At Rise of AI 2019 in Berlin, Sofie Quidenus-Wahlforss, CEO and founder of omni:us, describes her company's journey applying narrow AI to insurance claims handling and argues that general intelligence remains far off, so startups must focus on a specific, measurable use case. She traces her path from a document-scanning robot she built in 2004, which digitized over 980 million documents, through the realization that most digital documents like PDFs are not machine-readable, to founding omni:us and eventually narrowing its ambition from a broad, horizontal document-understanding platform serving 5-10 industries to a vertical focus on a single process: insurance claim handling. She explains that while standard approaches extract 7 to 14 data fields from a claim, omni:us technology extracts around 200, enabling a shift from process-driven to data-driven insurance. She describes the technical architecture of a reusable, scalable training engine combined with an industry-specific natural language processing layer, and closes by describing the cultural difficulty, for both her company and its insurance customers, of pushing large-scale transformation.
Key takeaways
- Quidenus-Wahlforss's early document-scanning robot digitized over 980 million documents, becoming a key data asset for later AI work.
- omni:us narrowed its ambition from a horizontal, multi-industry document platform to a vertical focus on a single process: insurance claim handling.
- Standard claims processing extracts 7 to 14 data fields; omni:us's combined vision and NLP approach extracts around 200 data fields per claim.
- The company's reusable technical engine is about 95% transferable across domains, but achieving high confidence scores requires a domain-specific language layer built for each industry.
- omni:us can offer pre-trained models needing only a few hundred documents to fine-tune per customer, rather than requiring millions of new documents.
Detailed notes
From a document-scanning robot to narrow AI
Quidenus-Wahlforss recounts starting her first company 14 years before this talk, building a robot that automatically found the right pressure to turn a book page and used a camera and inter-robot communication to digitize documents; over 160 units were sold worldwide, and the robot eventually digitized over 980 million documents. This work taught her two lessons that shaped her career: access to unique training data is the real bottleneck for AI startups, and most digital documents, including PDFs, are not machine-readable despite appearing intelligent to humans. Reflecting on general AI, she notes that a system beating the world's best Go player still cannot play chess, and argues genuine general intelligence, applying one learning to an unrelated situation the way a toddler can, remains distant, while narrow AI trained specifically for one task is what is achievable today.
From horizontal ambition to vertical focus on claims
Quidenus-Wahlforss describes omni:us's early temptation to stay horizontal, claiming to understand documents across 5 to 10 industries and a 200-billion market, arguing the same underlying document types like invoices recur everywhere. She says this was a painful choice for a founder unwilling to let go of opportunity, but the company ultimately decided to vertically focus on the insurance industry and, within it, on a single process: claim handling, aiming to become the leading technology for that specific transformation. She raised 4.2 million on an early pitch built mainly on credibility and access to data, but notes investors now expect concrete numbers and process specifics rather than general AI claims.
Why insurance and why claims
She explains the company chose insurance because it affects everyone, yet claims processes are slow, opaque and frustrating for customers who expect the same digital speed they get from their smartphones. She cites insurtech peers and customers such as Wefox, Lemonade and Coya, competing alongside larger incumbents like Allianz and AXA, and describes the industry as moving from skepticism toward the vision omni:us proposed two years earlier to active demand. Standard document processing extracts 7 to 14 data fields from a claim; omni:us's combined computer vision and natural language processing approach extracts around 200 data fields, enabling a shift from a process-driven to a data-driven claims operation, with benefits in speed, transparency and product design, not just automation.
Technical platform: a scalable engine plus a domain layer
Quidenus-Wahlforss describes two components any AI startup product needs: a scalable, reusable technical engine for fast experimentation, training, retraining and ground-truth annotation, and an end-to-end product layer built on deep domain-language understanding, since natural language processing depends on the specific vocabulary of an industry. omni:us started with a dataset of over 30 million documents, letting it offer customers pre-trained models that need only a few hundred documents to fine-tune, rather than requiring new customers to supply millions of documents themselves. She mentions collaborating with the Computer Vision Center in Barcelona as an example of the research partnerships a startup needs since it cannot build everything like a research center.
Q&A highlights
Asked how much of the omni:us architecture would be reusable in a different domain such as electronic healthcare records, Quidenus-Wahlforss says the technical platform, the engine enabling fast experimentation and ground-truth annotation, is about 95% reusable, but achieving a confidence score above 95% requires training a domain-specific language layer for each new industry, which is why the company focused so narrowly rather than trying to serve multiple domains at once.
Quotes
Neural nets are out there since the 60s, and the bottleneck, especially for startups, is the access to unique training data.
Sofie Quidenus-Wahlforss
We give back time because we automate the process of claim handling.
Sofie Quidenus-Wahlforss
Welcome to the uncomfort zone at omni:us.
Sofie Quidenus-Wahlforss
Questions answered in this talk
Why did omni:us focus on insurance claims instead of a broader document-understanding market?
According to Sofie Quidenus-Wahlforss at Rise of AI 2019, omni:us initially considered staying horizontal across many industries but found that achieving high confidence scores requires a domain-specific language layer, so it chose to vertically focus on one process, insurance claim handling, within one industry.
How much more data can AI extract from an insurance claim compared to standard processing?
Sofie Quidenus-Wahlforss states at Rise of AI 2019 that while typical approaches extract 7 to 14 data fields from a claim, omni:us's combined computer vision and natural language processing technology extracts around 200 data fields.
Is an AI document-processing platform reusable across industries like healthcare?
Sofie Quidenus-Wahlforss says at Rise of AI 2019 that about 95% of the underlying technical engine is reusable across domains, but reaching a high confidence score in a new domain such as electronic healthcare records requires training a new, domain-specific language layer.