How AI Uses Data to Predict Outcomes: From Car Accident Claims to Business Decisions

Someone gets rear-ended on the way to work. A few weeks later, a settlement offer shows up. Most people picture an adjuster reading their file and picking a fair number. In many cases, though, software has already weighed in before that offer goes out.

For anyone studying data science or AI, this is one of the clearest real-world cases of a model shaping a decision that lands in a person’s mailbox.

How Common Is AI in Auto Insurance Claims?

More common than most drivers realize. In 2022, the National Association of Insurance Commissioners (NAIC) published a survey of 193 private passenger auto insurers. It found that 88.6% of them were using, planning to use, or exploring AI or machine learning.

Claims was one of the busiest areas. 135 companies reported AI/ML models in use for claims work. Here’s how they used them:

• Informational resource for adjusters: 96 companies
• Claim assignment decisions: 58 companies
• Image evaluation (such as vehicle damage photos): 55 companies
• Settlement amount determination: 50 companies
• Claims approval: 9 companies
• Claims denial: 0 companies

There’s a detail here that data students will appreciate. The NAIC’s definition of AI/ML left out traditional statistical models like generalized linear models. Older rules-based valuation tools don’t show up in these numbers either. So the real footprint of automated decision support in claims is probably wider than the survey shows.

Where It Started: Rules Engines Like Colossus

Long before anyone talked about generative AI, insurers were using software to standardize injury valuations. The most widely discussed example is Colossus. According to the American Bar Association, Computer Sciences Corporation developed it to calculate settlement recommendations for bodily injury claims, and Allstate first deployed it in the 1990s.

The basic mechanics look like an expert system:

• The adjuster answers prompts generated by the software about the injury, the treatment and the prognosis.
• The system matches those inputs against its rules, often called “value drivers.”
• It produces a recommended range, which the adjuster can accept or override.

The ABA’s analysis points out some important limits. Items like lost wages and out-of-pocket costs had to be entered by hand, and so did any percentage of fault assigned to the injured person. Garbage in, garbage out applies here like anywhere else.

Transparency has been the long-running criticism. One insurance coverage attorney describes the proprietary value drivers, “reportedly numbering in the thousands,” as “hidden behind trade-secret protections” (Houston Harbaugh, 2026). Starting in 2005, class actions such as Hensley v. Computer Sciences Corp. alleged that insurers used the software to systematically undervalue injury claims. Several of those cases ended in settlements, which are not findings of wrongdoing. Courts have viewed the tool differently, too. In one case the ABA cites, a review found no evidence of improprieties but recommended stronger management oversight.

How a Claims Model “Reads” an Injury

First, it helps to know whose model this usually is. An injury claim is made against the person who caused the crash, but in many cases it’s their insurer that evaluates and pays it. So the software scoring your injury typically belongs to the other side. Think of it as a feature set. A claims model doesn’t feel your neck pain.

It sees structured data.

Medical bills and records use standardized code sets, like ICD-10-CM for diagnoses and CPT for procedures. These codes, along with dates of service, provider types and treatment notes, become the model’s view of the injury. The above-mentioned NAIC survey found that 63 insurers used medical data in their claims models. Vehicle data (123 insurers) and telematics (21 insurers) showed up as well.

This is where a basic data principle becomes very personal. Missing data isn’t neutral. A model can’t impute what nobody wrote down. Here’s how that can play out:

What the person lived through	What the file might show
Weeks of pain that got worse at night	One ER visit, no follow-up notes
Stopped physical therapy because of work hours	A “gap in treatment” with no explanation
Headaches mentioned to the doctor in passing	No headache diagnosis code
Two weeks of missed shifts	No wage documentation in the file

By its own logic, the model is doing its job. It scores the data it was given. When the record is thin, the output may not match what the person actually went through. That’s a big reason analysts and regulators keep pushing for human review and explainability.

