Teaching AI to Read Compliance Papers, Not to Decide
A two-stage AI pipeline reads SMS compliance documents field by field, while a deterministic rule engine — never the model — makes every approve, reject or review decision.
To send SMS in Türkiye, brands must prove sender identity, authorization, and operator agreements through a messy pile of documents — clean PDFs, blurry phone photos, bilingual text, and decades-old scans with handwritten signatures. A team replaced the manual review of these documents with an AI pipeline, but deliberately kept the model out of the decision-making loop: it only perceives, never judges.
The architecture splits perception from decision. A vision model extracts structured fields — company name, sender ID, date, signature presence — each backed by a verbatim quote from the source page, with no verdict field in its output schema at all. A separate, deterministic rule engine then evaluates these fields against critical and non-critical rules, collecting every result instead of stopping at the first failure; any critical failure rejects, any non-critical issue routes to manual review.
This separation matters to engineers because it preserves explainability, reproducibility, and testability — properties that can't be sacrificed in compliance-adjacent systems. The shared perception layer scales to new document types or providers through configuration rather than retraining, and the resulting system treats 'manual review' as a first-class third outcome alongside approve and reject, rather than forcing every case into a binary pass/fail.