AI-CHECK · Path 1: Understand and use safely
Checking and Improving AI Results
One day dedicated to the skill that decides between benefit and damage: systematically checking AI results. Sources and figures, plausible errors, counter-checks and the question of when a result needs sign-off.
- Duration
- 1 training day
- Group
- 4 to 12 people
- Format
- In-house · Remote
- Language
- German or English
- Lead time
- from 2 weeks
- Price
- €7,900 flat, plus VAT
Delivery from four actually registered participants. The date is confirmed in writing after agreeing scope, prerequisites and availability.
Who this track is for
Teams that already use AI drafts and now need the quality assurance to go with them: case handling, assistants, business departments, controlling-adjacent roles. Also a good follow-up day after AI-WORK or AI-LIT-LAB.
Who it is not for
Not for teams without any AI use, they start with the AI Literacy Praxislabor (AI-LIT-LAB). Not for the technical evaluation of AI systems with metrics and gates, that is Evaluation, Observability and Guardrails (AI-EVAL).
Starting situation
Drafts now come from AI, but checking is gut feeling: one person reads everything three times and saves no time, another adopts results unchecked and risks the error in the customer document. What is missing is a shared, fast and demonstrable check routine.
This exists after the track
- A checklist for AI results, practised on several case classes and adapted to your team
- A catalogue of plausible errors: invented sources, shifted time periods, silent assumptions
- Practised counter-checks: recalculation, source comparison, counter-questions to the model
- A documented escalation and sign-off rule for critical results
- A check-protocol template for cases requiring evidence
Prerequisites
First experience with AI drafts in daily work, for example from AI-WORK, AI-LIT-LAB or your own practice. A laptop with a browser is enough.
Preparation before the track
Before the day, collect three anonymised examples from your work where an AI result was wrong, incomplete or suspiciously smooth. No example at hand? We provide synthetic cases for your industry.
What's included
- 1 training day with Dino Bordonaro, in-house or remote
- Synthetic check cases with built-in, documented errors
- Checklist template and check-protocol template for continued use
- Certificate of participation and training records
- One 60-minute remote Q&A session four weeks later for contested cases
Agenda
Why plausible is not correct
The error classes of language models shown on real patterns: invented sources, wrong figure references, silent assumptions, a convincing tone. Exercise: three prepared outputs each contain one error, who finds it unaided?
The checklist, step by step
Sources, figures, completeness, faithfulness to the task, tone, data provenance. Every check step gets a concrete hand movement and a time budget. Checking must be faster than writing it yourself, or nobody will do it.
Counter-checks that work
Recalculating figures, comparing claims against sources, counter-questioning the model when unsure. Exercise on synthetic cases: which counter-check exposes which error type?
Your cases, round one
The anonymised examples you brought are worked through with the checklist. What would the list have caught, what not? The list is sharpened exactly where it fails on your own material.
Escalation and sign-off
Not every result needs four eyes, some need six. Decision tree: what do I sign off myself, what does a colleague check, what goes to the manager? The rule becomes written and team-ready.
Check protocol for the record
For cases requiring evidence: the lean protocol documenting what was checked. Exercise on your own case, after which the template is ready to use.
Team rule and next step
The adapted checklist becomes a team agreement: which check steps are mandatory from tomorrow, who maintains the list, when is it sharpened. Everyone leaves with a concrete application plan.
Exercises and lab share
Around two thirds of the day is checking work on synthetic cases with documented built-in errors and on the examples you brought along.
Platforms
Delivered with the AI tool approved in your organisation, in whichever version is available. Check cases are synthetic, real customer data is neither required nor accepted.
Transfer evidence
You pass if you find the built-in errors in the final case using the checklist and justify the escalation decision. The result is recorded in the training records.
Artifacts you take home
- Team-specific checklist for AI results
- Error catalogue of plausible mistakes with counter-checks
- Escalation and sign-off rule as a team agreement
- Check-protocol template for cases requiring evidence
Optional extensions
- Better Briefs, Better AI Results (AI-CONTEXT), so fewer errors arise in the first place
- Evaluation, Observability and Guardrails (AI-EVAL) for systematic technical quality assurance
- Regular 60-minute Q&A sessions for contested cases from daily work
Boundaries
The track trains the human checking of AI results. It replaces neither professional review nor legal advice nor a technical evaluation pipeline.
Frequently asked questions
We already check everything. What is new here?
Reading everything is not checking, it is spending time without a system. The checklist replaces reading three times with targeted moves per error class. That is faster and demonstrably catches more, especially the convincingly worded mistakes.
Does the checklist work for every AI tool?
Yes, it checks results, not tools. Whether the draft comes from a chat assistant, a search integration or an internal system: sources, figures, assumptions and faithfulness to the task remain the same check questions.
What distinguishes this day from AI-EVAL?
AI-CHECK trains people to check individual results in daily work. AI-EVAL builds the technical quality assurance of an AI system: test cases, metrics, gates. One does not replace the other, they interlock.