The “Build vs. Buy” Matrix for Claims AI: When Building is Actually the Right Choice
If you are a claims executive, your inbox is likely overflowing with urgent pitches from vendors promising to “revolutionize your claims workflow” with AI.
The pressure to buy is real. You are watching competitors announce partnerships, your board is asking for an AI strategy, and your adjusters are drowning in administrative work. In this environment, waiting 18 months for internal IT to build a custom solution feels like a strategic failure.
This often leads to “shadow procurement”: business units buying software without IT oversight just to get moving. The hesitation from IT is also valid. They know that many of these platforms are actually just thin user interfaces sitting on top of public models like OpenAI’s GPT (also known as a GPT wrapper).
So, who is right? Neither. The conflict comes from treating every AI opportunity as the same type of problem. To solve this, you don’t need a ban on vendors. You need a decision matrix on when and how to use them.
The Decision Matrix: Renting Speed vs. Buying Expertise
When a new use case pops up, how do you decide whether to sign a vendor contract or put it on the internal engineering backlog? You should evaluate two variables: the complexity of the problem and the data advantage.
1. The Low Complexity “Utility” Zone: Buy
- Examples: Meeting transcription, email drafting, basic document summarization.
- The logic: Your ability to draft a settlement letter slightly better than your competitor is not a strategic moat; it is a commodity task. Speed is the only metric that matters here.
- The decision: Don’t waste expensive data science hours building a custom email drafter. Buy a secure, compliant wrapper that integrates with your email provider. Pay for the UI and get it deployed in weeks.
2. The High Complexity “Expert” Zone: Buy
- Examples: Litigation prediction, fraud detection, medical severity scoring.
- The logic: This is the “data gravity” trap. You could build a fraud model internally, but your model will only ever learn from your claims. A specialized vendor sees fraud patterns across 50 different carriers. Their model is smarter because their dataset is vastly larger than yours will ever be.
- The decision: Buy. Do not try to outbuild a vendor with a massive “data moat.” Partner with vertical AI companies that offer deep, domain-specific intelligence that you cannot replicate in-house. This isn’t a wrapper; it’s a specialized brain.
3. The High Internal Specificity “Idiosyncratic” Zone: Build
- Examples: Internal claim routing rules, adjuster assignment logic, proprietary settlement authority matrices.
- The logic: These problems rely entirely on your specific internal culture, legacy systems, and risk appetite. A vendor model won’t know that Steve in Florida handles high-touch VIP claims.
- The decision: Build. This is your “secret sauce.” Keep the logic that defines your specific operational DNA in-house.
Demystifying the “Wrapper” vs. “Vertical AI”
The danger in procurement isn’t buying software; it’s mistaking a wrapper for vertical AI.
- A wrapper sells you an interface; its value is in workflow enhancements.
- Vertical AI sells you a pre-trained, expert model; its value is in the domain intelligence it offers.
If a vendor claims to do litigation prediction but is just sending your case facts to OpenAI with a generic prompt, you are buying a wrapper at a vertical AI price.
However, if that vendor has a database of 10 million past court verdicts and uses that proprietary data to predict your case outcome, they are offering expertise. That is an asset you should acquire.
How to Vet the Vendors
Whether you are buying a wrapper for speed or vertical AI for expertise, you must vet the vendor to ensure security. Ask these three questions in the first conversation:
1. “What is your data retention policy with the model provider?”
You need a “zero-day retention” guarantee. This means that if the vendor sends your claim data to OpenAI/AWS/Azure for processing, the data is deleted immediately after the answer is generated. It must never be used to train the public base model.
2. “What is your proprietary IP?”
This is the litmus test.
- Wrapper answer: “We have great prompt engineering.”
- Vertical AI answer: “We have a proprietary consortium database of 50 million claims that trains our models (and/or grounds our predictions).”
3. “Are you model agnostic?”
The AI landscape is constantly changing. A good vendor abstracts the model layer, allowing you to swap out the engine (e.g., GPT-4 to Claude to Llama) as technology advances. You want to buy their application and their data, not be locked into their specific choice of LLM.
If they are selling “expert” prices but give “wrapper” answers, walk away.
Stop Fighting, Start Sorting
The tension between Claims Ops (“We need tools now”) and IT Security (“We need control”) disappears when you agree on this matrix.
Shadow Procurement happens when Ops feels like IT is the “Department of No.” By adopting a “Build vs. Buy” strategy, IT becomes a partner.
You can say: “We will buy this email tool because it’s a commodity. We will buy this litigation predictor because the vendor has better data than we do. But we will build the routing engine ourselves because that requires our internal logic.”
It is not about avoiding vendors. It is about knowing the difference between renting a tool and hiring an expert.
