Cost Per Claim: AI in Healthcare and Insurance, By the Numbers

In healthcare and insurance, the unit of work is the claim: a health insurance claim for a medical service, an auto insurance claim for an accident, a workers' compensation claim for an injury. Measuring cost per claim adjudicated is the foundation of understanding whether AI is actually improving margin in these operations.

A mid-sized health insurance company might process 50,000 health claims per month. An AI system that can adjudicate routine claims (determining whether a claim is covered, calculating the patient responsibility, sending approvals) saves money if the cost per claim (including all hidden costs) is less than the cost of a human adjuster. A human claims adjuster costs roughly $75,000 per year in salary plus benefits, plus another $30,000 in overhead, for a fully burdened cost of ~$1.40 per claim (assuming 75,000 claims processed per year per adjuster). AI should be cheaper.

The Full Cost Stack for a Claim

Consider a $4,200 health insurance claim for a knee surgery. An AI system needs to:

1. Look up the claim history and coverage rules (embedding lookup in the insurance company's database): $0.08
2. Run fraud detection integrations (checking for duplicate claims, comparing against known fraud patterns, calling external fraud-detection APIs): $0.12
3. Make the initial adjudication decision (API call to Claude or GPT-4 to evaluate coverage): $0.18
4. Perform prior authorization checks (if the procedure requires pre-approval, integrate with the provider network): $0.10
5. Generate the explanation of benefits (EOB) (create a customer-facing letter explaining the decision): $0.06
6. Compliance review and sign-off (a human compliance officer reviews the decision, especially for edge cases or high-dollar claims): $0.45
7. Logging and observability: $0.03

Total cost per claim: $1.02.

If the human adjuster baseline is $1.40 per claim, the AI system saves $0.38 per claim, or about 27%. For 50,000 claims per month, that's $19,000 in monthly savings.

But the math changes if you adjust variables. If you increase human review to 20% of claims (to handle denials, complex coverage scenarios, high-dollar claims), the human review cost jumps to $0.90 per claim, and your total cost per claim rises to $1.53—now you're losing money. If you reduce human review to 5%, your cost per claim drops to $0.68, and you're saving $0.72 per claim.

The ratio of automated claims to human-reviewed claims is the biggest lever in the cost model.

Real-World Examples: Three Common Scenarios

Scenario 1: Routine health claims at a large health insurance plan. Claims: 80,000 per month. AI auto-approval rate: 65%. Human review: 35%. Cost per claim: $0.15 (API and embeddings) + $0.08 (fraud detection) + $0.03 (observability) + ($0.45 * 0.35) = $0.45 per claim. Baseline human cost: $1.40 per claim. Savings: $0.95 per claim * 80,000 = $76,000/month savings.

Scenario 2: Auto insurance claims at a regional insurer. Claims: 12,000 per month. AI handles liability assessment and fraud detection. Most claims still require human adjuster review (higher complexity). AI auto-approval rate: 25%. Human review: 75%. Cost per claim: $0.22 (API, embeddings, integrations) + ($0.50 human review * 0.75) = $0.59 per claim. Baseline human cost: $1.40 per claim. Savings: $0.81 per claim * 12,000 = $9,720/month savings.

Scenario 3: Prior authorization (pre-approval) workflow at a health plan. Prior auths: 5,000 per month. AI pre-determines which requests are likely to be approved based on medical necessity guidelines. All require final human sign-off. Cost per prior auth: $0.20 (API and coverage lookup) + $0.50 (human review, since all are reviewed) = $0.70 per prior auth. Baseline human cost (if done manually without AI): $1.50 per prior auth. Savings: $0.80 per prior auth * 5,000 = $4,000/month savings.

In all three scenarios, the AI system's value comes from handling the volume of routine work efficiently, while human reviewers focus on edge cases and high-complexity decisions. The cost per claim is lowest when the AI auto-approval rate is highest, but that auto-approval rate is constrained by risk tolerance and compliance requirements.

