About This Episode
In Stratola Spectrum Episode 2, Dinesh Chandrasekhar, Chief Analyst, Stratola, speaks with Shuveb Hussein, CEO and co-founder of Unstract, about the next wave of Intelligent Document Processing and why unstructured data is still the biggest gap in the modern data stack.
Shuveb frames the core problem as structural: modern data stacks handle structured data well, but enterprises still struggle to operationalize unstructured documents at scale. He positions Unstract as “unstructured data ETL,” meaning you can drop documents into storage like S3 and get structured JSON out, delivered into systems like Snowflake, Redshift, or BigQuery. Instead of solving workflow automation end-to-end, Unstract focuses on the foundational layer: turning messy document variants into a consistent schema that engineering teams can use downstream.
They discuss where demand is strongest: regulated industries that ingest documents from outside their control. Shuveb says most traction comes from BFSI (banks, financial services, insurance), followed by healthcare, with a long-tail across other verticals.
The episode also dives into why LLMs disrupt traditional IDP. Classical IDP and OCR workflows often require manual annotation and struggle with real-world variability. Shuveb argues LLMs behave more like humans: they can read a document without needing fixed coordinates, and directly output JSON against a target schema, even when formats vary across hundreds of templates (example: bank statements from 100 different banks).
But LLM-based document extraction introduces new constraints: input quality and preprocessing matter, hallucinations exist, and cost can compound at scale. Shuveb explains that before the LLM step, documents often need robust text extraction and layout-aware representation (tables, checkboxes, radio buttons). To reduce hallucination risk, he describes a consensus approach: using multiple models to converge on an answer, preferring null output over wrong output.
For large documents (60–100 pages common in insurance and mortgage), they discuss context window limitations and the tradeoff between full-context extraction (higher accuracy, higher cost) versus chunking with retrieval (lower cost, potentially lower accuracy). Shuveb’s stance is business-value driven: compare AI cost against the human cost and latency it replaces.
They close on data security and enterprise adoption barriers, including contractual restrictions against sending customer data to external LLMs, plus approaches like PII masking/anonymization and running models inside cloud boundaries (Bedrock, Vertex AI, Azure OpenAI) or via self-hosted open-source models. Shuveb notes a practical gap: open-source models often underperform for complex extraction workloads compared to frontier models.
Key Takeaways
Unstract positions itself as “unstructured data ETL,” converting messy documents into structured JSON that can land in Snowflake, BigQuery, Redshift, etc. The focus is structuring and schema normalization across many document variants, not downstream automation.
The biggest traction is in regulated industries that cannot control inbound documents. Around 60% BFSI and 25% healthcare, where banks, insurers, and providers receive highly variable external paperwork.
LLMs are transforming IDP by reducing dependency on manual annotation and rigid templates. Like humans, LLMs can “read” for meaning across formats, enabling extraction even when layouts differ widely.
Accuracy depends heavily on preprocessing and document representation before the LLM step. Extract clean text, preserve layout, and represent tables, checkboxes, and form elements in a model-friendly way to improve results.
There is a clear tradeoff between cost and accuracy as document length grows. Sending full context boosts accuracy and can still be cheaper than manual work; chunking with RAG reduces cost but may reduce accuracy, so techniques like prompt consolidation, summarization, and token shaving matter at scale.