01
Classify with a confidence score
Every asset is classified from the schedule, slip, SOV or loss narrative with a confidence score and an XAI trace attached to the output.
ClassifierIQLive
The assets, classified
— and remembered
A SOV might list 400 properties in twelve formats, and a slip uses shorthand every Lloyd’s broker understands and no generic model does. ClassifierIQ classifies the assets within, and remembers every correction your team makes.
<60sper classification
97.4%accuracy, live instance
100%of corrections retained
The reality
What this looks like
in most firms today
- A schedule of values arrives with 400 line items and no two describing an asset the same way.
- Classification depends on shorthand your underwriters know and no off-the-shelf model does.
- The same correction gets made by the same analyst every quarter, and is never captured.
The institutional knowledge that makes classification correct at your firm currently lives in three people’s heads. It leaves when they do, and it is unavailable to everyone else in the meantime.
Where the existing tools stop
Where tools stop
Generic classifiers apply a market-average taxonomy, which is not the taxonomy your pricing team actually uses.
Worse, they do not remember. Correct the same misclassification a hundred times and the hundred-and-first arrives identical.
What it takes
Classification is only useful when it converges on how your firm thinks about risk, which means every correction has to be captured and applied.
How it delivers value
How ClassifierIQ earns its place
in the workflow
02
Correct in a review queue
Low-confidence items route to a human queue. The underwriter or pricing analyst corrects the classification and the mapping is recorded at that moment.
03
Remember, and apply next time
The correction is stored in the memory loop. The next similar asset is classified using your firm’s mapping rather than a market average.
04
Class-specific agents you configure
Build, publish and manage classification agents per class of business, each tuned on your taxonomy and improvable through the same feedback loop.
Capabilities
The specification, without decoration
- Classification from schedules, slips, SOVs, loss notes and loss narratives
- Confidence score and XAI reasoning trace attached to every single output
- Human review queue for low-confidence items, with the threshold under your control
- Memory loop that captures and applies every human correction
- Batch history with batch ID, confidence, duration and full result drill-down
The approach
How it actually works
Written for the technical evaluator rather than the buyer. If you are going to put this in front of an architecture review board, this is the section they will read.
Extraction before classification
Line items are extracted and normalised from the source document first. Classifying un-normalised text is where most accuracy is lost, long before the model is involved.
Confidence scoring drives routing
Each classification carries a calibrated confidence. Items above threshold complete automatically; items below it queue for review. You set the threshold, and you can move it.
The memory loop is a mapping store, not a fine-tune
Corrections are stored as explicit asset-to-category mappings rather than folded into model weights. That keeps them inspectable, reversible and explainable to a reviewer.
XAI reasoning in plain English
Every result carries its understanding of the asset, the confirmation signals it found, the code mapping applied and a justification for the confidence score.
Analytics on accuracy over time
Live KPIs track schedules run, assets classified, hours saved and accuracy trend, so the improvement from the memory loop is visible rather than asserted.
The breakthrough
What changes once it is running
Accuracy and speed figures are measured on a live instance. The reduction range is modelled, because it depends on how much of your book the memory loop has seen.
97.4%classification accuracyMeasured on a live instance, up from 94.1% the prior month
<60sper classificationIncluding the XAI reasoning trace
40–60%less manual classification effortModelled, and improving as the memory loop fills
100%of corrections retainedEvery human correction captured as an explicit mapping
