DataFab //  artificial proprietary intelligence T+00:00:00 ISO/IEC 27001:2022 · SOC 2 Type II — certified Raw egress  0 B 

DataFab  /  Utilities  /  Insurance claim handling

Governed utility · legal-expenses and written-claim insurance

The written claim, worked end to end.

Extraction vendors tag the claim and hand it on — and two in three written claims still land on a caseworker’s desk. DataFab is the layer that finishes the job: it reads the claim and every document behind it, finds what is missing, checks coverage against the policy, assesses the prospects, checks the fees, and drafts the grounded response — over the systems you already run, with a caseworker on every decision that matters.

Read-only, in placePer-step calibrated confidenceSource per statementHuman in the loop
The problem · 01

“Extraction stops at the start. Two in three claims still reach a desk.”

The incumbent tags and extracts for a token per-case fee — and it is worth exactly that, because extraction alone cannot finish a written claim. Unstructured solicitor letters, coverage judgement and statutory fee-checking stay manual.

So a large majority of claims fall out to a caseworker — and a manually-settled case costs a large multiple of the extraction fee, fully loaded. A senior executive at a peer legal-expenses insurer puts that multiple in the tens, at the top of the range for document-heavy legal lines, and independently confirms the thesis: extraction does not complete the job. Letters, coverage judgement and fee-checking stay manual.

The economics · 02

The value is not the extraction slice. It is the fallout pool.

Every claim moved from manual to straight-through removes a manual case worth many multiples of the extraction fee. The right measure of any solution is one number: how far it shrinks the fallout rate. The break-even fee sits at many multiples of anything realistic, so the headroom is structural, not marginal.

Fig. 01  //  the fallout pool is where the cost lives● illustrative proportions
Written claimsthe full intake · all life-areasIncumbent extractiontags + extracts · token fee / caseextraction alone does not finish the job35%handled straight-throughroutine · rules-clear · complete65%still falls to a caseworkerat ~40x the extraction fee, fully loaded(peer-validated: 20–75x)the cost is not the extraction — it is the two-thirds that extraction cannot finish
ScenarioFallout falls toStraight-through rises toManual workload removedManual-handling cost
Conservative45%55%−31%−31%
Base case35%65%−46%−46%
Ambitious25%75%−62%−62%

Against a 65% starting fallout rate. Every ten points of fallout removed cuts the manual workload by roughly fifteen per cent. Illustrative model output from a worked example, not a quote and not a measured result.

Fig. 01a  //  the fallout pool, as the deck draws it◆ illustrative
Written claimsthe full intake · all life-areasIncumbent extractiontags + extracts · token fee / caseextraction alone does not finish the job35%handled straight-throughroutine · rules-clear · complete65%still falls to a caseworkerat ~40x the extraction fee, fully loaded(peer-validated: 20–75x)the cost is not the extraction — it is the two-thirds that extraction cannot finish
Fig. 01b  //  the floor and the combined-ratio north star
expense~29 bpsloss~57 bps~86 bpsThe board runs on thecombined ratio. The incumbenttouches only the expense half.The loss half — coverage accuracyand routing — is twice the size,and only a claim-understandinglayer can reach it.the loss half is ~2x the expense half — and only a claim-understanding layer reaches it

What it does · 03

DataFab works the claim — not just reads it.

One governed pass, from intake to a decision-ready file. Where information is missing, it drafts the reach-out. Where the rules are clear, it resolves. Where judgement is needed, it prepares the judgement — coverage points matched to clauses, prospects grounded in cohorts of similar claims, fees checked against the statutory schedule — and routes the file to a caseworker with everything attached.

01
per-field confidence

Reads & extracts

The claim and every document behind it — letters, forms, invoices, scans — read with a calibrated confidence on every extracted field rather than one score for the document.

02
missing info & reach-out

Finds the gaps

Completeness is checked against what the decision actually requires, and where something is missing the reach-out is drafted for approval.

03
policy points & exclusions

Checks coverage

Claim facts matched to the specific policy clauses and endorsements that apply. Where cover is partial, excluded or conditional, the reason is flagged and attached.

04
prognosis of success

Assesses prospects

Merits and likely outcome drawn from comparable resolved cases and the clauses in play — supporting the decision, not making it.

05
statutory fee schedule

Checks the fees

Solicitor invoices checked line by line against the statutory schedule. Over-charges, duplicates and out-of-scope items surfaced before payment, each carrying the rule it failed.

06
grounded

Drafts the response

The response is drafted with every statement traceable to the page and passage it came from.

07
a person decides

Prepares & routes

A decision-ready file with everything attached — grounded, attributable, defensible. The utility prepares the decision; a caseworker makes it.

The architecture · 04

One governed layer, over your systems.

Data stays in place. A governed fabric reads your systems read-only, document intelligence and similarity services ground each claim, and the existing decisioning core is adapted — with confidence, trace and governance attached to every decision.

