Amos

DemosClaims desk

Claims deskEach claim is a damage photo, the customer's note and the policy that applies. Amos answers six questions in one call, boxes the damage on the photo, and gives every answer a probability. Pick a claim to see it next to GPT-6.1 Sol.

recorded run

  • Claims

    8

    six questions each, one call

  • Agreement with Sol

    79%

    38 of 48 answers

  • Median time per claim

    3.41 sSol 5.57 s

    round trip, both called from one machine

  • Cost per 1,000 claims

    $2.21Sol $5.55

    Amos on its own GPU, Sol at Bedrock prices

Amos 1 Large through the public API, GPT-6.1 Sol on Bedrock, each given the same photo, note, policy and options. Amos's cost is its H100 at $6.88 an hour over the 0.863 claims a second it served on that GPU. Sol's cost is $2.00 in and $10.00 out per million tokens over the 1,913 in and 172 out it used per claim. Boxes agree when they overlap by half or more.

Phone dropped on the pavement
Amos, sure 0.69Sol, dashed

Photo: Sygmoral, CC BY-SA 4.0, via Wikimedia Commons

C-2041

Phone dropped on the pavement

Customer note

My phone slipped out of my jacket pocket onto the pavement on Saturday. It still turns on, but the glass is cracked across the top of the screen.

Policy excerpt

Gadget cover. Accidental damage to the insured phone, including a cracked screen, is covered anywhere in the world. Excess: $50 per claim.

Amos 1 Large · 3.04 sGPT-6.1 Sol · 4.32 s
  • Covered by the policyYes0.97Yes0.99
  • SeverityModerate0.68Moderate0.83
  • What was damagedElectronics0.96Electronics1.00
  • Fraud signalsNone0.93None0.96
  • Next stepApprove0.84Approve0.94
  • Where the damage isBoxed0.69Boxed, overlap 0.56
The request each claim gets: a photo, six questions, one callShow
POST /v1/systemone
{
  "model": "amos-1-large",
  "state": "Claim … Customer note: … Policy excerpt: …",
  "images": [
    "<the photo, base64>"
  ],
  "questions": {
    "covered": {
      "type": "noul",
      "instructions": "Is this damage covered under the policy excerpt?"
    },
    "severity": {
      "type": "score",
      "instructions": "How severe is the damage in the photo?",
      "criteria": [
        "Cosmetic",
        "Moderate",
        "Severe",
        "Total loss"
      ]
    },
    "item": {
      "type": "choice",
      "instructions": "What was damaged?",
      "criteria": {
        "vehicle": "A car, scooter or other vehicle",
        "building": "Part of a home or building",
        "electronics": "A phone or other electronic device",
        "parcel": "A parcel or the goods shipped in it",
        "clothing": "Shoes or clothing"
      }
    },
    "fraud": {
      "type": "multi",
      "instructions": "Which fraud signals does this claim show?",
      "criteria": {
        "story": "The note's story is inconsistent or implausible",
        "wear": "The damage looks like pre-existing wear or long use, not one event",
        "mismatch": "The photo does not match what the note describes",
        "none": "None; the claim looks genuine"
      }
    },
    "next": {
      "type": "choice",
      "instructions": "What should the claims desk do next?",
      "criteria": {
        "approve": "Approve and pay the claim",
        "adjuster": "Send an adjuster to inspect",
        "photos": "Ask the customer for more photos",
        "decline": "Decline the claim"
      }
    },
    "where": {
      "type": "region",
      "instructions": "Where is the damage in the photo?"
    }
  }
}