AI Can Write Code. What Still Goes Into Development?

Generated Code Is a Useful Starting Point

Many agency projects now start with a design, an exported HTML file, or an AI-generated prototype. That gives the team something concrete to evaluate and can reduce the time needed to explore an initial direction.

The prototype usually represents one version of the interface under one set of assumptions. Production software also has to work with real content, real users, different devices, changing data, existing applications, and business rules that may not have been included in the original prompt.

This difference matters when estimating work. AI may reduce the time needed to create a first draft, while the estimate still needs to cover the path from that draft to a reliable result. Making both parts visible helps clients and developers agree on where the tool saves time and where implementation or verification is still needed.

AI can accelerate the first draft. Development turns that draft into a dependable feature.

What the Research Actually Says About AI Productivity

The honest answer is not that AI always saves time or that it never does. The result depends on the task, the developer, the codebase, the tool, and how the team reviews and delivers the work.

In GitHub's controlled study of 95 professional developers, the group using Copilot completed a defined JavaScript HTTP server task 55 percent faster. A larger set of field experiments at Microsoft, Accenture, and another Fortune 100 company found a 26.08 percent increase in completed tasks across 4,867 developers, with larger gains among less-experienced developers.

A different randomized study from METR followed 16 experienced open-source developers working on real issues in mature repositories. With early-2025 AI tools, they took 19 percent longer, even though they believed the tools had made them faster. METR later reported that late-2025 public agents produced small productivity benefits of roughly 4 to 20 percent in its newer work, while warning that selection effects made the size uncertain.

These findings are not contradictions that need one winner. A focused greenfield task is different from changing a system someone has maintained for years. A developer who needs help with unfamiliar syntax may gain more than an expert who must repeatedly correct suggestions that do not understand the repository's history. The useful conclusion is that AI productivity should be measured in the actual workflow, not assumed from a demo.

• AI is often strong at boilerplate, first drafts, repetitive transformations, test scaffolds, documentation, and exploring unfamiliar APIs.
• The size of the advantage can vary when a task depends on deep repository context, unwritten business rules, careful judgment, or many connected systems.
• Perceived speed is not always actual speed. Prompting, reading, correcting, and verifying are part of the task even when they are less visible than typing.
• The relevant measure is not lines of code generated. It is useful work completed without increasing defects, rework, security risk, or maintenance cost.

Why an HTML File Is Not a Finished Shopify Page

Shopify is a good real-world example. An HTML mockup can show headings, images, buttons, and a layout. A live Shopify section has to become part of the theme's architecture. Shopify uses Liquid to generate markup, JSON for configurable templates, sections and blocks for merchant editing, plus CSS, JavaScript, locale files, app blocks, and dynamic data sources.

The developer has to decide what should be editable, which values come from the product or a metafield, how the section behaves when content is missing, and whether it can be reused across products. A purchase interface also has to respect variants, availability, quantity rules, discounts, selling plans, cart behavior, and any apps already changing the product page.

Then the page has to work beyond the screenshot. It needs responsive behavior, keyboard access, visible focus states, valid controls, image sizing, performance checks, translation support, and testing in the actual theme. Shopify's own documentation treats performance and accessibility as work that must be implemented and tested. None of that is contained in a static HTML export.

Claude, ChatGPT, Copilot, or another coding assistant can help write Liquid, schema, CSS, JavaScript, and test cases. The developer supplies the store context, chooses the architecture, reviews the output, tests the interactions, and resolves conflicts. The combination can create meaningful time savings while preserving the work needed for a production-ready section.

Speed and Quality Need the Same Workflow

Using AI does not automatically improve or reduce quality. The outcome depends on whether faster generation is matched by clear requirements, review, testing, and feedback.

DORA's research describes AI as an amplifier of the system around it. Teams with clear standards, small changes, automated tests, fast feedback, and healthy review practices can turn generated code into useful output. Teams with unclear ownership and weak quality controls can generate more change than they are able to understand or safely release.

The 2025 Stack Overflow Developer Survey shows why verification remains necessary. More developers distrusted the accuracy of AI output than trusted it. Sixty-six percent said their biggest frustration was receiving solutions that were almost right, and 45 percent said debugging AI-generated code took more time.

