and a change to payment authorisation do not require identical scrutiny. Teams can automate repeatable checks and still retain accountable people for decisions that exceed the evidence or authority available to the workflow. An automated reviewer is another source of evidence, not an automatic substitute for a service owner’s judgement.
Give engineers a path to develop judgement
The work of understanding users, designing interfaces, checking failure modes and operating services does not belong to one job title. Engineers need technical depth and opportunities to practise it alongside product and domain discovery.
That includes early-career engineers. If an agent produces the first draft, a junior engineer can still investigate the requirement, explain the design, add a missing failure test and follow the change into production with a mentor. The point is not to manufacture busywork. It is to make the learning previously embedded in small tasks deliberate and visible.
Measure the result the service accepts
Use a small set of measures together: lead time from request to accepted change; time waiting for review and dependencies; escaped defects and recovery; and whether the change achieved its intended user outcome. Segment by change type and risk. A shorter review queue means little if rework or production incidents rise elsewhere.
Compare similar work over a stated period, and treat small samples cautiously. No single metric captures developer productivity, and model-reviewer agreement or comment acceptance is not proof of correctness. Pair measures with the experience of people who build, review and operate the service.
The Contextual Agentic Enterprise is Contextual Agentic’s framework for connecting context, authority, evidence and outcomes across people and systems. The same questions are useful in AI-assisted delivery: what was the change intended to do, who could approve its effects, and what evidence survives when it runs?
A useful next step is to trace one recent change from idea to operation. Find its longest wait, identify the decisions that depended on tacit knowledge, and choose one way to improve the evidence available at that point. Then measure what happens to the whole service, not just the speed of the next generated diff.
Application requires organisation-specific architecture, security, privacy, legal, regulatory, operational, safety, accessibility, commercial and governance review, together with appropriate approvals.
Sources
- Becker et al. (2025), Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity: https://arxiv.org/abs/2507.09089
- METR (24 February 2026), We are Changing our Developer Productivity Experiment Design: https://metr.org/blog/2026-02-24-uplift-update/
- DORA (2025), State of AI-assisted Software Development: https://dora.dev/research/2025/dora-report/
For more on how context and verification change software delivery, read The AI-Native SDLC: What Moves When Agents Write the Code. The book The Contextual Agentic Enterprise explores how context, authority and evidence connect to enterprise workflows.
Related
The AI-Native SDLC: What Moves When Agents Write the Code
When agents produce most of the code, the bottleneck moves from typing to context and verification. Five things shift: the artifact, the review, the tests, the roles and the bottleneck, and the evidence for each is thinner than the slogans suggest.
When Code Becomes Cheap, Design Becomes Critical
Coding agents answer the hard design questions by default, and differently each time. A letter on what an architecture review of forty generated adapters should ask.
Enterprise Agentic AI Adoption: Why Contextual and Agentic Have to Arrive Together
Most adoption programmes buy one half of the name. The agentic half without the contextual half is a fast demo that cannot be let near a system of record; the contextual half without the agentic half is knowledge management with a chat window. The two are one requirement, and the reason is in the nature of the machine.