Production-grade AI workflows, adopted by your team, with a measured financial return.

No. If you already know the one process that needs work, start here. Week one runs the same check on that process. The review is for when you’re weighing several, or don’t know where to start, and its fee counts toward the sprint.

That’s what week one checks, before any work on the AI workflow.
• If something you already pay for handles it, or a simple template does the job, you hear that plainly.
• You pay €2,900 for that week and nothing more. It’s a real finding, not a setback.
• If week one says go, your fixed price is an offer valid for 30 days. You can still stop at €2,900.

• Useful information, not a setback: a failed first attempt has a cause worth finding before you try again.
• You start from “A pilot that never went live”, and week one decides whether to finish or rebuild it.
• Finishing usually costs less than starting from scratch: a pilot brings data access, cases and a working path.
• If week one says rebuild, you get the from-scratch quote before going further.

• A decision-maker who can say yes or no, about 2 hours a week.
• One expert who does the work today: 4 to 8 hours in total, mostly in week one.
• Access to the systems involved, agreed in writing before week one.
• About 30 to 100 past cases with known right answers, depending on where you start.
• Today’s numbers, or permission to measure them in week one: volume, time per case, error rate.
• One IT or security contact for two short reviews.
• Time for the people who do the work to join rollout sessions.
• If your AI workflow is live: access to its logs, and short voluntary interviews with 3 to 5 people who stopped using it.

• The quality bar covers what I control. Go-live is judged on accuracy and safety on your held-back cases, logging and fallbacks.
• Rollout sessions and a named owner are checked at the results report.
• If I miss the bar, I keep fixing for up to two more weeks at no charge.
• Still missed? You choose. Keep the AI workflow running with your staff approving each case: the go-live payment is due, and the final 30% is waived.
• Or stop: nothing further is due.
• Usage and financial value are measured and reported 30 days after go-live, not guaranteed.
• If the cause is on your side, such as late access, missing cases, changed requirements or a quality bar set against my advice, the timeline moves instead.

In four parts, each tied to work you can check:
• €2,900 when you sign, for week one.
• 30% of the rest when you sign off the test cases and quality bar, around week two.
• 40% at go-live, once the AI workflow meets the quality bar.
• 30% after the results report, 30 days after go-live.
• A sign-off counts as given if you don’t reply within 5 business days.
• If go-live slips on your side by more than 10 business days, the go-live payment is due anyway.
• Each invoice is due within 14 days. The exact terms are agreed in your proposal.
• Prices exclude VAT.
• Running costs (AI model use, hosting, licences) are paid to your providers, never marked up by me.
• Typical model cost per case is small: under 1 US cent in my synthetic benchmark.
• Week one estimates your cost per case and per month, and I set a spend cap.

Week one checks all three.
• Which data the AI workflow uses, where it’s processed, and the data processing agreement you need.
• Access is agreed upfront. Sanitized examples are shown before anything real is shared, and nothing proprietary leaves your environment without agreement.
• If you have a works council, a short note on how usage is tracked: per team, not per person.
Tracking usage can need works council approval, and the test is narrower than “any AI touching employees.” Under German law, co-determination is triggered when a system is technically capable of monitoring an individual’s behaviour or performance, whether or not that’s the intent. A 2024 Hamburg labour court decision found no co-determination right where staff used ChatGPT through their own accounts and the employer had no access to the usage data. That’s why usage here is tracked at team level by design, not per person, unless that kind of tracking is already part of someone’s role and they’ve agreed to it. AI used in hiring, pay or termination decisions is a separate, near-certain trigger regardless of tracking level. I’ll help you check the specific case with HR and your works council early, which German law expects anyway. This is general orientation, not legal advice.

The decision stays with your team, as with any tool, but every flagged case shows why it was flagged and every approval is logged. In week one we agree who is accountable for which cases. General orientation, not legal advice.

• I lead and do the work.
• Advisors help on positioning and finance.
• Specialists join only where named in your proposal.
• Code and documents sit in your accounts, so the work doesn’t depend on me.
• Unlike a permanent hire, the sprint is scoped to one process, and you see real work before committing to more.
• You know the fixed price at the end of week one, and cost per case is tracked from day one.
• Need someone to own AI beyond one process? That’s the Fractional AI lead.

• Yes. Week one agrees who builds what. Your developers can build alongside and review everything, and the price doesn’t change.
• You own the code, prompts, test cases and dashboards once paid, and your team is trained to run them.
• The sprint is scoped not to compete: week one decides build, buy or connect for each gap, checked against your roadmap and stack.
• Tools follow the process: Claude Code and Codex for development, LangChain, CrewAI and n8n for orchestration, Supabase and FastAPI for infrastructure, and Langfuse or LangSmith for logging. Clients who need EU-hosted or self-hosted logging get Langfuse.

• Usage is tracked from rollout, so a stall shows up early, not months later.
• The fix targets the actual cause. Training assumes the problem is know-how; often it’s trust, a poor fit with the process, or one bad early experience.
• If the real reason is that the AI workflow doesn’t help, you hear that plainly, not a push for more training.

It depends on what the AI could get wrong, not whether it’s customer-facing. For products, including AI systems, placed on the market or put into service on or after 9 December 2026, the directive covers harm to people: death or injury (including medically recognised psychological harm), damage to property that isn’t used exclusively for professional purposes, and destruction or corruption of data not used for professional purposes. It doesn’t cover damage confined to your own business property or records. Three questions worth asking: could a malfunction plausibly injure someone? Could it destroy or corrupt a person’s own data that isn’t used for work? Could it damage property that isn’t purely a business asset? If any answer is yes, being able to show how the system was tested, monitored and updated matters more for anything placed on the market or put into service on or after that date. The sprint builds traces, test results and an audit trail in from the start, which is the evidence this directive rewards. This is general orientation, not legal advice, so bring these three questions to your legal team rather than starting from scratch.