"When we requested an estimate, the man-hours came back at half of what they used to be. While that's great, we can't tell whether it's actually reliable"—we hear this uncertainty often from clients outsourcing software development. Behind this lies a fundamental shift in how development is carried out. The assumption that humans write every single line is breaking down, and "AI-first development," where AI agents draft requirements and handle parts of implementation, has become a realistic option.
A prime example is the evolution of Jira, an issue-tracking tool used by many development teams. It has rolled out functionality where AI automatically drafts requirements definitions, assigns the broken-down tasks to AI coding agents like Claude Code and GitHub Copilot, and produces review-ready pull requests. While we previously discussed key insights clients should grasp before comparing tools in Why You Shouldn't Start with No-Code/Low-Code in the AI Agent Era, this article focuses squarely on how those commissioning custom development should navigate this shift.
What changed: Non-humans appearing among "assignees"
Until now, development was predicated on humans reading requirements, humans designing, and humans writing code. Jira's new workflow replaces parts of that sequence.
- AI drafts requirements definitions: From rough requests, AI identifies and breaks down the list of necessary tasks.
- AI agents can be assigned to tasks: Claude and others appear in the assignee options, summonable via comments or assigned through automated rules.
- Agents create pull requests end-to-end: Reading acceptance criteria and target repositories, they modify code in isolated environments and submit changes in a reviewable format.
Supported tools include Claude Code, Cursor, and GitHub Copilot, with OpenAI Codex support announced as well. The key point is that the center of gravity has begun shifting from "AI helping with what humans write" to "humans reviewing what AI writes." The actual loop design used by development teams is also covered in Development Workflows Using Claude Code.
What is promising and what is risky for the commissioning side
This transformation brings both advantages and risks to clients.
| Dimension | Advantages | Points of caution |
|---|---|---|
| Speed | Initial implementation accelerates, producing prototypes faster | "Fast" does not mean "done"; verification steps do not decrease |
| Cost | Potential to reduce man-hours for routine tasks | A lower estimate does not guarantee that quality assurance is included |
| Requirement | Drafts appear quickly, making it easier to start discussions | AI drafts are merely starting points; humans must guarantee correctness |
A major pitfall is that when AI drafts requirements, articulate-looking documents appear so quickly that teams may push forward without properly scrutinizing the substance. When implementation begins on ambiguous requirements, rework costs actually increase. The approach to commissioning without over-engineering was outlined in Scope Design to Avoid Over-Engineering; even in an era where AI generates drafts, deciding "what not to build" remains a human responsibility.
Three things clients must keep a firm grip on in AI-first development
Even as methodologies evolve, the client's scope of responsibility does not vanish. In fact, the following three points become more critical than ever.
- Deliver crystal-clear acceptance criteria: AI agents operate by reading acceptance criteria and design standards. If criteria are vague, deliverables will be vague. Concretely verbalizing "what must be achieved to consider it complete" directly governs quality.
- Do not underestimate reviews: The faster and more abundantly AI produces code, the more valuable human review becomes. Rather than assuming "it's fine because AI built it," make who verified it and how a condition of accepting deliverables.
- Choose partners who can articulate what AI handled and what humans owned: Is the vendor merely selling "cheaper and faster with AI"? Judge partners by whether they can explain the scope delegated to AI and how they assured overall quality.
In short, what clients must evaluate is not "whether AI is being used," but who takes responsibility for final quality when AI is used. Commissioning work based solely on speed and low cost when this responsibility is murky is dangerous.
Test small to gauge what to delegate
Rather than commissioning a mission-critical system on an AI-first premise from day one, testing on a small scale is safer. Start with low-impact features or internal tools to run through a complete cycle from requirements drafting through implementation and review, evaluating the quality of deliverables and the thoroughness of explanations. Once confidence is established, expand the scope of delegation. This "test small and scale up" approach is a core best practice across all forms of outsourcing, not just AI.
Start by writing a one-line "definition of done" for your next order
AI-first development does not take work away from clients. What gets eliminated is routine manual effort; what remains is the work of deciding "what to build" and "what constitutes completion." The next time you outsource something, try writing a single sentence in your own words before drafting a requirements doc: "When this is accomplished, it is done." The clarity of that single line directly shapes deliverable quality, even in AI-driven development.
If you need guidance on commissioning AI-driven development, validating estimate reasonableness, or verifying quality assurance workflows, feel free to reach out via GleamHub's Development, AI, and Automation Consultation. From verbalizing requirements to determining where to leverage AI and where humans must provide guarantees, we work alongside you tailored to your project.
Sources
- Atlassian Jira Adds Features for Automatic AI Requirements Drafting, Task Assignment to AI Agents like Claude and Copilot | Publickey
- Introducing Claude Agent for Jira - Inside Atlassian
- Atlassian Extends AI Reach of Jira Into Agentic Engineering Workflows - DevOps.com
- Work with AI agents in Jira | Jira Cloud | Atlassian Support
