THE ASK: Build an AI Developer Onboarding Assistant for SAP’s internal tooling at 4.5 engineer-months ($450K all-in) to reduce new developer ramp-up time by 60%.
THE BET: We believe 80% of new engineers will complete their first commit within 3 days (vs. 14-day baseline) by using auto-generated, personalized onboarding guides that adapt to codebase and team changes.
THE ROI EQUATION:
2,500 new hires/year × 10 hrs saved × $85/hr (blended eng cost) = $2.125M/year recoverable
(source: SAP HR hiring forecast 2025, time-savings from pilot study n=22, eng cost from Regional Cost Benchmarks)
If adoption is 40%: $850K/year
Kill criteria: If <50% of new hires generate a guide within 30 days or time-to-first-commit exceeds 7 days at D90, halt investment.
WHAT THIS IS: An AI-driven onboarding guide generator that surfaces role-specific setup steps, codebase context, and task checklists.
WHAT THIS IS NOT: A real-time code assistant or replacement for human mentorship.
COMPETITOR APPROACHES:
- Notion: Static template wikis require manual updates (users "hire" for documentation)
- Linear: Task tracking for onboarding tickets (users "hire" for accountability)
- Jira Service Desk: Ticket-driven Q&A (users "hire" for request tracking)
| Capability | Notion | Linear | This Product |
|---|---|---|---|
| Auto-personalized guides | ❌ | ❌ | ✅ (unique) |
| Real-time sync with codebase | ❌ | ❌ | ✅ |
| First-task recommendations | ❌ | ✅ | ✅ (AI-curated) |
| WHERE WE LOSE | Content ecosystem | Task integration | ❌ vs Linear’s Jira sync |
Our wedge is contextual codebase awareness because only we ingest repo structures and CI/CD configs to generate precise setup steps.
WHO / JTBD: When a new software engineer joins a 50+ person SAP product team, they need to configure their environment, understand the codebase structure, and complete a first commit without relying on fragmented wikis or interrupting senior engineers.
THE GAP: New engineers can access generic onboarding docs but cannot get team/role-specific guidance. This forces them to:
- Manually reconcile conflicting setup guides (avg. 6.8 hrs wasted, n=45 new hires)
- Ask 15+ basic questions via Slack/email in week 1 (source: Confluence survey, Q1 2025)
- Delay first commits by 14 days avg (source: Git commit timestamps, n=120)
COST OF WORKAROUNDS:
| Symptom | Frequency | Cost |
|---|---|---|
| Senior eng Q&A time | 5 hrs/new hire | 12,500 hrs/year × $85 = $1.06M |
| Delayed productivity | 14 days ramp-up | 35,000 lost eng-hours/year = $2.97M |
| Setup errors | 22% require IT rework | $380K/year in IT labor (source: IT ticket analysis) |
JTBD STATEMENT: "When I join a new SAP team, I want a single, always-current guide with my exact setup steps and first tasks so I can commit code without scavenging docs or bothering colleagues."
CORE FLOW:
- Inputs: New hire’s role (e.g., "Java Backend"), team ID, repo access permissions
- AI Engine:
- Ingests team’s Confluence/Git/CI/CD docs via SAP Internal Graph API
- Extracts setup steps, codeowners, and recent "good first issues"
- Output: Personalized guide with:
- Environment setup (OS-specific commands)
- Code tour (key services diagram + entrypoint files)
- Week 1 checklist (3 tasks max)
- Contact list (tech lead, onboarding buddy)
┌───────────────────────────────────────────────┐
│ SAP Onboarding Assistant: [Engineer Name] │
├───────────────────────────────────────────────┤
│ ✅ ENVIRONMENT SETUP │
│ - Install: Java 17, Docker, team CLI v3.2 │
│ - Config: db-settings-dev.yaml (link) │
│ │
│ 🗺 CODE TOUR: Order Processing Service │
│ - Entrypoint: OrderService.java (link) │
│ - Architecture: [Diagram] │
│ │
│ 📋 WEEK 1 TASKS │
│ 1. Fix log formatting in #ISS-204 (link) │
│ 2. Add test for OrderValidator (template) │
│ │
│ 👥 KEY CONTACTS │
│ - Tech Lead: Maya R. (Slack) │
└───────────────────────────────────────────────┘
Phase 1 — MVP (6 weeks):
US#1 — Generate Environment Setup
- Given new hire’s OS + team ID
- When guide generates
- Then setup steps include 100% correct package names/versions from team’s CI config
If fails: Senior eng receives alert to manually correct
Validated by QA against 20 team configs
US#2 — Recommend First Task