The Hidden Cost: Compliance and Fraud Risk

One cost that often gets overlooked is the cost of error. If an AI system incorrectly denies a valid claim, the insurer faces an appeal, potential regulatory scrutiny, and reputational damage. Some companies factor in an error reserve: 0.5% of denied claims get appealed and overturned, costing ~$100 in reprocessing and overhead per appeal. For 50,000 claims per month, that's $250,000 in error reserves.

This is why human review of high-dollar claims is non-negotiable. A $4,200 knee surgery claim that the AI incorrectly denies is expensive to fix. A $85 routine office visit that the AI denies is less costly to appeal. Many health insurance companies scale human review by claim dollar value: 100% review for claims over $5,000, 10% sample review for claims under $1,000, and risk-based review for the middle band.

This changes the cost model: instead of a uniform human review percentage, you use a threshold-based approach that keeps compliance risk manageable while minimizing cost per claim.

Comparing Against Your Current Baseline

To evaluate whether AI is worth deploying in your claims operation, compare the cost per claim to your current baseline:

Current process baseline: $1.40 per claim (human adjuster, fully loaded). AI-enabled process: $0.45 per claim (with 65% auto-approval, 35% human review). Savings: $0.95 per claim. For 500,000 claims per year: $475,000 in annual savings.

Adjust these numbers for your operation: your human baseline might be $1.20 or $1.60 depending on geography and staff productivity. Your AI auto-approval rate might be 40% or 80% depending on your claims mix and your risk tolerance. But the framework is the same.

The Economics: Where AI Makes Sense

Not every claim is economical to process with AI. A routine health claim for a $100 office visit has a low complexity and a quick human processing time. AI might save 30 seconds (worth maybe $0.04 in labor), which doesn't justify the full cost stack ($0.65 in AI). But a complex claim for a $25,000 surgery with prior auth requirements, coordination of benefits, and fraud-check integrations might take a human 15 minutes to fully adjudicate. AI that handles it in 2 minutes saves 13 minutes ($3.25 in labor), easily justifying the full cost.

This is why many insurance companies don't automate 100% of claims. They use AI for routine, high-volume claims where the cost per claim is below the labor savings, and they route complex claims to humans. This creates a tiered processing model:

• Tier 1 (Auto-approved, no human review): Routine claims, low dollar value, clear coverage. 60% of volume. Cost per claim: $0.35. Fully automated, no human touch.
• Tier 2 (Human review optional): Moderate complexity, edge cases on coverage. 30% of volume. Cost per claim: $0.65 (AI + partial human review). Routed to human only if AI is uncertain.
• Tier 3 (Human primary, AI assist): Complex claims, high dollar value, multiple coverage questions. 10% of volume. Cost per claim: $1.80 (mostly human, AI assists with documentation and data lookup). Human adjuster primary decision-maker, AI gathers context.

This tiered approach is more nuanced than "use AI for everything" or "use AI for nothing," and it's where real economics come in.

Vendor Benchmarking

When evaluating AI vendors for claims adjudication (or any claims processing), ask them for cost per claim and insist they break it down: API cost, integration cost, human review cost, compliance cost. If they quote only API cost and call it done, they're not being honest about the full economic picture.

Also ask: What's your accuracy rate on auto-approved claims? What's the human audit rate (what percentage of auto-approved claims do humans spot-check)? What's your fraud detection false-positive rate (what percentage of legitimate claims does your AI flag as fraudulent)? These quality metrics matter because a 99% accurate auto-approval system has different economics than a 95% accurate one. The 5% tail error becomes a cost (rework, appeals, regulator complaints).

Smarter Technologies, a leader in healthcare claims automation, has published work on cost-per-claim models that shows how to structure the trade-off between automation rate, human review, and quality. Their approach aligns with the Runrate framework: measure the full cost stack, connect it to business outcomes and quality metrics, and use that data to make vendor and process decisions.

To learn more about how to structure cost per claim allocation across business units or customer segments, see the article on AI chargeback and showback or return to the pillar article on AI cost attribution.