Offer · AI decision audit
Which decisions in your process can a fast model take over?
Your agent or chatbot writes a sentence to answer a question that has two possible answers. You pay for every word, wait seconds, and cannot tell which answers to trust. I will test it on your data.
What we do
- Map the decisionsWe walk one process together, such as ticket handling, complaints or document flow, and list every point where a person or a model picks from a known set of answers.
- A test set from your dataA few hundred real cases, labelled by the people who do the work today. I bring the protocol: answers frozen before the test and an independent check of the labels.
- Test the modelsFast decision models, including ones that run on your own hardware, compared with a model that thinks before answering and with what you use today. We measure accuracy, the share of decisions a model can take alone at your acceptable error rate, latency and cost.
- Report and recommendationWhich decisions to hand to a fast model, at what confidence threshold, what stays with a person or a thinking model, and how it fits into your system.
What you keep
A report with numbers and a recommendation, the test set, and the scripts to rerun the test whenever the model or the process changes.
Your data
Your data can stay in-house: some decision models, such as basal, run on your own hardware. Cloud models are tested only with your consent and only on data you are allowed to share with an outside provider.
Why I know this works
I ran the same kind of test on about 400 Polish decisions from customer service and twelve other fields. The results, charts and limits are in my write-up, and the data and scripts are public on GitHub. I am not affiliated with TypeSafe AI or the authors of basal, so the model is chosen for your case.
How to start
Book a short call and tell me which process has the most repetitive decisions. We agree on timing and price on that call. I run the audit myself, with the Prosit team for implementation.
Book a 30-minute call →