Fig. 02  //  five tiers, read-only at the bottom◆ nothing copied out
ONE GOVERNED LAYER OVER YOUR SYSTEMS. THE DECISIONING CORE IS ADAPTED, NOT REPLACED.05Human surfaceswhere adjusters workAdjuster workbenchTabular compareLink analysis & OSINTMedia miningConversational dialogueNotification centreReach-outDrafting04Claims decisioning coreexisting platform, adaptedDecision engine & rule chainsReserves recommendationCase managementWorkflow engine & SLAFee & invoice checkingDeadline & limitation monitor03Intelligence & similarityunderstand each claimOCR & schema-bounded extractionPolicy point matchingCohort creationFraud & anomaly detectionProspects of success02Knowledge Fabricsystems integrationGoverned connector layer, MCPMetadata catalogue & claims ontologyEntity resolution to golden record01Your data sourcesdata stays in placePolicy administrationClaims and first-notification intakeDocument storesHistorical claims and decisionsCROSS-CUTTING — ATTACHED TO EVERY DECISIONConfidence engine · chain trace and evidence pack, append-only and replayable · governance: role-based access, operational modes, human review gates
The closed loop

When the decision engine flags missing or partial information, reach-out drafts the request; the claimant’s response re-enters through intake and the case is automatically re-evaluated — no manual re-keying.

Why it can be trusted · 05

Trust is the architecture, not a promise.

Claims teams have been burned by black-box systems with a blanket accuracy figure printed on the box. This is built the other way up — every output carries its own evidence, so a caseworker can verify any statement in seconds and an auditor can replay any decision end to end.

Calibrated

Per-step confidence

Not one score for the document — a calibrated confidence per extracted field and per reasoning step, so review effort goes exactly where the uncertainty lives.

Grounded

Source per statement

Every statement in every draft traces to the page and passage it came from. Nothing is asserted that cannot be shown.

Gated

Human in the loop by design

Coverage calls, prospects, declines and payments wait for a person. Autonomy is a dial set per step — and every gate is logged.

In place

Read-only

A governed connector reads your policy, claims and document systems where they live. Nothing is migrated, nothing copied out.

Explainable

Reason codes, not verdicts

The decision engine returns the rule chain and reason codes behind every recommendation — reviewable, contestable, explainable.

Current

Re-evaluates on new data

When the missing document arrives, the case re-runs automatically. The file is always current, never stale.

Live demo  //  document intelligence & OCR● pick a page, run the extraction
SOURCE OF RECORD — NOT MIGRATED
SOURCE OF RECORD — NOT MIGRATEDFLAGGEDFLAGGED
SOURCE OF RECORD — NOT MIGRATEDFLAGGEDFLAGGEDFLAGGED
Awaiting run
Mean confidence
Processing time
Backend used
Source SHA-256
Routing decision — resolved before dispatch
Fig. 03  //  confidence rolls up, never averages away◆ amber marks soft confidence
ONE CONFIDENCE PER DECISION. MANY PER CASE. ROLLED UP, NEVER AVERAGED AWAY.SIGNALS · RAW, PER ARTEFACTOCR field confidence0.96Document completeness0.92Entity match0.95Clause similarity0.83Endorsement scan0.79DECISIONS · ONE PER DECISIONExtraction accepted0.95Policy affiliation0.95Clause match0.81Cohort assignment0.88DETERMINATIONS · CLAIM LEVELCoverage: partial0.74Reserves estimate0.79OUTCOME · CASE ROLLUPWeighted by decision criticality — routed to a personHuman reviewAMBER MARKS SOFT CONFIDENCE — THE CASEWORKER SEES WHERE THE WEIGHT IS THIN BEFORE THEY DECIDE

The value, two ways · 06

A committed floor. A combined-ratio north star.

The floor · per-case efficiency · committed

Cost-competitive on day one.

  • Manual workload reduced by roughly a third to two thirds, depending on the scenario
  • The committed, forecastable floor — not the upside case
  • The fee it replaces is a fraction of the fully-loaded manual cost it removes
  • Provable on your own data during the pilot
  • The efficiency case alone justifies the deployment

The north star · combined ratio · where we steer

A combined-ratio partner.

  • The board runs on the combined ratio, and the loss half is twice the expense half
  • Better coverage decisions and better routing move the loss ratio itself
  • Directional — proven on your data, then shared in the gain
  • The difference between a software vendor and an underwriting-margin partner
  • It is why the pricing follows the ratio, not the case count

Directional figures are model output for illustration, not contracted. Magnitudes depend on your case mix, book and configuration.

What it unlocks · 07

The fabric, and what it unlocks.

The Knowledge Fabric is a governed foundation that reads across your systems and resolves them into one current, shared view — policies, claims, documents and history, connected through a single ontology and a golden record. Nothing is re-platformed into another silo; every process draws on the same grounded picture, so a decision made in one place stays consistent with every other.

The Knowledge & Agentic Studio sits on top of it — an environment to compose and run agencies, governed teams of agents each serving one purpose. A new process is assembled from reusable capabilities rather than built from scratch, and every decision inherits the layer’s confidence scoring, evidence trail and human-review gates. Each use case is configuration on the same foundation — defensible from day one, and faster to stand up than the last.

