AI coding in client work, honestly.

AI coding in client work, honestly.

We are a team of two or three people shipping production sites with AI assistance daily. It helps enormously. It also creates a specific new bottleneck that vendors rarely mention.

Start with the numbers we trust. The 2025 Stack Overflow Developer Survey found 16.3 percent of developers using AI tools reported increased productivity, while 41.4 percent reported little impact. The Faros and DORA AI productivity report found teams using AI heavily completed 21 percent more tasks and merged 98 percent more pull requests, but review times jumped 91 percent with no faster delivery overall. Gartner predicts 90 percent of developers will use AI coding assistants by 2028, up from 14 percent in 2024.

More output, slower review, same ship date. That matches our experience exactly.

Where AI actually helps us

Three tasks, consistently.

Understanding code. IBM's CHI 2025 study of its internal assistant (669 developers surveyed, 15 in usability testing) found comprehension topped generation as the primary use case. Same here. Pointing an assistant at an unfamiliar codebase and asking for the data flow beats an afternoon of file spelunking. Summaries of diffs before review are similarly useful.

Boilerplate and scaffolding. Routes, schemas, query wiring, form states, repetitive variants. A Sanity schema with ten document types and eighteen typed queries is exactly the kind of structured repetition where generation earns its keep, provided TypeGen verifies every line after.

Prototyping direction. First pass of a component, three layout options in an hour, throwaway spikes to test an idea. Speed of exploration matters more than polish at this stage.

Where it slows us down

Debugging complex runtime issues, security sensitive code, and reviewing generated output. Communications of the ACM summarized the divide in July 2026: acceleration on boilerplate and prototypes, deceleration on debugging, security and review of AI written code.

The mechanism is simple. Generated code arrives fast and confident. Every line still needs reading. A 400 line component produced in a minute takes twenty minutes to review properly, and skipping that review is how subtle bugs ship. The ICSE 2026 paper on developer perspectives put the open question directly: does AI generated code need more, less, or different review, and can reviewers even tell which lines are generated. Our answer after a year: it needs different review, slower and more skeptical, especially around data fetching, auth boundaries and caching.

Our rules for directed work

We run AI like a fast junior with infinite stamina and no judgment. Useful with supervision, dangerous without it.

1. Small diffs only. One component or one query per generation. Large generated rewrites get rejected regardless of quality because they cannot be reviewed well.
2. Types verify everything. Sanity TypeGen plus TypeScript strict plus lint plus build on every change. If the assistant breaks a query type, the pipeline catches it before a human debates it.
3. No generated code touches auth, payments or access control without a second human reading every line. This rule has no exceptions.
4. The author owns every line. IBM's study named it shared responsibility between people and systems. In our studio the split is simpler: if your name is on the commit, you tested it. "The model wrote it" is never an explanation in a postmortem.
5. Rework rate is the metric. Not lines generated. We track review burden, bugs per release and confidence after deploy. When rework climbs, generation scope shrinks.

The workflow in practice

A typical 0on0 task runs: human writes the brief in the ticket, assistant drafts the scaffold, human narrows and rewrites the interesting parts, types and tests verify, human reviews the diff line by line, small commit ships. The assistant never opens a pull request alone. It never merges.

Total time saved on straightforward builds: roughly a third, mostly in scaffolding and lookup. Total time saved on hard debugging: near zero, sometimes negative once bad suggestions are counted. Blended across a project: meaningful but not miraculous. Anyone promising tenfold output is selling the demo, not the maintenance.

What stays human

Taste, scope and refusal. The decisions that matter are what to build, what to delete, and when the page is done. No model has ever told us a section was unnecessary. Every deletion on our site came from a person arguing with another person. That arguing is the studio. Automation handles the typing. Humans keep the judgment.

Takeaways

• Expect the documented split: more tasks started, much longer review, delivery speed unchanged unless review is redesigned.
• Generate scaffolding and summaries. Hand write security boundaries and data flow.
• Enforce small diffs, type verification and human ownership of every commit.
• Measure rework rate and review burden, not generation speed.

Sources: Stack Overflow Developer Survey 2025; Faros AI and DORA AI Productivity Paradox Report via CACM July 2026; Weisz et al., Examining the Use and Impact of an AI Code Assistant, CHI 2025 and arXiv 2412.06603; Beyond the Commit: Developer Perspectives on Productivity with AI Coding Assistants, ICSE SEIP 2026; The Impact of LLM Assistants on Developer Productivity systematic review, ACM TOSEM 2026; CACM How to Cross the AI Code Productivity Divide, July 2026; Gartner AI coding adoption forecast via CACM.