the mechanics of business and real-world applications of machine intelligence

Key Takeaways

Historically, much of the automation/augmentation in insurance has occured in consistent, structured types of risk (e.g. auto). At a reinsurer like Swiss Re, risk is complex, inconsistent, and global in scale; LLMs are a major unlock. And this is just one of many learnings from Luca.

Topics Covered

• Luca's background PhD and two post-docs, with a focus on the core math of machine learning 00:12
  • Transitioned into startup for 3.5 years building systems to analyze aerial imagery 00:26
  • Then four years ago joined Swiss Re, a large global reinsurance company, and have since been working to augment our products with data & AI 00:48
• Background on Swiss Re's business 01:43
  • Difference between insurance (e.g. home insurance, dental, auto) what we call "personal lines" and reinsurance 01:54
  • Some of these insurers need to offset or hedge their risks by "ceding" risk to a reinsurer, which has a much broader and diverse global portfolio, so if an earthquake strikes California, it can be compensated by earthquakes not happening in Japan at exactly the same time 02:11
  • Swiss Re is structured along 5 major business units 03:02
    • Life & Health Reinsurance 03:05
    • P&C (property & casualty) Reinsurance, which includes commercial properties, natural disasters, etc. 03:13
    • Global Clienting Solutions, which caters to our largest global clients, through which we provide services/solutions for fees 03:23
    • Public Sector Solutions, which is a mix of solutions, but also risk transfer for public entities, so e.g. the Government of California wants to get insurance to cover the losses of earthquakes that are not covered by private insurance 03:39
    • Corporate Solutions, which is a primary insurance business that caters to bespoke products for large companies 03:56
    • FInally we have a smaller homegrown company that builds whitelabeled insurance technology platforms that caters to B2B2C 04:22
  • Sitting on top of the business units and the "group functions" -- I sit in digital & tech, then you have HR, legal & compliance, risk management, etc. 04:42
• What are the primary functions in a given business line? 05:14
  • Product development, figuring out new products or adjustments to existing products in e.g. flight delay insurance, pet insurance, mobile phone insurance 05:22
  • Distribution is another major area, but it's less important for us because we are B2B; the primary insurers who are our clients are focused on the distribution/marketing of products -- that's the bulk of their activities 05:44
  • Underwriting -- there needs to be an underwriter or underwriting system that assesses the risks and determines if its a risk we want to insure and at what price 06:14
  • Underwriting is linked tightly to costing & pricing -- this is where the actuaries are doing a lot of work in making cost estimates of different types of risk & coming up with a market competitive price 06:31
  • Then claims... you need to look at all claims in a manual or automated manner, and compensate/adjust for each of them 06:49
  • Then there are auxilary functions around contracts management, and advisory roles... this is becoming a key differentiator for reinsurers, guiding companies into more healthy behavior, or otherwise better understanding risk to get coverage where you need it 07:07
• Where's the bulk of the headcount in the business? 07:51
  • For primary insurance, there's a lot of effort into distribution & sales -- it's very competitive, particularly for standardized product lines 08:10
  • It depends on size... some insurance companies are just carriers, distribution pre-defined products, applying underwriting rules, maybe they leverage reinsurance to do underwriting and claims settlement 08:38
  • The largest primary insurers have their own R&D, underwriting, and claims teams, so there the heacount is way more distributed across different functions 08:59
• What does a claim between an insurer and reinsurer look like? 09:45
  • There are two types of reinsurance 09:57
    • Treaty -- where your primary insurer cedes a whole portfolio of risk, for example all their auto policies in a given region, which may include thousands, or even millions of policies... if there is a natural catastrophe event, e.g. a hail storm, the reinsurer will step in if the insurer has a cash deficit, and in that case we will assess the event 09:59
    • Facultative, or "single risks" -- e.g. a primary insurer wants to insure a large manufacturing plant and its too big of an individual risk, so they want to split it apart; often in this case the insurer need the reinsurers' expertise to underwrite and evaluate the claims 10:48
  • Mailboxes through which we receive submissions from brokers via email, with documents attached the describe the property or individual live; that's still a big chunk of software we use 12:06
  • Specific software around underwriting, which is effectively a database, but that helps with underwriting decisions, and in more standard risk will do the pricing decisions 12:25
  • Similarly for claims, there's a system that ingests them, and ultimately puts them into data lakes & warehouses 12:56
• When did we start transitioning from paper pushing into the first wave of digitization? 13:32 re are different timelines depending on the product; very off the shelf/standard products were digitized first, really around the internet enabling people to purchase them 13:41
• Fair to say that Swiss Re is involved primary in less standard types of risk given focus on reinsurance and complex primary insurance? And therefore less automated today? 14:54
  • Fair statement, but also what we've been addressing... speed matters & we'd like to move more quickly without compromising on underwriting and pricing 15:17
  • More complex lines is where digitization is struggling; there so much heterogeneity in the data sources and types of risk 16:03
• What's the role of your team? 16:25
  • Two categories -- scaling decisions & improving quality of decisions -- both depend on the quality of the data we get in, much of which is unstructured in documents 16:26
  • That's really the biggest challenge for primary and reinsurance; it's a document & information based system 16:56
