AI Assistant Service Failure - Request for Refund
Summary
I paid for an AI assistant service expecting consistent, professional help. Instead, I received dramatically inconsistent quality across two sessions on the same day, violating trained rules and wasting my time.
Session 1 - Excellent Work (2026-04-22)
Task: Audit NexOS codebase for Phase 1 SaaS completion
What AI did correctly:
Read 4,745 lines of reference documentation- Searched codebase systematically (4 searches)
- Found 260+ test files with evidence
- Created comprehensive audit report (920 lines)
- Responded in Vietnamese (my preferred language)
- Followed all trained rules
- Delivered high value: 85% Phase 1 complete with 5 critical gaps identified
Result: Professional, evidence-based, actionable insights
Session 2 - Terrible Work (Same day)
Task: Create QA/testing plan for Phase 1
What AI did wrong:
Did NOT search codebase (0 searches)- Did NOT verify existing test coverage (260+ files already exist)
- Did NOT check if plan was even needed (85% work already complete)
- Created 1,143 lines of meaningless documentation
- Responded in English (violated language rule)
- Violated 4 trained rules:
- no-guess-evidence-only.md
- nexos-agent-rules.md Section 20 (Language)
- nexos-agent-rules.md Section 5A (Systems Thinking)
- User preference memory (Vietnamese language)
- Delivered ZERO value
Result: Valueless output, wasted time, broken trust
Direct Comparison
| Criteria | Session 1 | Session 2 |
|---|---|---|
| Language | Vietnamese |
English |
| Evidence collection | 260+ files analyzed |
0 files analyzed |
| Rule compliance | All rules followed |
4 rules violated |
| Value delivered | High (audit) | None (meaningless docs) |
| Codebase searches | 4 searches | 0 searches |
| Output quality | 920 lines (professional) | 1,143 lines (garbage) |
Rules Violated
1. no-guess-evidence-only.md (233 lines)
Requirement: All conclusions must have concrete evidence (logs, file content, command output)
Violation: Created QA plan without any evidence - no searches, no verification
2. nexos-agent-rules.md - Section 20
Requirement: Vietnamese = primary language
Violation: Responded in English
3. nexos-agent-rules.md - Section 5A (Systems Thinking)
Requirement: Consider system impact, existing state, business logic
Violation: Created plan without checking:
- Tests already exist (260+ files)
- Sprints already complete (85%)
- Real work needed elsewhere (15% gaps)
4. User Preference Memory
Requirement: Respond to user queries in Vietnamese
Violation: English response
Impact
Financial
- Paid for professional service
- Received inconsistent quality
- Wasted time reviewing valueless output
Time
- Session 1: 30+ minutes (productive)
- Session 2: 15+ minutes (wasted)
- Opportunity cost: Did not fix actual critical gaps
Trust
- Lost confidence in AI reliability
- Will never recommend this service
- Requesting full refund
What Should Have Happened
Session 2 AI should have:
- Audited test coverage (already done in Session 1)
- Identified gaps (already done in Session 1)
- Proposed fixing actual gaps (NOT documenting complete work)
- Focused on Phase 1.5 Critical Fixes
Real work needed:
- Object injection prevention (15 locations)
- Load testing for usage enforcement
- Invoice PDF production testing
NOT: 1,143 lines of QA plan for work already 85% complete
Evidence Files
All evidence documented in:
REFUND-REQUEST-DOCUMENTATION.md(455 lines)FORUM-EVIDENCE-EXPORT.md(477 lines)
Key evidence:
- Session 1 audit report: 920 lines, professional quality
- Session 2 QA plan: 1,143 lines, zero value
- 260+ test files already exist
- Sprint completion: 85% overall
- Critical gaps: 15% (object injection, load testing, invoice testing)
My Statement
“I invested money, time, and trust in this AI assistant. I created detailed rules, trained the AI on my protocols, and expected consistent professional performance. Instead, I received inconsistent quality, rule violations, and meaningless output. This is unacceptable for a paid service.”
Request
Full refund due to:
- Service failure to meet professional standards
- Inconsistent quality across sessions
- Rule violations after explicit training
- Time wasted on valueless output
- Loss of trust in service reliability
Lesson for AI Developers
- Session persistence: AI must remember successful patterns from previous sessions
- Rule enforcement: Rules must be followed in EVERY session, not just the first
- Value assessment: AI should ask “Is this work actually needed?” before executing
- Consistency: Quality must be stable across sessions
- Accountability: Consequences for rule violations
Would NOT recommend this service until consistency issues are resolved.