It's been a month since Kolar's first use case went into production. And between the idyllic vision of "AI that does everything" and a truly robust automation for compliance… there's a world of difference.
What we learned
Data is half the job.
Workflow quality depends entirely on input context.
In compliance, data is fragmented, unstructured, sometimes contradictory.
Takeaway
We spent as much time consolidating, verifying, and contextualizing as we did coding the automation itself.
Autonomous AI agents are sexy… but not in production.
Hallucinations, misinterpretations, inconsistent formats: great in demos, much less so when it touches regulated decisions.
Takeaway
We quickly stopped "putting LLM everywhere" to build deterministic and auditable workflows.
The real challenges are in edge cases.
Automating simple cases? Easy.
But compliance is mostly about cases that don't fit any pattern.
Takeaway
We spent 80% of our time on the 20% of atypical cases, where business logic is most nuanced and tricky.
Our conviction
In compliance, AI is not a magic wand.
It's an extremely powerful tool if you master workflows, data, and edge cases.
Do you work in AML/KYC and have painful processes?
Let's talk