Underwriting Software Problems: The 7 That Cost Real Money, and How to Avoid Them

Why does rating parity testing get cut, and what does it cost?

Parity testing is invisible in a demo, it consumes weeks, and it produces no new features, so it is the first item traded away when a fixed price needs to come down. It is also the only thing that makes underwriters use what you build.

The reason is specific to specialty insurance. A habitational or casualty rater is not a formula, it is thirty or forty tabs of protection class lookups, catastrophe loadings, schedule credits and judgement factors accumulated over years, with the reasoning for several of them living in nobody's memory. A developer reading that spreadsheet produces something that is nearly right, and nearly right is the worst outcome, because the differences show up account by account with no pattern and every one of them is an argument on the underwriting floor.

The fix: put parity in the schedule as a named phase with an exit criterion, not a test task. Take several hundred recently bound accounts across the full range of the book, including the awkward ones, run them through the existing rater and the new engine, and reconcile every difference until you can explain it or eliminate it. Some differences will reveal that the spreadsheet was wrong, which is a finding worth having, and each one should be a documented decision rather than a silent correction. A proposal without parity in the schedule is not a plan, and the schedule attached to it is fiction.

What goes wrong with submission data and clearance?

Brokers do not send data. They send an ACORD 125, a statement of values in whatever column layout their agency prefers, and five years of loss runs as scanned documents. Generic extraction handles the ACORD form adequately and then falls apart on a seven hundred row statement of values with merged cells, subtotal rows that look like locations, and construction classes buried in free text.

Clearance is the second failure and it costs bound business. Two wholesalers submit the same insured through different retailers, both get quoted, sometimes at different prices, and the broker notices before you do. That conversation costs more than the account. Under it sits a data problem: insured names arrive spelled four ways, addresses arrive with and without suite numbers, and a shared spreadsheet is doing the matching.

The fix: tune extraction against the actual paper your top twenty producers send, since they carry most of the volume, and treat everything else as manual entry with a queue rather than pretending coverage you do not have. Score every extracted field and route low confidence values to a human before they touch a rate. Clear every submission against the live book on arrival using fuzzy matching on insured name and address, and store the match decision so the same account does not need clearing twice. Never auto accept an extracted total insured value into rating, because a misread subtotal row prices an account you did not intend to write.

Why do the policy admin and data vendor integrations break after launch?

Because the workbench and the system of record were connected by human hands for years, and the integration inherits every inconsistency that hid inside that arrangement. A bound quote pushed into Duck Creek or Vertafore AIM has to match their expected structures for class codes, states, coverage parts and endorsement forms, and yours have drifted. The push works in testing on clean accounts and fails in production on the multi state schedule with a manuscript endorsement.

External data feeds break differently. A property risk score service returns nothing for an address it cannot geocode, and if the rating path treats a missing factor as a neutral factor rather than as an exception, the account prices as though the exposure were average. Nobody sees it, because the quote looks normal.

The fix: reconcile the reference data before the integration is built, not during. Map class codes, states and coverage structures explicitly and keep the mapping as versioned data with an owner. Make a missing external factor a blocking exception that routes to an underwriter with the reason, never a silent default. Monitor the push as a queue with retries and a visible failure list, because a bound account sitting in a failed integration is an unissued policy and eventually a coverage argument. Ask for the integration to be scoped in the first release rather than deferred, since deferring it means keeping the rekeying you were trying to remove.

What happens when authority and referral evidence is not covered?

An underwriter carries authority by premium, total insured value, class and state. Anything above it refers upward, and today that referral is a forwarded email answered with a line of agreement, with no record of the rate version used, the terms attached or the conditions imposed. Policy admin referral workflows fire at issuance, which is after the pricing decision, so they see none of it.

Then the Lloyd's coverholder audit or the fronting carrier's annual review asks for evidence of referral controls, and a compliance manager spends two weeks reconstructing threads. The finding lands on the binding authority renewal, which is the part that actually costs money, because a restricted or delayed authority is a restriction on the book itself.

The fix: encode the authority matrix in the workbench per underwriter and per program, and make a quote that breaches authority unreleasable rather than flagged. Route it to the correct approver with the full rating detail attached, and write the approval, the conditions and any override back onto the quote record with who, when and why. Every quote should carry an immutable snapshot of the rate version, the inputs, the factors applied and the overrides. The evidence pack for the next audit then generates in an afternoon instead of consuming a fortnight.

Should you build custom or configure what you already own?

If you run one or two programs, your rating stays close to standard structures, and volume sits under a few thousand submissions a year, buy. A vendor workbench such as Federato or hx Renew, or even a disciplined single owner spreadsheet process with real version control, will serve you at a fraction of the cost. Configuration beats construction when your process is close to the market default, and pretending otherwise is how agencies sell projects.

Whatever you do, do not replace policy admin. Duck Creek, Vertafore AIM and OneShield are the system of record for issuance, billing and claims, and replacing one is a multi year programme. A workbench ships in months precisely because it leaves that system alone: it owns intake, clearance, rating and referral, then pushes bound business across.

Build when three or more programs each carry their own rater and every rate change becomes a version control incident, when submission volume has you hiring assistants to rekey, when an audit has already produced a finding on referral documentation, or when per seat vendor pricing across the team, compounded over the years you will hold the book, exceeds the cost of owning the asset outright.

How do hidden costs get into the quote?

Four items push this category to the top of its band and none of them is obvious at proposal stage. Rater count and depth, because a twelve tab general liability model converts in weeks and a forty tab property catastrophe rater with external lookups does not, and a quote priced against the first does not cover the second. Extraction ambition, since parsing statements of values across dozens of broker formats is genuinely hard and the honest answer involves a manual queue for the long tail.

Policy admin API condition is the third: a current Duck Creek instance integrates far faster than an ageing installation where the interface is a scheduled file. And bordereaux template count is the fourth, because each carrier wants its own layout with its own totals, and every additional template is real work that recurs whenever that carrier revises the format.

The fix: hand over your actual raters and three months of real submissions before asking for a number, and treat any firm that quotes without looking at them as quoting a different project. Ask for parity testing, extraction and each bordereau template priced as separate lines, then ask what the number becomes when the fourth program comes on.

What separates a build that works from one that fails here?

Make them draw the data model. Submission, clearance, quote, bind and endorsement across the lifecycle, with effective dated rate tables and versioned rating snapshots. A team that models a quote as one mutable row has built you an audit failure, because the record of how an account was priced will be overwritten by the next revision of that account.

Demand a rating parity plan in the proposal, with a stated number of accounts and a defined exit criterion. Check integration scar tissue by asking specifically about Duck Creek and Vertafore AIM interfaces, ACORD data standards, and feeds such as Verisk or HazardHub, and listen for war stories rather than logo slides. Probe compliance instinct by asking how they would evidence a coverholder audit or a fronting carrier review: authority controls, override logs, retained rating snapshots. A blank look there costs you a binding authority later.

Then settle ownership before work starts. You should hold the source code, the rating tables and the cloud accounts, with work for hire language agreed in advance. Your rating logic is the intellectual property the binding authority depends on, it shows up in a valuation, and it should sit on your balance sheet rather than inside somebody else's subscription.