Report
AI-supported quality assurance prior to release
For our own approval process, we have developed a verification agent that checks figures, cross-references sources against our local database and challenges assumptions, using documents with deliberately embedded errors to calibrate the system.
Let’s discuss your project
Contact us directly and let's talk about a possible collaboration.
Background
Calculation errors, incorrectly cited sources and unstated assumptions are often only spotted by the client. The dual-review principle catches many of these, but it ties up senior staff’s time and becomes less effective towards the end of a project, as the same people have long since become familiar with the text.
Our Approach
-
1
Development of a verification agent that checks figures itself, tests every claim against locally stored sources and specifically challenges assumptions.
-
2
Calibration using documents containing deliberately introduced errors, tested blind, to highlight the hit rate and false positives.
-
3
Issued as a release form with a work list, indicating the severity of each finding based on its impact on the key result.
Result
The agent reliably detects the built-in critical errors without fabricating any findings. The second characteristic is the more challenging one in practice: a tester who imagines errors costs more time than they save. Every deliverable is therefore released with a list of resolved errors. Technical approval remains the responsibility of humans; the agent finds errors, but is not responsible for the results.