Generative AI Is Changing Both Sides of the Table

The newer wave of AI goes beyond rules engines. Large language models can summarize hundreds of pages of medical records, flag inconsistencies, and draft correspondence. Insurers aren’t the only ones using them.

CCC Intelligent Solutions reported in August 2026 that bodily injury made up 52.3% of combined bodily injury and vehicle damage payout dollars in 2025, up from 44.4% in 2022. The same report cited Thomson Reuters data showing that 41% of surveyed law firms now use generative AI, up from 28% in 2025. CCC calls AI a “force multiplier” in turning records, bills, and photos into organized injury demands.

So the picture in 2026 looks like this. Models on one side summarize and score the claim. Models on the other side organize the evidence. What sits in the middle is the same thing it’s always been: the quality of the underlying records.

Who Checks the Algorithm?

Regulators have started to step in. In December 2023, the NAIC adopted a model bulletin on insurers’ use of AI systems (Sullivan & Cromwell summary). As of August 31, 2026, the NAIC lists 24 states and jurisdictions that have adopted it, including Massachusetts.

The Massachusetts Division of Insurance issued Bulletin 2024-10 on December 9, 2024. It expects insurers to maintain a written AI program with governance, risk controls, and internal audit, plus controls designed to reduce “Adverse Consumer Outcomes.” The key line is this: “Actions taken by Insurers in the Commonwealth must not violate M.G.L. c. 176D, regardless of the methods the Insurer used to determine or support its actions”.

Chapter 176D lists unfair claim settlement practices. Among them are refusing to pay claims “without conducting a reasonable investigation based upon all available information,” and failing to promptly explain the basis for “the offer of a compromise settlement”.

For future data professionals, this is where the career opportunity sits. Model governance, bias testing, documentation, and explainability are exactly what these rules ask for. If that interests you, BIA’s Legal Analytics course and Generative AI & Agentic AI program in Boston cover the skills behind this kind of work.

What This Means If You’re the Injured Person

There’s a practical takeaway here for anyone who has been in a crash. Getting checked out promptly, telling your doctor about every symptom and following through on treatment all help the record reflect what actually happened. Keeping track of missed work and out-of-pocket costs helps too. If a settlement offer still seems low, you can ask how it was calculated, and this guide on how to respond to a low settlement offer walks through common next steps. Every claim is different.

The Bottom Line

AI is already part of how claims get handled, and simpler versions of it have been around since the 1990s. For data learners, it’s a live case study in features, missing data, explainability, and regulation.

For anyone looking to build these skills, an artificial intelligence course can provide a foundation in areas such as machine learning, data analysis, AI applications, and model evaluation.

Frequently Asked Questions

1. How does AI help insurance companies evaluate car accident claims?

AI can analyze information such as medical records, vehicle damage images, claim details, and other structured data to identify patterns and support claims-related decisions. Depending on the system, AI may assist with claim assignment, image evaluation, settlement recommendations, or other parts of the claims process.

2. What type of AI is used in insurance claims?

Insurance companies can use different forms of AI and machine learning, including predictive analytics, classification models, computer vision, natural language processing, and rules-based systems. Newer applications may also use Generative AI to summarize documents and organize large amounts of information.

3. Why is data quality important in AI decision-making?

AI models depend on the information provided to them. Missing, inaccurate, or incomplete data can affect the model’s output. This is why data collection, data cleaning, feature engineering, model evaluation, and human review are important parts of responsible AI implementation.

4. What can I learn in an artificial intelligence course?

An artificial intelligence course can cover areas such as Python, data analysis, machine learning, deep learning, natural language processing, computer vision, Generative AI, model evaluation, and responsible AI. Practical projects can also help learners understand how these technologies are applied to real-world problems.

5. Can AI completely replace humans in insurance claims?

Not necessarily. AI can automate analysis and provide recommendations, but important decisions may still require human oversight. Human review is particularly important when decisions involve incomplete information, complex circumstances, fairness concerns, or significant consequences for individuals.