Autopilot
// A day with Autopilot

What this looks like in practice.

Not from the GM’s perspective. From the perspective of the people who actually use the tool: junior professionals, service technicians, junior engineers. Three concrete day flows with real tool vocabulary — no glossy case studies.

// Junior tax advisor · 29 · StB firm 115 employees

A junior tax advisor in a Mittelstand firm

Employed professional · 18 own clients · case work for 30 more

The day

Tuesday morning, 8:30 am. Inbox: 40 client emails. 60% are recurring topics: status updates on tax returns, deadline reminders, document requests, BFH ruling enquiries. Classic: 12-15 minutes per email × 25 emails = four hours of the morning gone before the actual mandate work starts.

How it goes with Autopilot

  1. 08:30

    Open the inbox

    In Outlook, next to each client email there is a „Draft suggestion" button from Autopilot. Click on the first email.

  2. 08:31

    AI draft appears

    Draft with greeting, client context from Mandant-Memory (last deadline, ongoing proceedings, firm’s standard reply style), reference to specific paragraphs or BFH rulings with case numbers.

  3. 08:33

    Human review

    Account lead from the AC team has already cross-read — fact check, tonality, style are done. Two legal-professional decisions remain for me.

  4. 08:34

    My sign-off

    I check the two legal anchors (case number + subsumption), adjust if needed, click „release". My name is under the email as the professional — that is non-delegable.

  5. 08:34

    Send

    Email goes out with GDPR-compliant pseudonymisation — no personal data has left the firm via US cloud.

Sign-off block of the email
Reviewed by
L. Wagenknecht (StB) · Professional sign-off
Modell
Claude Opus 4.7 (EU hosting via Hetzner)
Risk-Class
limited · EU-AI-Act
Anmerkung
Case reference § 233a AO + BFH X R 32/22 verified. Client data pseudonymised before API call.

What does NOT happen

  • Client data does not go to OpenAI in the US — pseudonymisation runs before the API call.
  • The AI does not decide legally — the two legal anchors I check, otherwise I don’t sign off.
  • My performance is not measured — how fast I click, how many emails I release per day, appears in no report.
  • Senior colleague (62) does not suddenly have to use AI — her 30 own mandates run unchanged without Autopilot, as long as she doesn’t want to.

What changes specifically

  • Client correspondence from 12-15 min to under 4 min per email — four morning hours become two.
  • The two saved hours flow into complex mandates that are real value creation.
  • Professional liability documented per email with sign-off — auditor and professional indemnity insurance have the audit trail.
  • Partner round sees: the firm has arrived in the AI era without senior colleague being overrun.
// Service technician · 44 · special-purpose machinery 380 employees

A service technician in field service

Industrial master mechatronics · 24 years in the company · 200+ travel days/year

The assignment

Tuesday noon, 1:40 pm. At the customer in Italy, in a hall without stable network. Machine is dead — controller error code E47 on the display, a code I haven’t seen yet. Five years ago I would have called my senior colleague who knows these by heart. He retires in a year.

How it goes with Autopilot

  1. 13:40

    Capture error code

    I type „E47 + axis drive stops after cold start" into the company tablet. Tablet has the offline-capable knowledge base available locally — even if hall WiFi gives nothing right now.

  2. 13:40

    AI answer with sources

    Answer: „E47 = encoder error axis 2. Probable cause: connector X12 loose after transport (3 documented incidents Q1 2025). Repair: visual inspection + connector re-lock, diagnostic run F11."

  3. 13:40

    Check source link

    Below the answer: service manual page 247, plus two comparable ticket reports from 2025 with case numbers, plus photo of the connector. Sources are linked directly — I can verify myself, not „the AI said so".

  4. 13:42

    Repair

    Connector X12 was loose — as suspected. Re-locked, diagnostic run F11. Machine runs.

  5. 13:50

    Report

    Service report for the workshop wiki: „E47 → X12 → confirmed by diagnostic run F11". Goes into the wiki, instantly available for the other 27 colleagues on the next assignment.

