funnel-analytics

Funnel Analytics

Pure reasoning skill. Reads structured pipeline + application data and produces funnel analysis with actionable insights.

Built for: Recruiter or recruiting lead running weekly / monthly / quarterly reviews of how the funnel is performing.

When to Use

• Inside /report command
• "How's the funnel looking?"
• "Where are we losing candidates?"
• Before quarterly planning to decide where to invest

Phase 0: Inputs

Read recursively:

• clients/<cliente>/jobs/*/pipeline.md (snapshot of each vacancy)
• clients/<cliente>/jobs/*/candidates/*/application.md (per-candidate state + history)
• clients/<cliente>/jobs/*/reports/index.md (past reports for trend)

Confirm:

• Period: week / month / quarter / YTD / custom
• Scope: single client / all clients / specific vacancy
• Mode: funnel / interviewer-calibration / source / diversity / quality

Phase 1: Funnel Mode (default)

Stage-by-stage conversion

Pipeline canonical stages:

Sourced → Applied → Screening → Phone → Technical → Onsite → Offer → Accepted → Hired

For the period, compute:

• Volume entering each stage
• Volume passing to next stage
• Pass-through rate (= passed / entered)
• Median time in stage

Output table

| Stage transition | N entered | N advanced | Pass rate | Benchmark | Status |
|---|---|---|---|---|---|
| Applied → Phone screen | 150 | 30 | 20% | 10-20% | 🟢 |
| Phone → Technical | 30 | 12 | 40% | 30-50% | 🟢 |
| Technical → Onsite | 12 | 6 | 50% | 40-60% | 🟢 |
| Onsite → Offer | 6 | 2 | 33% | 30-50% | 🟢 |
| Offer → Accepted | 2 | 2 | 100% | 80-90% | 🟢 |

Benchmarks (rough — adjust by domain)

Stage transition	Healthy range
Sourced → Applied (outbound)	5-15%
Applied → Phone screen	10-25%
Phone → Technical	30-50%
Technical → Onsite	40-60%
Onsite → Offer	25-50%
Offer → Accepted	80-95%

Outside these ranges → bottleneck or opportunity. Diagnose:

• Lower than range at applied → phone: bad sourcing quality OR over-strict screening
• Lower than range at phone → tech: scorecard not calibrated to JD
• Lower than range at onsite → offer: panel calibration drift OR debrief discipline weak
• Lower than range at offer → accept: comp mis-calibration OR slow process OR weak close

Time metrics

For the period:

• Median time-to-fill (req approved → offer accepted)
• Median time-to-hire (applied → accepted)
• Median time-in-stage per stage

Reason code analysis

Tabulate rejection reason codes for each stage:

| Stage | Top 3 reason codes | Volume |
|---|---|---|
| Phone screen | skills-gap (12) | level-mismatch (8) | comp-mismatch (5) |
| Technical | failed-assessment (8) | code-quality-concern (4) | comm-concern (2) |
| Onsite | level-mismatch (4) | judgment-concern (3) | culture-decompose (2) |

If "comp-mismatch" is high at phone screen → comp band is wrong vs market. If high at offer → calibration is wrong, not market.

Phase 2: Interviewer Calibration Mode

For each interviewer in the period:

• Number of interviews conducted
• Score distribution (avg + std deviation)
• Final-decision alignment (% of times their recommendation matched panel decision)
• Score variance vs other interviewers on the same candidates

Identify outliers:

• Hawk (consistently low scorer) — risk of false negatives, scaring off hires
• Dove (consistently high scorer) — risk of false positives, weak hires through
• Inconsistent (high std deviation) — needs calibration training

Recommend: shadow interviews for outliers, debrief facilitation, calibration training.

Phase 3: Source Mode

For each sourcing channel:

• Candidates entering pipeline
• Candidates reaching phone screen
• Candidates reaching offer
• Hires
• Cost (if known — outbound is recruiter time, paid is direct $)

Output:

| Source | Candidates | Hires | Hire rate | Cost/hire | Quality (post-hire) |
|---|---|---|---|---|---|
| Referrals | 12 | 3 | 25% | $0 | High |
| Outbound LinkedIn | 80 | 2 | 2.5% | recruiter hours | Mixed |
| Inbound (careers page) | 60 | 1 | 1.7% | $0 | High |
| Agency | 8 | 1 | 12% | $25k | High |
| ... | | | | | |

Decisions:

• Double down on top performers
• Kill or fix the bottom (if cost > value)
• Investigate quality drift if source has high hire rate but low post-hire performance

Phase 4: Diversity Mode (only with explicit consent)

⚠️ Only run if client has documented:

• Candidate consent for demographic collection
• Demographics stored separately from decision-maker access
• Aggregated reporting only, never individual

For each demographic dimension collected:

• Pipeline composition at each stage
• Stage-by-stage drop-off rates

If drop-off rates differ significantly by demographic at a specific stage → that stage has bias signal. Investigate:

• JD language (entry-funnel bias)
• Screener bias (phone stage bias)
• Interviewer bias (panel stage bias — drill into specific interviewers)

Do not report individual data, ever. Only aggregated, only above minimum cohort size (10+).

Phase 5: Quality of Hire Mode

Cross-reference hires with performance data (from HRIS, 6-12 months post-hire):

• Performance review rating
• Retention (still here at 12m? 24m?)
• Voluntary attrition (did they leave or were they let go?)
• Promotion velocity (above/below cohort norm?)

Compute:

• Correlation between scorecard score and post-hire performance
• Correlation between interviewer's recommendation and post-hire performance (which interviewers' calls predict?)
• Source → quality (which sources produce hires that stick + perform?)

This is the highest-value analysis but the hardest to run. Most agencies don't have HRIS access — but if you do, this is the loop that improves the entire process.

Phase 6: Insights and Recommendations

Don't just report numbers. For each report:

• Top 3 insights — concrete patterns, not platitudes
• Top 3 recommended actions — specific, owner-assignable

Examples:

• ❌ "Funnel conversion is healthy." (no signal)
• ✅ "Phone → Tech pass rate dropped from 45% (last quarter) to 28% (this quarter), driven by 12 'level-mismatch' rejections on the SDR role. Hypothesis: JD was rewritten last quarter and over-sells seniority. Action: Recruiter to refresh JD + recalibrate phone screen rubric this week."

Phase 7: Output

Write to clients/<cliente>/reports/<mode>-<YYYY-MM-DD>.md:

# Funnel Report — [Cliente] | [Period]

## Executive summary
[3-5 bullets — health snapshot + top issue]

## Volume + conversion
[full funnel table]

## Time metrics
[table]

## Reason code analysis
[table]

## [Mode-specific deep-dive section]

## Top insights
1. [...]
2. [...]
3. [...]

## Recommended actions
| Action | Owner | Due |
|---|---|---|
| [...] | [...] | [date] |

## Notes on data quality
- [Period had N applications — small sample, treat % with grace if <30]
- [Missing reason codes on N records — fix process going forward]

Append to clients/<cliente>/reports/index.md:

• [Date] — [mode] report — [1-line takeaway] — link

Anti-patterns to refuse

• Report without reason codes — refuse, ask user to backfill
• Diversity report without explicit consent flag — refuse
• Quality-of-hire without HRIS data — flag, can't be done
• Conversion rate analysis with <10 candidates in funnel — flag as noisy, don't lead decisions
• "Funnel is healthy" with no specifics — push back, force concrete insight