- Given engineer’s role (backend/frontend)
- When guide generates
- Then recommends 1 "good first issue" from Jira with ≤3-day estimated effort
P1: ≥95% match to eng capability
Out of Scope (Phase 1):
| Feature | Why Not Phase 1 |
|---|---|
| Real-time guide updates | Requires event streaming (Phase 1.1) |
| Multi-repo dependency mapping | Needs cross-repo analysis (Phase 1.2) |
| Custom task sequencing | AI training data insufficient |
PRIMARY METRICS:
| Metric | Baseline | Target (D90) | Kill Threshold | Method |
|---|---|---|---|---|
| Avg time-to-first-commit | 14 days | ≤5 days | >10 days | Git timestamp analysis |
| Senior eng Q&A volume | 15 queries/new hire | ≤5 | >12 | Slack/Jira ticket audit |
| Guide utilization rate | N/A | ≥70% | <40% | Assistant login logs |
GUARDRAIL METRICS:
| Metric | Threshold | Action if Breached |
|---|---|---|
| Setup error rate | ≤5% | Pause rollout, fix doc sourcing |
| P95 guide gen latency | <8s | Optimize AI model |
WHAT WE ARE NOT MEASURING:
- Total guide views (vanity; doesn’t indicate usefulness)
- User satisfaction scores (lagging; use behavior instead)
- Number of tasks completed (gaming risk; focus on commit time)
RISK 1 — Outdated Source Docs
- Failure Mode: AI uses stale Confluence page, causing setup failures
- Probability: High | Impact: High
- Mitigation: Add "last updated" timestamps + alert doc owners if >30d old (owned by Content Team)
- Detection: Monitor "Report outdated guide" clicks (threshold: >5/week)
RISK 2 — Low Senior Eng Adoption
- Failure Mode: Tech leads skip updating "good first issues"
- Probability: Medium | Impact: Medium
- Mitigation: Auto-pull Jira tickets tagged "beginner" (owned by PM)
- Detection: <50% of guides show Jira tasks at D30
COMPLIANCE RISK — GDPR Data Mining
- Failure Mode: AI scans PII in code comments during ingestion
- Probability: Low | Impact: Critical
- Mitigation: Exclude .md/.txt files; legal review by 2025-10-15 (owned by Legal)
- Consequence: If GDPR clearance not obtained, disable code scanning for EU teams
KILL CRITERIA (within 90 days):
-
50% of new hires create manual setup tickets
- Time-to-first-commit fails to drop below 10 days
- IT spends >10 hrs/week fixing Assistant-induced setup errors
Decision: Scope of codebase analysis
Choice Made: Parse only entrypoint files and CI/CD configs (not full repos)
Rationale: Full-repo scanning adds 3+ weeks latency; rejected due to security/compliance overhead
Decision: Update trigger for guides
Choice Made: Nightly rebuilds (not real-time)
Rationale: Real-time requires event-stream integration (6+ weeks effort); daily updates cover 95% of changes (source: repo churn analysis)
Decision: Fallback when AI confidence <85%
Choice Made: Show "Verify with Tech Lead" flag + link to legacy docs
Rationale: Prevents incorrect instructions; rejected "blocking user" as too disruptive
Decision: Phase 1 integrations
Choice Made: Confluence + GitLab + Jira (exclude ServiceNow)
Rationale: Covers 80% of teams (source: SAP tooling survey); ServiceNow adds 4 weeks for auth
BEFORE/AFTER NARRATIVE:
Before: Lena (new backend engineer) spent 3 days finding conflicting Java version requirements across 4 wikis. She asked 12 Slack questions about deployment scripts, delaying her first fix (#ISS-204) to day 16. Her tech lead spent 45 minutes explaining the code structure.
After: Lena logs into the Assistant on day 1. It generates: (1) exact Java 17 install commands from her team’s CI file, (2) a code tour highlighting OrderService.java, (3) #ISS-204 as her first task. She commits a fix on day 2. The tech lead reviews only the PR.
PRE-MORTEM:
"It is 6 months from now and this feature has failed. The 3 most likely reasons are:
- Critical teams’ repos used unsupported config formats (e.g., Bazel), making 30% of guides unusable
- Senior engineers refused to maintain "good first issues," leaving task recommendations empty
- Legal blocked code scanning for 40% of EU teams, fragmenting adoption
SUCCESS LOOKS LIKE: New hires commit code on day 3 without asking "where do I start?" in Slack. Tech leads report 80% less onboarding interruptions. The COO cites the Assistant in a board meeting as cutting new-engineer productivity loss by half."