The first use case

Intake

Written claims are read, checked for completeness and assembled into a decision-ready file. Missing information is identified and the follow-up drafted automatically. The handler receives a case that is already reasoned.

Same layer

Compliance

Obligations under the applicable regime are checked against each matter as it proceeds. Breaches, gaps and required disclosures are surfaced with the rule they relate to. Every check is logged and evidenced, so an audit trail exists by default.

Same layer

Coverage checking

Every claim is matched to the specific policy clauses and endorsements that apply. Where cover is partial, excluded or conditional, the reason is flagged and attached. The handler opens a coverage view that is already reasoned, not raw.

Same layer

Fee checking

Solicitor invoices are checked line by line against the statutory fee schedule. Over-charges, duplicates and out-of-scope items are surfaced before payment. Each flag carries the rule it failed, so review is fast and defensible.

The same layer, reused · 08

More on the same layer.

Each reuses the same fabric, studio and governance — configuration on one foundation, not a new build.

Deadlines

Deadline monitor

Legal deadlines and limitation periods are tracked across the open book. The system alerts before a deadline is at risk — not after it is missed. Every deadline is tied to its source document and the obligation behind it.

Reserving

Reserves

New claims are clustered with historically similar cases to inform the reserve. The recommendation arrives with the comparable set it was drawn from. Reserving becomes consistent and evidenced, not case-by-case judgement alone.

Integrity

Fraud & anomaly

Claims are screened for patterns and inconsistencies against the historical book. Suspicious or contradictory cases are flagged for a person, with the anomaly explained. Nothing is auto-declined — the signal simply routes a case to review, sooner.

Merits

Prospects of success

For contested matters, the merits and likely outcome are assessed to inform coverage and routing. The view draws on comparable resolved cases and the clauses in play. It supports the decision; a person still makes the call.

External

Media mining

Counterparties and third parties are screened across open and licensed sources. Relevant findings — sanctions, litigation, reputational signals — are surfaced with their source. Each result is attributable, so a person can verify it before it informs a decision.

Dialogue

Ask the case

Any case can be questioned in natural language across the connected systems. Answers come grounded in the case record, with the evidence behind them. It turns the fabric into something a handler can simply ask.

How it lands · 09

Four gates. Prove one path deeply — not many thinly.

Quality is proven offline before a line of integration is built; the integration path and its full cost are signed off before the engineering starts; and the one legacy read-only path is proven concretely — with radical cost transparency, so no effort ever surfaces late.

01
gate

Quality, offline

Extraction, coverage and drafting quality scored against a ground-truth set of decided claims — before any integration work begins. If it does not clear the bar here, nothing else starts.

02
gate

Integration path priced

The path and its full cost are signed off before engineering starts. Radical cost transparency, so no effort surfaces late.

03
gate

One read-only path, proven

A single legacy system connected read-only and proven concretely — the integration evidence for everything that follows.

04
gate

Exit criteria met

Agreed up front, measured on your rules and your book, and signed off before scope extends.

Fig. 04  //  four gates — prove one path deeply◆ radical cost transparency
1Offline / sandboxquality vs ground truthbase parity + added value, provenon a controlled setLOW EFFORT2Architecture & costthe integration patheffort estimate, pricing &operating model — fullytransparentLOW-MED3Read-only integrationone legacy path, proventhe core engineering: prove onepath deeply, not many thinlyMED-HIGH4Overall evaluationthe pilot casequality x integration x economics,jointly assessedLOWthe expensive engineering lands only after the quality proof and the economics sign-off

What it needs · 10

To prove it on your book, the pilot needs:

Phase 1 · quality

A ground-truth case set

A representative set of decided written claims and their documents — the baseline the extraction, coverage and drafting quality is scored against.

Phase 1 · rules

Your rule base & policy versions

The coverage rules, policy wordings and statutory fee-schedule versions the utility must reason against — so parity is measured on your rules, not ours.

Phase 3 · integration

One read-only path

Read-only access to one legacy system, proven concretely — the integration evidence for everything that follows.

The north star

Your underwriting numbers

Loss ratio, average payout, coverage-error and routing costs — the data that converts the combined-ratio direction into a measured, provable case.

The ask

A scoped pilot on one line of business — quality proven offline first, integration priced transparently, exit criteria agreed up front. Weeks to the quality proof, not a platform programme.

Bottom line · the proposition

Start where extraction stops. Price on the ratio it moves.

One line of business is the beachhead; the full written-intake stream — every life-area — is the prize. The utility clears the efficiency floor from day one, and the same layer then reaches the number the board actually runs on: the combined ratio, where the loss side is twice the expense side and only a claim-understanding layer can touch it. Win the fallout, prove the loss-side impact on your data, and the vendor relationship becomes an underwriting-margin partnership.

Stage 01

The beachhead

One line, one legacy path, quality proven offline — value in weeks.

Stage 02

The platform

The same layer scales to the full written intake across every life-area.

Stage 03

The partnership

From per-case fee to gain-share on the combined ratio it improves.

Next step

Bring one line of business and a set of decided claims.

Quality is proven offline against your own ground truth before a line of integration is written. Exit criteria are agreed before the pilot starts.