Sign-off block of the knowledge-base answer
Reviewed by
A. Müller (Account Lead AC, machinery)
Modell
Claude Opus 4.7 · Service-RAG mode
Risk-Class
limited · EU-AI-Act
Anmerkung
Source: service manual v3.4 p.247 + 3 ticket reports from the plant (2025-Q1). No comparable case older than 18 months considered.

What does NOT happen

  • Nobody sees that I am in Italy right now — my route data is not captured, my hotel WiFi is not logged.
  • My timesheet is not augmented by AI — when I repair the machine, when I take a break, is not part of data collection.
  • I am not evaluated on whether I am faster than other service technicians — there are no person-level performance metrics.
  • I don’t have to „learn it myself" — the system has grown with real incidents that my senior colleague reviewed when writing them up.

What changes specifically

  • E47 error solvable without Sergio’s knowledge — also in 2 years, when he is retired.
  • Response time at the customer: from „call senior colleague and wait 20 min" to „30 seconds on the tablet".
  • 40 years of senior knowledge moves into the system — preserved, not replaced. The next service technician generation learns with it.
  • My workshop wiki, which I „should be maintaining anyway", gets better with every assignment. Maintenance happens as a side-effect, not as an extra evening task.
// Junior engineer · 27 · engineering firm 180 employees

A junior engineer in structural engineering

Master Structural Civil Engineering (TU Berlin) · 4 active projects · RFEM + Allplan

The day

Thursday morning, 9:15 am. Project lead handed me a school building mandate yesterday — structural design roof, snow load design per EC1. Location: terrain height 380m, snow load zone 2 per map. Normally I would now spend 4-5 hours with Beuth search, old application examples and Excel tools before I even start the calculation.

How it goes with Autopilot

  1. 09:15

    Norm query

    I type into the office tool: „Design value snow load zone 2, terrain height 380m, EC1 with national annex DE".

  2. 09:16

    Answer with source link

    Answer: „sk = 1.06 kN/m² for zone 2 at H=380m. Calculation per DIN EN 1991-1-3 / NA, table NA.1, factor μ1 = 0.8 for roof slope <30°, formula see eq. (NA.10)." Every element is clickable — DIN reference linked directly.

  3. 09:18

    Verification

    I click the DIN-EN-1991-1-3 link, see table NA.1 with the zone-2 value, see eq. (NA.10) myself in the norm excerpt. Take the value into my RFEM model — with the certainty that this is not hallucinated.

  4. 09:40

    Calculation in RFEM

    Instead of 5h research I have the load case in the model 25 min later. The free time flows into the actual structural analysis.

  5. 10:00

    Senior review

    My senior mentor reviews the calculation in the afternoon — he checks the subsumption, not the DIN research. Instead of waiting 2 hours for my preparation, he now has the finished model to review directly.

Sign-off block of the norm answer
Reviewed by
Dr.-Ing. T. Hartmann (Specialist civil engineer, 22 yrs practice)
Modell
Claude Opus 4.7 + Eurocode-RAG (DIN-Verlag licence)
Risk-Class
limited · EU-AI-Act
Anmerkung
Source: DIN EN 1991-1-3:2010-12 + DIN EN 1991-1-3/NA:2019-04, table NA.1. Value sk = 1.06 kN/m² for zone 2, H=380m verified against NA calculation tool.

What does NOT happen

  • The AI does not compute the structural safety verification — my professional liability as an engineer is non-delegable.
  • Invented DIN table numbers do not end up in the plan — every cited value is clickable via the source link.
  • Client data and city specifications do not go to US servers — all processing runs EU cloud Hetzner with GDPR-compliant pseudonymisation.
  • My senior mentor is not replaced — his review becomes more valuable because he can focus on the subsumption instead of norm lookups.

What changes specifically

  • Norm research: 5h → 40min per project. With 4 projects = 16h junior time free per week.
  • This free time flows into engineering-grade work I studied for.
  • Senior-mentor reviews become content-deeper because the lookup work is gone.
  • I become visible in the office as „tech champion" — career differentiation that helps me in the partner round later.
// Want to see the day yourself

Try it in the trial or talk to us.

7-day trial without a card: you walk through 5 pre-filled quick-win templates, exactly these workflows in your industry. Or directly a 30-minute first call if you as GM/partner want to set up the mandate.