The Challenge
Two problems sit on either side of every development ticket. Upstream, tickets written by project managers describe outcomes in client language, which is correct for the client and insufficient for the developer. Downstream, estimates are produced from experience and instinct, which works until the estimator is unfamiliar with the codebase.
AI can plausibly fill both gaps, and that plausibility is the danger. An assistant with no access to the repository will confidently invent file paths, cite line numbers that do not exist, and produce an estimate with no evidence behind it. The output looks like rigor. It is not.
BRIO needed a workflow that could:
- Produce consistently structured tickets across a mixed portfolio of technologies.
- Ground technical detail in the client’s real code rather than in inference.
- Estimate work in a way that reflects AI-assisted implementation honestly.
- Fail loudly instead of guessing quietly.
The Approach
BRIO built a three-stage ticket lifecycle. Ticket creation captures the work in a structured format and runs a definition-of-ready check. Triage turns that ticket into an implementation brief anchored to specific files in the repository. Estimation then sizes the work against that brief. Each stage is a discrete step, so a ticket can move through the full sequence or stop wherever the evidence runs out.
The governing design decision is a hard requirement, not a preference: the triage and estimation stages check for a checked out repository and refuse to run without one. If no project root is detected, the workflow stops and says so, explaining that a fabricated technical description is worse than an empty field. That refusal is the feature that makes the rest of the output trustworthy.
The Solution
Structured Ticket Creation: Tickets are created against the firm’s full work item hierarchy, from workstream and initiative down through epic, feature, task, bug, and sub-task. A definition-of-ready check surfaces gaps before the ticket reaches a developer. Tickets can be posted directly or staged in a spreadsheet for review, which matters on engagements where a human approves every write.
Code-Anchored Triage: Triage reads the ticket, its comments, its acceptance criteria, and its history, then explores the repository and writes a technical brief into a dedicated Technical Description field. The project manager’s original description is never overwritten. When the ticket carries an unresolved architectural ambiguity, triage posts a handoff comment naming the specific blockers instead of papering over them.
Evidence-Based Estimation: Estimates are itemized into three lines, AI implementation time, human review time, and quality assurance testing time, each carrying a confidence rating traceable to something found in the ticket or the code. Separating the review line from the implementation line is what keeps AI-assisted estimates honest; the work does not disappear, it moves.
Continuity Between Sessions: Each stage checks whether the previous one already ran and reads any handoff comments left behind, so a ticket blocked last week can be picked up this week once a stakeholder has answered the open question. The audit trail lives on the ticket where the team already works.
Technical foundation
- Custom ticket fields for Technical Description and Acceptance Criteria
- Project root detection across Git, Salesforce DX, Node, Python, and Rust conventions
- Repository convention files read at runtime for project-specific standards
- Connector-based integration with the project tracker for reads and writes
- Applied across Salesforce and Apex, .NET, React, and WordPress codebases
The Impact
Anchoring the workflow to the repository changed the quality of what came out of it:
- Estimates carry an explicit confidence rating and cite the evidence behind every line.
- Review effort is estimated separately rather than absorbed silently into implementation.
- Tickets arrive at developers with file-level technical detail already attached.
- Ambiguities surface at triage, before they become mid-sprint discoveries.
The lesson generalizes past ticketing. AI becomes dependable in delivery work at exactly the point where it is given real access and permission to say no.
