The most expensive conversation in insurance claims is the one where an assessor says 'this is hail damage' and the insurer says 'that looks like wear and tear.' Both parties are looking at the same photograph. Both are applying professional judgment. Neither can point to an objective metric that settles the question.
This is the dispute cycle, and it adds an average of 12 business days to claim settlement in the Australian property insurance market. Twelve days of back-and-forth emails, supplementary photo requests, second opinions, and internal escalations, all because two professionals disagree about what they see in a photograph.
What a confidence score actually means
When ARIS Detect analyses a roof image, every detected finding receives a confidence score between 0 and 100%. This score represents the model's statistical certainty that the area contains the identified damage type, hail impact, cracked tile, missing material, corrosion, or any of the other categories in the detection taxonomy.
A 96% confidence score on a hail damage detection means the model has identified visual features that match hail impact patterns with very high certainty. A 72% score means the features are present but less definitive, perhaps the angle is suboptimal, the resolution is lower in that area, or the damage pattern shares characteristics with another damage type.
Critically, the score is not the final word. The assessor reviews every detection, adjusts or removes false positives, and can add findings the model missed. But the score provides a shared numerical framework that both parties, assessor and insurer, can reference.
Changing the nature of the conversation
Before confidence scores, a typical dispute conversation looked like this: 'I inspected the roof and found 14 hail impact points on the north-facing elevation.' The insurer reviews the photos and responds: 'We can only confirm 8 of those as consistent with hail. The remaining 6 appear to be pre-existing deterioration.'
With confidence scores, the same conversation becomes: 'The AI detected 14 impact points on the north-facing elevation. 11 scored above 90% confidence. 3 scored between 70% and 85%. I reviewed all 14 and confirmed each as hail damage based on the following characteristics...'
The insurer can now make a risk-based decision. Do they challenge the 3 lower-confidence detections? Perhaps. But the 11 high-confidence detections are essentially unchallengeable, the AI model, trained on over 100,000 roof images, has identified them with near-certainty, and the on-site assessor has confirmed them.
Quantifying the impact
We analysed dispute rates across 12,000 claims processed through ARIS Detect in 2025 compared to equivalent manually-assessed claims from the same insurer portfolios. The findings were significant.
Claims where all damage findings had confidence scores above 85% experienced a 67% reduction in disputes compared to the manual baseline. Claims with mixed confidence scores (some above 85%, some between 60% and 85%) still saw a 41% reduction. Even the presence of the scores, regardless of their values, changed how disputes were initiated and resolved.
The reason is behavioural as much as technical. When an insurer sees a detection with a 96% confidence score, confirmed by an on-site assessor, the burden of proof shifts. Challenging that finding requires counter-evidence, not just a different opinion. The score creates an anchor that both parties negotiate around rather than starting from subjective positions.
The calibration question
Assessors frequently ask: 'What confidence threshold should I use?' This is the wrong question, because the threshold depends on the claim context.
For straightforward residential hail claims, many assessor teams set the AI Assistance slider to show all detections above 60% and manually review each one. This catches edge cases that might otherwise be missed and gives the assessor complete visibility into what the model found.
For complex commercial properties or disputed claims, some teams review every detection regardless of score, using the confidence value as one input alongside their own assessment. The score does not replace judgment, it informs it.
For catastrophe events with high claim volumes, teams often set a higher threshold (80%+) to focus assessor review time on the findings most likely to require human confirmation. Lower-confidence detections are flagged for secondary review if time permits.
Standardisation across assessor networks
Perhaps the most significant impact of confidence scores is on consistency across assessor networks. When an insurer engages a panel of 40 independent assessors, report quality and damage identification vary enormously. Some assessors are conservative; others are aggressive. Some photograph everything; others photograph selectively.
With ARIS Detect, every assessor starts from the same AI detection pass. The model does not have good days and bad days. It does not rush the last inspection of the afternoon. It processes every image with the same parameters and produces the same scores regardless of who uploaded the imagery.
This does not eliminate assessor variation, human judgment still plays a role in confirming, rejecting, and supplementing AI findings. But it provides a consistent baseline that narrows the range of variation across the network. When disputes do arise, both the assessor and the insurer can point to the same evidence, scored by the same model, as the starting point for resolution.
The path forward
Confidence scores are not a magic bullet for dispute resolution. Edge cases will always exist, unusual roof materials, ambiguous damage patterns, pre-existing conditions that the model has not been trained on. But for the 80% of claims where the damage is identifiable and the question is simply 'how much and how certain,' a numerical confidence framework reduces friction materially.
The insurers who have integrated ARIS confidence scores into their claims workflows report faster settlement times, lower dispute escalation rates, and higher assessor satisfaction. The assessors who use confidence scores report spending less time writing justification narratives and more time on the parts of the job that actually require human expertise.
Damon Smith
Founder, ARIS Detect