AI Data Ingestion for Underwriting: Lessons for FinTech

Property underwriting has had a quietly brutal workflow problem for years: too much of the “risk work” is actually data typing. In IAG’s case, underwriters were touching seven different systems and could spend up to half a day just ingesting partner-provided documents before they even started shaping an offer.

That’s not a niche insurance annoyance. It’s the same bottleneck banks hit in credit assessment, fintechs hit in onboarding, and trading teams hit in pre-trade checks: good decisions don’t happen until the data is clean, structured, and in the right place. If your risk engine is smart but your data ingestion is slow, your whole business is slow.

IAG’s recent rework of high-volume ingestion for its intermediated brands (CGU and WFI) is a useful case study for anyone building AI in finance and fintech. The headline isn’t “LLMs in underwriting.” The headline is: modern risk operations start with ingestion, not dashboards.

Why “data ingestion” is the real underwriting platform

Data ingestion is the point where operational reality meets your risk models. Get it wrong and your fancy analytics don’t matter—because the inputs arrive late, incomplete, or inconsistent.

IAG’s underwriters were receiving three key document types from partners (think brokers): schedules that ranged from a couple of assets to thousands of locations and asset types, plus risk transfer information and previous history. This variety is exactly why ingestion is so painful: templates differ, fields shift, and the “truth” is scattered across PDFs, spreadsheets, and email attachments.

Here’s the stance I take: if your underwriters (or credit analysts) spend hours re-keying data, you don’t have a risk function—you have a data-entry function wearing a risk badge.

For financial services leaders, the parallel is immediate:

• Credit scoring systems can’t move faster than income/expense verification ingestion.
• Fraud detection performs best when transaction, device, and identity signals are normalized quickly.
• Algorithmic trading relies on low-latency, high-quality market and reference data feeds (ingestion at speed).

The result is the same everywhere: improve ingestion and you improve throughput, decision quality, and customer experience.

The practical target: accuracy that’s good enough to automate

IAG set a clear operational goal: automatically extract and ingest partner documents into underwriting tools at around 98% accuracy.

That number is doing a lot of work.

In regulated, high-stakes finance workflows, “pretty good” isn’t good enough. A model that extracts fields at 85–90% accuracy might demo well, but it creates a bigger downstream problem:

• Staff waste time hunting for errors.
• Exceptions spike.
• Trust collapses.
• Teams revert to manual processes “just to be safe.”

IAG’s experience is a familiar pattern: early approaches were OCR-driven, then they tested AI and large language models to improve extraction. Their proof-of-concept reportedly started around 68% accuracy—useful as a learning moment, not as automation. Within a couple of months of close work with their platform partner, they pushed toward 96–98%.

What actually changes when you hit ~98%?

At that level, you can redesign the workflow around automation rather than “assistive” tools.

A practical way to think about it:

1. Below ~80%: AI is a suggestion engine; humans do most work.
2. 80–95%: AI is a copilot; humans verify a lot.
3. 95–99%: AI becomes the default path; humans handle exceptions.
4. 99%+: you can start optimizing for speed and straight-through processing at scale.

Most finance teams underestimate how dramatic the shift is between stage 2 and stage 3. The “last mile” is where you earn automation.

Underwriting and credit scoring are closer than they look

Insurance underwriting and bank credit scoring are cousins. Both are about pricing risk, controlling exposure, and making fast decisions with imperfect information.

The hidden connection is that both domains depend on document-heavy ingestion:

• Underwriting: asset schedules, claims history, risk transfer schedules.
• Credit: payslips, bank statements, invoices, tax documents, asset proofs.

If you want better risk modeling, you start by feeding your models better inputs. That’s why IAG’s program matters beyond insurance: it’s a reminder that AI in finance isn’t only models—it’s pipelines.

The fintech lesson: don’t build risk on a messy intake layer

I’ve found that many fintech product roadmaps obsess over model choice (“Should we use gradient boosting or an LLM?”) while ignoring the reality that:

• data arrives late,
• data arrives in weird formats,
• and data arrives with missing context.