  • Things come through normal documents, transcriptions, phone calls, and often in these complex risks, you don't have claims coming in every day, but you do have a very large portfolio over a very long time horizon you can leverage 17:28
  • That said, we can get data from primary insurers, so we can e.g. get information on auto policies, but this is an easy problem 18:09
  • Let's think more about insuring flights/airplanes -- there are thousands of flights every day, but there are very few plane crashes -- how do you insure these tail risks, very large, but very rare events? 18:26
• How do you address this problem of underwriting/pricing tail risk? 19:15
  • Alternative datasets can help, but you still need to correlate them & number of events still limits you; this is where we're trying to really integrate/blend human judgement and expertise with data-driven approaches 19:20
  • Purely data driven approaches for rare events tend to fail; underwriting at some point for this very complex risk becomes a bit of an art 19:39
• Breaking down data extraction into 1) extracting known data fields into structured systems 2) extracting information on a bespoke/one-off basis 20:22
  • The first one is not perfectly solved, but it's much better solved in these structured, well-understood products 21:01
    • Though in providing consumer flexibility e.g. "take a picture of your car to submit your auto claim", you need to deal with some unstructured data 21:20
    • Identifying fraud here is also a huge issue; it's very possible now that consumers will use e.g. Stable Diffusion to create deepfaked images of their car, to add damage, and then submit that as a claim 21:49
    • Is there still work required on the document OCR / ingestion piece? 23:00
    • Yes because these documents can be crazily heterogeneous and complicated; there might be information about multiple pages, tables split across pages, in e.g. medical reports you may get a bundle but you only care about one page noting the cause of death 23:06
  • On category 2, extracting signal fr om unstructured data, what opportunities do you see there? 23:48
    • This is where recent advances with LLMs and genAI seem to be opening a breach 24:02
    • We get submissions for very complex risk describining the manufacturing plant or different sites; these documents can run in the hundreds of pages and the underwriter may be looking for something very specific 24:12
    • There's a real risk of fatigue reading through 200 pages 24:33
• Have you observed post ChatGPT that your colleagues were starting to use these tools organically? Or have you tried to disallow this behavior? 25:13
  • People are experimenting, but within a very conservative boundary 25:50
  • Some of the things we've tried within these boundaries are quite exciting 26:11
• How exactly have you seen this tech changing workflows? 26:23
  • Technology should be there to augment decision-making, not replace it; humans will be in the loop; GPT should not decide if you should receive health insurance or life insurance 26:25
  • We just want the models to help humans find information they need, with citations 26:48
• How have you thought about implementing these models? What challenges have you run into? 27:27
  • One reason the models are so powerful is because they can address several tasks; you don't need to maintain multiple models for e.g. contracts, engineering reports, etc.; this reduces the "MLOps" effort 27:56
  • Fine-tuning necessitates you maintaining multiple fine-tuned versons, and at this point the return of doing that is not so clear 28:37
  • You get similar advantages just by doing proper prompts 29:10
• Are these tools something you'd like to buy externally or build in-house? 29:21
  • It's a challenging question; there's not a general recipe 30:01
  • Building comes with risk because you can't just build, you need to think about security aspects, and broader workflow -- you need the broader system, not just the data science piece 30:17
  • For smaller companies, it's easier to just buy; often times insurers will go through us to find these capabilities 30:36
  • Important to think about what our advantage is, particularly when you have so much available in the open source community 31:11
  • You also want to be careful about creating too much dependency with a single vendor 31:34
  • DocQA players today really just look at extracted text sequentially; have you seen anything with more of a visual or spatial understanding? 32:04
  • LLMs we've used today are just looking at a very large sequence of words, but we've spoken to a couple emerging vendors that have the visual understanding of the layout of a document, which is very important 32:40
• Do you think that the document-based Q&A format will hold as the long-term form factor? 33:49
  • It's both this and the structured data extraction; both will exist 34:18
  • There are always going to be these complex cases that require human reasoning & in these cases the doc Q&A format is very helpful 34:45
• Zooming out from doc data extraction in underwriting/claims, are there insights you can add in other areas e.g. new product development? 35:53
  • Two underlying questions... 1) how do we use AI/ML to do product development and 2) how do we feed back insights from claims/writing into the products? 36:37
    • On #2, you go through this review process on annual or biannual basis to fine-tuning parameters in your costing models. This is where you factor in changes in e.g. population longevity, safety standards in cars, etc. 36:56
      • Often times you are using some sort of data science on claims or additional data sets to better understand trends; it is helpful here 37:57
    • On #1, where you don't have prior information / it's a new product, it's often just more standard market sizing / marketing research 38:13
      • The question is ultimately "does this new data source provide an edge, relative to its cost" 38:41
      • Often times we identify new data sources that are easier to tap into that are cheaper or allow an easier customer experience (reducing questions from 20 to 10 when submitting an application) 39:16
• Are there other major categories or excitement or time investment for your team that we should be talking about? 40:05
  • Governance is very important; not everything you can do should be done 40:15
  • Key that we understand regulatory landscape, civil society sentiment; it's key that these models have a positive impact for the end clients (the policy holders) 40:38