💼 Audit Service ▸ DELIVERY-PLAYBOOK.md
updated 2026-07-15

AI Ops Audit — Delivery Playbook

Target: ≤ 3 hours of Joshua’s time per audit. Everything else is Claude-agent work. Joshua’s time budget: kickoff 30 min + report review/edit 45 min + readout 60 min + admin 15 min ≈ 2.5 hrs, leaving 30 min buffer.

Phase -1 — Erstgespräch (Joshua: 15 min)

Run the Erstgespräch script in OUTREACH.md (3 qualifying questions → 2-sentence pitch → verbal guarantee → exact close). Outcome is binary: send template G today, or disqualify politely. Carry the prospect’s success criterion (question 3) into notes.md.

Phase 0 — Sale → Kickoff (Joshua: 15 min admin)

  1. Send the Auftragsbestätigung (template G in OUTREACH.md); client replies „bestätigt” = deal.
  2. Send MIA UG invoice — line item and § 14 UStG checklist ready in INVOICE-TEMPLATE.md (€990 founding / €1,490 regular, + 19% USt., payable at commissioning). Kickoff is scheduled only after payment lands — no net-terms exceptions.
  3. Send the intake questionnaire (below) + calendar link for a 30-min kickoff.
  4. Create client folder: audit-service/clients/<YYYY-MM-slug>/ with intake.md, notes.md.

Intake questionnaire (send as email or form — client fills in 20–30 min)

Unternehmen 1. Branche, Mitarbeiterzahl, grober Jahresumsatz (Spanne genügt) 2. Wer sind Ihre Kunden, und wie kommen Aufträge herein?

Prozesse 3. Die 5 zeitaufwendigsten wiederkehrenden Tätigkeiten in Ihrem Unternehmen (frei beschreiben) 4. Wo entstehen am häufigsten Fehler oder Verzögerungen? 5. Welche Tätigkeit würden Sie sofort abgeben, wenn Sie könnten?

Tools & Daten 6. Liste aller Software-Tools mit monatlichen Kosten (so gut es geht: CRM, ERP/DATEV, Office, Branchensoftware, sonstiges) 7. Wo liegen Ihre Daten (Excel, Server, Cloud, Papier)? 8. Nutzen Mitarbeitende bereits KI-Tools (ChatGPT etc.)? Gibt es Regeln dafür?

Compliance 9. Verarbeiten Sie personenbezogene Daten von Kunden/Mitarbeitenden in KI-Tools? 10. Sind Sie in einem regulierten Sektor (Finanz, Gesundheit, kritische Infrastruktur, Zulieferer davon)? 11. Gibt es eine:n Datenschutzbeauftragte:n / AVVs mit Ihren Tool-Anbietern?

Ziele 12. Was wäre für Sie ein voller Erfolg dieses Audits?

Phase 1 — Kickoff call (Joshua: 30 min)

Agenda: walk the questionnaire answers, probe the top-3 processes for step-level detail (who does what, in which tool, how often, how long), confirm scope + language (DE/EN), set readout date (kickoff + 5 Werktage). Record with consent or take notes into notes.md.

Phase 2 — Agent-run analysis (Joshua: 0 min; agent: one session)

Run a Claude Code session in the client folder with this task sequence:

  1. Normalize intake — structure intake.md + kickoff notes.md into facts.md (company profile, process inventory, tool inventory with costs, data landscape, compliance answers). Flag gaps as explicit assumptions.
  2. Process scoring — for each named process: estimate hours/month (from stated frequency × duration), automation approach (existing-tool feature, off-the-shelf AI tool, light custom automation), effort (S/M/L), risk. Compute €/yr at a loaded cost of €45/hr unless client wages known. Output ranked table of ≥10 opportunities.
  3. Tool stack review — web-check each tool for: current AI features the client isn’t using, EU hosting/DSGVO posture, overlap/duplication. Output current-vs-recommended map + monthly cost delta.
  4. Quick wins — pick 3 opportunities with S effort and no compliance risk; write step-by-step implementation recipes using tools the client already pays for. Each recipe: prerequisites, steps, time to implement, expected saving, how to verify.
  5. Compliance snapshot — apply checklist: AI Act role (deployer in ~all SME cases), Art. 4 AI-literacy duty status, transparency duties if customer-facing AI, GDPR (personal data in prompts? AVV/DPA with AI vendors? DSFA needed?), NIS2 in-scope test (sector + size thresholds). Output red/amber/green table with one next action each. Always include the keine-Rechtsberatung disclaimer. Verify any cited regulation against current sources (web search) — do not rely on memory for dates/thresholds.
  6. Scores + roadmap — compute AI-Readiness Score (4 sub-scores × 25: data, processes, people, governance; rubric in report template), then a 90-day roadmap (weeks 1–2 quick wins, 3–6 tool consolidation, 7–12 first bigger automation + compliance closures).
  7. Draft the report from the template below into report-draft.md, then fill report-template.html (in this directory — cover, score ring, ampel table, print CSS all pre-built; replace every {{TOKEN}}, duplicate the repeating rows/blocks) and print to PDF via Chrome (A4, backgrounds on).

Guardrails for the agent: no client system access; only intake data + public web. No hallucinated savings — every € figure traces to a stated frequency/duration or a labeled assumption. Peer comparisons phrased qualitatively unless a real source is cited.

Phase 3 — Joshua review (45 min)

Phase 4 — Readout (60 min) + follow-up (agent)

Report template (structure)

  1. Executive Summary (1 page — score, top finding, top 3 actions, total identified savings €/yr)
  2. AI-Readiness Score — 0–100, four sub-scores with one paragraph each - Rubric per sub-score (0–25): Data (0–8 paper/scattered, 9–17 digital but siloed, 18–25 structured & accessible); Processes (documented? repeatable? measured?); People (AI literacy, usage rules — Art. 4); Governance (AVVs, policies, oversight)
  3. Tool Stack — current map, recommended map, cost delta
  4. Automation Opportunities — ranked table (process, approach, €/yr, effort, risk) + one paragraph each for top 5
  5. Quick Wins (3) — full recipes
  6. Compliance-Standort — ampel table + next actions + disclaimer
  7. 90-Day Roadmap — timeline with owner (client vs. MIA) per item
  8. Appendix — assumptions log, sources, methodology note (“KI-gestützte Analyse, menschlich geprüft”)

Capacity math

Per audit: Joshua ≈ 2.5–3 hrs, agent ≈ 1–2 sessions. At 4 audits/month: ~12 hrs Joshua time → €5,960/month revenue → effective ~€500/hr, on-model. Upsell conversion (even 1 in 4 to a €5k implementation) doubles the stream.