02/05·Pilot rollout

Pilot rollout.

We take one clearly defined workflow and build the agent for it. After three to four weeks it runs live on your infrastructure. With tests, logging, an eval set so you can see how good it is, and a runbook so your IT can take it over.

One workflow, three weeks, in production.

PIL-WF-3W

·

EUR 14,000 net (12 to 18k)

·

3 to 4 weeks

·

remote, 1 person, weekly 30-minute sync

0X/05·Methodology anchor

Where in the Six-Step-Plan map.

[00]

method.step0Title

method.step0Sub

[01]

method.step1Title

method.step1Sub

[02]

method.step2Title

method.step2Sub

[03]

method.step3Title

method.step3Sub

[04] ●

method.step4Title

method.step4Sub

[05]

method.step5Title

method.step5Sub

Step 4 in three weeks: one agent architecture, in production.

Volle Methode ansehen →

02/05·Detail

// promise

After three to four weeks one workflow runs in production, is evaluated, documented and handed over to an internal owner.

// process

Week 1

Setup, eval gold-set defined together, architecture finalised.

Week 2

Build, MCP/skill implementation, first end-to-end runs against gold-set.

Week 3

Hardening, edge cases, logging, deploy to client infrastructure, training 1.

Week 4

Stabilisation, drift observation, training 2, handover (optional).

// deliverables

+Productive workflow (Claude skill, MCP server or n8n integration, depending on audit architecture).
+Evaluation suite: at least 20 gold-set examples, automatic drift detection, run report.
+Runbook (PDF plus Markdown): operations, error patterns, rollback procedure, escalation path.
+Handover screencast (15 to 25 minutes).
+Status report (REP-YYYY-MM-DD-NNN, Janis brand) with before-and-after metrics.

exit criterion

Workflow has been running productively for at least 5 working days with less than 5 percent eval regression. Runbook and screencast delivered. Internal owner named and trained.

// tech footprint

Claude (Sonnet 4.6, Opus 4.6 if deep reasoning is required), MCP server, optionally n8n, Drizzle/PostgreSQL for state, Fly.io or Hetzner for deploy.

03/05·Reference profiles

Wer das typischerweise kauft.

Tax or law firm with audit deliverable in hand

branche: Tax or legal advisory, approx. 50 to 150 FTE.
größe: 50 to 150 FTE, ideally directly following a workflow audit.
anlass: Audit report on the table, management wants visible results within four weeks.
outcome (geschätzt): Workflow runs productively every day with five to twenty users. Processing time per case reduced from 4 to 8 hours to 20 to 60 minutes. Two documented eval runs as baseline.

HR or architecture consultancy with email triage

branche: B2B consultancy with high frequency of structured text work.
größe: 80 to 250 FTE, in-house IT of two to five people.
anlass: Recurring bottleneck in a clearly scoped process (quotes, supplier requests).
outcome (geschätzt): 60 to 80 percent of cases automatically classified and pre-answered. Capacity effect estimated at 0.5 to 1.2 FTE.

04/05·Testimonial

template, zur Freigabe durch Kunde

[TEMPLATE]

„After three weeks we had a Claude workflow in real operation, with eval suite and runbook. The documentation is structured so that our in-house developer continues to maintain the skill on their own."
Head of IT, B2B service provider, approx. 150 employees.

05/05·What you do not get

Does not fit.

×We don't yet know exactly what we want (audit first).
×More than two workflows in one sprint (scope creep, fixed price breaks).
×Compliance sign-off is still pending in week one.