A services business was spending 19 hours every week manually reviewing employee time recordings in Harvest. Quixas built a five-stage AI review pipeline that reduced that to 2 hours, an 89% reduction in review time.
The Problem
The operations assistant was losing nearly half their working week to a single repetitive task:
- Exporting Harvest recordings, reviewing each entry for completeness, accuracy, correct project tagging, and policy compliance, then drafting individual messages to employees whose entries needed correction
- Manual review at this volume introduced inconsistency. The same type of entry might be flagged one week and missed the next, depending on fatigue or workload
- Employees waited for the assistant to work through the full review cycle before being notified about corrections, by which point the context for the entry was days old
What We Built
AI-Powered Recording Review
Claude evaluates every Harvest entry across multiple dimensions: description completeness, time reasonableness, project accuracy, and policy compliance. Each entry receives a structured rating and a confidence score. No entry is skipped, no entry is reviewed inconsistently.
Automated Classification and Routing
Based on Claude's output, entries are classified into three categories: Approved (auto-cleared), Flagged (employee notified automatically), and Escalated (routed to the assistant with full context). Only the roughly 10% of genuinely ambiguous entries reach a human.
Multi-Channel Notifications
Employees receive immediate Slack or email alerts when their entries are flagged, with specific guidance on what needs correcting. The assistant gets a daily digest of escalations. Management receives automated weekly reports covering approval rates, common issues, and compliance trends.
The Pipeline
Stage 1: Automated Data Ingestion — n8n pulls all new Harvest recordings via API
Stage 2: AI Review by Claude — each recording evaluated against review criteria
Stage 3: Classification and Routing — Approved, Flagged, or Escalated
Stage 4: Notifications and Alerts — employees, assistant, and management notified
Stage 5: Human-in-the-Loop — assistant reviews only genuinely ambiguous cases
Stack: n8n (self-hosted) · Claude AI (Anthropic) · Harvest API · Slack / EmailThe Result
19 hours of manual recording reviews reduced to 2 hours per week. An 89% reduction in review time, with 100% of recordings now reviewed by AI on every cycle.
89% reduction in weekly review time. From 19 hours of manual review every week to 2 hours of human escalation review. Approximately 68 hours per month returned to the operations assistant for higher-value work.
Consistent, standardised evaluation on every entry. Every recording evaluated against the same criteria every time. No variation based on fatigue, workload, or inconsistent interpretation.
Employees notified immediately on flagged entries. Automated Slack and email alerts fire the moment an entry is flagged, with specific guidance on what needs correcting.
Management visibility without manual reporting. Weekly automated reports covering total recordings reviewed, approval rates, common issues, and compliance trends.
"We were spending 19 hours every week on something that should not have required a person. Now it takes 2. The quality of the reviews is actually better, because the criteria never changes."
Operations Director, Services Business
If your team is spending hours on reviews that follow predictable criteria, learn about our process automation and how we eliminate repetitive operational work.
