AI Development Workflow in 2026. From Prompt to Production.
The workflow has settled into a shape. Spec, plan, generate, verify, ship. The teams that skip verify are the ones still firefighting.
8 min read
- Cycle time per ticket
- 1-2 days
- from 3-5 days
- AI code reviewed
- Always
- Sometimes
- Rework rate
- Under 10%
- from 25-30%
Two years ago the question was whether AI could write usable code.
That question is closed.
The open question is what happens between the generated code and the merge button.
Teams that added a verification layer got faster.
Teams that removed review to go faster ended up slower, just later.
§01 · The problem
Where the time actually goes.
- 01Generated code merged without review
- 02No spec before the agent starts working
- 03Tests written by the same agent that wrote the code
- 04No record of which code was AI-generated
- 05Prompt changes untracked across the team
- 06Speed measured, quality assumed
AI does not remove the review step. It moves review from writing to reading, which most teams have not staffed for.
§02 · The mechanism
The AI workflow.
- 01
Ticket converted to a written spec
Generative AI - 02
Engineer approves the spec
Human - 03
Agent plans the change
Agentic AI - 04
Code and tests generated
Agentic AI - 05
Automated review before human review
AI Agent - 06
Human reviews the diff and merges
Human sign-off
A person closes the loop: the workflow ends on a human or a check against a defined outcome.
- Generative AI
- Model drafts something a human will edit
- Human
- A judgement call only a person makes
- Agentic AI
- Model plans and executes a multi-step task
- AI Agent
- Model extracts, classifies, matches, or routes
- Human sign-off
- Final approval before anything ships
§03 · The difference
Before vs after.
| Before | After |
|---|---|
| Ticket handed straight to the agent | Spec written and approved first |
| No spec, so scope drifts mid-task | Scope fixed before generation starts |
| Code generated and merged same day | Agent plan reviewed before execution |
| Review skipped because it looked fine | Automated checks run before human review |
| Bugs surface in the next sprint | Human reviews a bounded, documented diff |
| Nobody can trace what the agent changed | Every AI change traceable in the log |
§04 · The measure
Time saved.
| Task | Before | With AI |
|---|---|---|
| Spec creation | 3 hrs | 20 min |
| Implementation | 2 days | 4 hrs |
| Review cycle | 1-2 days | Same day |
| Rework rate | 25-30% | Under 10% |
Continue reading
Tell us what you’re building.
Share a few details about what you need, and we’ll get back to you with a quotation.