A smarter sequence is:

1. Standardize ingestion and schema first.
2. Add extraction + validation automation next.
3. Then improve modeling.

This mirrors how modern credit platforms mature: first you unify customer data, then you automate verification, then you introduce more advanced decisioning.

What “commercial enablement” really means in finance ops

IAG described its broader effort as a “commercial enablement” program—reducing admin tasks and manual controls. That phrasing is worth stealing, because it’s the right framing for AI operations work.

Too often, AI projects are sold as “innovation.” Underwriting teams don’t need innovation. They need capacity.

When underwriters spend half a day on ingestion, you’re paying expensive risk talent to do the work of a brittle integration layer. Fixing ingestion creates capacity in three places:

• Speed to quote / decision: fewer delays between submission and offer.
• Growth throughput: more submissions handled per underwriter.
• Colleague experience: less grind means better retention and better judgment.

Banks can translate this directly:

• Less manual data entry in business lending means faster approvals.
• Better ingestion in AML investigations means faster case resolution.
• Cleaner onboarding ingestion means fewer KYC drop-offs.

This is the operational backbone of AI in finance.

How to implement AI ingestion without creating a compliance headache

If you’re aiming to use AI for data ingestion in underwriting, lending, or claims, treat it like a production risk system—because it is.

1) Design for exceptions, not the happy path

The reality: your extraction won’t be perfect, and documents will always be messy. Build a workflow where:

• the AI extracts fields,
• confidence scores drive routing,
• low-confidence items go to review queues,
• and humans can correct values quickly.

Your KPI shouldn’t only be “accuracy.” It should include:

• straight-through processing rate (what % needed no human touch),
• average handling time for exceptions,
• rework rate (how often a correction is later corrected again).

2) Validate against business rules, not just model confidence

A model can be confident and still wrong.

Finance teams should layer deterministic validations on top of AI extraction:

• asset values must be within sensible ranges,
• addresses must match known geographies,
• policy start dates can’t precede submission dates,
• entity ABNs/ACNs must pass checksum formats.

This hybrid approach (AI + rules) is how you keep automation safe.

3) Keep a full audit trail for every extracted field

If you can’t answer “Where did this number come from?” you will struggle with governance.

Best practice audit trail elements:

• document version hash,
• extraction timestamp,
• extracted value + confidence,
• source snippet / bounding box reference,
• user override history.

This is equally relevant in lending and insurance: auditors and risk committees care about traceability, not novelty.

4) Operationalize model monitoring like you would fraud models

Document formats drift. Partner templates change. A new broker uploads scanned images instead of digital PDFs.

You need ongoing monitoring:

• accuracy sampling by partner and document type,
• weekly exception trend reviews,
• automated alerts when confidence drops,
• retraining cadence tied to real-world drift.

The practical payoff is stability: automation stays trusted.

Why tackling the hardest ingestion first is often the right move

IAG noted they “unwittingly” trialled AI on one of their most complex ingestion tasks—and that success made the solution reusable across other areas like acquisition and claims.

That’s a good strategy more often than teams admit.

When you solve the hardest ingestion problem:

• you force your data model to handle real complexity,
• you expose edge cases early,
• and you avoid building a solution that only works in clean, toy workflows.

For fintechs, the equivalent is starting with:

• complex SME bank statements and accounting exports,
• multi-entity ownership structures,
• or cross-border identity documents.

If you can automate that, everything else gets easier.

What to do next if you’re modernising underwriting or credit ops

If your underwriting or credit team is still copying fields between systems, you’re sitting on a compounding cost: slow decisions, inconsistent risk inputs, and burned-out specialists.

A tight action plan for the next 60–90 days:

1. Map the ingestion journey end-to-end (documents → extracted fields → systems touched → decision made).
2. Pick one high-volume workflow where humans re-key the same fields repeatedly.
3. Define “automation-ready accuracy” (often 95–98%) and what happens below it.
4. Implement field-level auditability from day one.
5. Measure outcomes that matter: cycle time, straight-through rate, and exception handling time.

This is how you turn AI in finance from a demo into a measurable operating advantage.

The bigger question for the AI in Finance and FinTech series is simple: if insurers can modernise ingestion to free underwriters to focus on risk, what’s stopping banks from doing the same for credit teams and AML investigators?