Own system-development project | Robert James | Published October 8, 2026
A content team can delegate drafting and still leave its senior reviewer with a full reread of every article. I designed an AI-assisted QA workflow around that bottleneck: apply safe routine fixes, surface unresolved decisions, and check the exact final version before delivery. This case describes system design and scenario testing, with production results still to be measured.
Project type: Own system-development and R&D project
Scope: Client rules, issue states, human review, and final-version checks
Testing: 55 defined cases across three versioned records
Status: Tested design; production performance unproven
The starting problem
Generic checks cannot fully account for a client's audience, approved sources, terminology, search brief, factual risk, or recurring feedback. A useful QA process needs those requirements before it can judge whether a draft is ready for delivery.
The design decision
The workflow starts with a client profile and article brief. It keeps detection, proposed fixes, applied changes, and successful rechecks separate. A recommendation is not a completed edit, and a clean check on an earlier draft cannot establish the quality of a later export.
How the workflow operates
The sequence is to inspect the draft, apply safe fixes, recheck the changed version, collect human decisions, apply those decisions, and check again. The final artifact is then frozen before export. Claims, sources, firsthand experience, and higher-risk content remain decisions that need appropriate evidence or review.
The reported testing scope
The testing program contains 5 founder-validation cases, 16 adversarial QA scenarios, and 34 security or red-team cases across three versioned records. These 55 defined cases describe testing scope, not a pass rate or a count of production customers.
What the design makes explicit
The workflow distinguishes an issue found, a correction proposed, a correction applied, and a successful recheck. It also ties final approval to the exact export version. The reported tests covered fictional affiliate, health, and legal settings across several AI tools and informed the emphasis on client-specific rules and evidence states.
The next measurement
A production pilot should measure review time, false positives, missed material issues, and whether senior reviewers need fewer full rereads. This case demonstrates system design and a defined testing program. Production time savings and reliability remain to be measured.
Discuss a similar project
If senior approval remains a bottleneck, I can map your client requirements and design a QA workflow that makes the remaining human decisions clear.