Almost-right code requires careful review because it can look plausible. It may pass a quick visual check while failing for a different product, screen size, user role, language, browser, or data state. Including verification in the workflow helps teams catch those gaps before they become rework or production issues.

That review load is also where this discussion meets AI fatigue. When generation becomes faster but every suggestion still requires a decision, the work can shift toward continuous checking and context switching. The answer is not to blame the tool or the developer. It is to give review clear boundaries and include it in the delivery plan.

How AI Is Changing Project Expectations

AI has changed expectations faster than many organizations have updated their delivery process. It is reasonable for clients and managers to expect efficiency gains. The challenge is identifying which parts of the project become faster and which responsibilities remain with the delivery team.

If an estimate includes discovery, integration, testing, review, and launch checks, the agency or developer should explain those items instead of presenting one unexplained number. The client can then see which tasks AI assists and which tasks protect the final result.

Reviewing generated code also requires enough knowledge to recognize subtle mistakes. DORA recommends dedicated work time for learning AI tools, clear acceptable-use policies, and stronger feedback loops. These practices help developers use the tools more effectively while giving organizations a more reliable way to measure the benefit.

For agencies, the strongest position is transparency. AI can be part of the workflow without becoming the only basis for a quotation or timeline. The goal is a clearer process and a better result, not simply a larger volume of generated code.

Production Work Still Needs a Human Owner

Code can be syntactically correct and still expose data, weaken authentication, mishandle permissions, trust unsafe input, or use a dependency incorrectly. For that reason, every production change needs a clearly identified person or team responsible for reviewing and approving it.

OWASP's 2025 guidance on inappropriate trust in AI-generated code says developers should understand all code they submit and review AI-assisted work for vulnerabilities using human review and security tooling. That is the same standard teams should apply to copied snippets, outsourced code, and human-written code. AI changes the source of the draft, not the need for accountability.

Responsible use also includes protecting private source code, customer data, credentials, and internal documentation. Teams need to know which tools are approved, what data may be shared, what permissions an agent receives, and who reviews its actions before it can affect production.

What AI Should Change in a Project Estimate

A good estimate should reflect the real workflow. If AI reliably shortens a known task, the estimate can improve. It should not be reduced by an arbitrary percentage just because an assistant is available.

The team first needs to define what done means. For a web feature, that may include approved responsive layouts, real content, browser and device checks, keyboard behavior, analytics, integration tests, performance limits, error states, documentation, deployment, and a rollback plan. AI can help with several of those items, but the acceptance criteria remain.

The fairest productivity measures look at cycle time, escaped defects, review time, rework, user outcomes, and maintainability. Counting prompts, generated lines, or hours spent typing rewards activity rather than value.

• Estimate from the requirement and definition of done, not from the speed of the first generated draft.
• Separate implementation time from integration, review, QA, deployment, and project communication.
• Use smaller releases so AI-assisted changes are easier to understand, test, and reverse.
• Track where AI actually saves time in your own projects, then update future estimates from evidence.
• Keep a named human owner for architecture, security, quality, and launch approval.

A More Useful Conversation About AI

Clients are right to ask whether modern tools can reduce time and cost. Agencies and developers should be able to explain where they help. The healthier conversation is specific: Which tasks are faster? Which checks are still required? What risks are being managed? What result is the team accepting responsibility for?

Both sides benefit when ordinary work is explained plainly and an impressive demo is treated as evidence of possibility rather than a complete project estimate. The scope, assumptions, acceptance criteria, and review process should remain visible.

AI is already part of professional development, and its capabilities will continue to change. The durable advantage is not pretending the tool does nothing or everything. It is knowing what can be delegated, what must be verified, and where human judgment creates the value the client is actually paying for.

The Goal Is Better Software, Not More Code

AI can make a capable developer faster, especially when the team already has clear standards and reliable feedback loops.

The best teams use AI to reduce repetitive effort and create more room for architecture, accessibility, performance, testing, communication, and the business problem behind the code. They do not remove those responsibilities from the estimate merely because the first draft arrived quickly.

A generated page is a starting point. A production feature is a promise that the result works, can be maintained, and was checked by someone willing to be accountable for it. That difference is still development.