Skip to content
Tarka
Rows of aged claim folders in a filing drawer

38%,

Rebuilding claims intake around Document AI and rules automation

Document AI and rules automation for a national health insurer's claims desk.

The context

The claims desk handled around 14,000 claims a week. Most arrived as scanned forms, discharge summaries and invoices, and every one was keyed in by hand before an assessor could look at it. Average handling time had crept up each year, and backlogs peaked every open enrolment period.

The insurer could not pause operations to rebuild the process. Any change had to sit beside the existing claims system, keep assessors in control of every payment decision and meet the regulator's record-keeping rules. Earlier automation attempts had stalled because nobody could say what they had actually improved.

Hands pressing a stamp onto a form beside a stethoscope

What we did

  1. 1

    Measure the baseline

    Before writing code we spent three weeks timing claims end to end. We instrumented the claims queue, sampled 1,200 files by hand and split handling time into intake, keying, checks and assessment, so every later gain had a fixed reference point.

  2. 2

    Document AI extraction

    We trained extraction models on the insurer's own forms, invoices and discharge summaries. Each field carries a confidence score: high-confidence fields go straight through, and low-confidence fields are routed to a keying screen with the source image alongside.

  3. 3

    Rules as code

    We moved policy and eligibility checks out of assessor checklists into a versioned rules engine. Simple, complete claims are prepared for one-click approval; anything unusual reaches an assessor with the failed rule and the evidence already attached.

  4. 4

    Phased release

    We released by claim type, starting with outpatient claims, and ran each phase in parallel with the manual process for two weeks. A phase went live only when its handling time and error rate beat the baseline.

The result

38%.

Faster claims processing

Average claims processing time fell 38% against the measured baseline, from 6.5 working days to 4. We tracked it on the same queue instrumentation we set up in week one, so the comparison covers identical claim types, volumes and working hours.

The real test was open enrolment, when weekly volumes rise by about 40%. The gains held through the peak without extra temporary staff, and assessors now spend their time on the complex claims that need judgement rather than on keying.

average claims handling time, down from 6.5
4 days
of fields extracted without manual keying
70%
extra temporary staff through open enrolment
0
“Tarka measured our claims baseline before writing a line of code, shipped in phases, and the gains held through open enrolment. That discipline is rare.”
Portrait of Laura BrennanLaura BrennanVP Claims Operations, a health insurer

Have a number you need to move?

Tell us what it is. We will tell you what we would measure, and how long a discovery would take.

Talk to Our Team