Runink FACE · Insurance, inside the Finance domain
This is a scenario for Runink FACE, the Fulfilment Autonomous Claims Engine. Insurance is not a separate product or a separate module: FACE types claims, reserves, premiums, deductibles, settlements and underwriting records into its Finance domain, on the reasoning that a claim is a reserve held against a policy.
In Short
- The file is assembled. It is not judged. The wording, the loss documents, the reserve movements and the authority limit that applies at that value arrive in one place, joined to the claim they belong to.
- Every reading says where it came from. Text lifted out of a document comes back with the file it was read from and the method used to read it, so a figure in the draft can be traced to a page rather than taken on trust. What it does not come back with is a per-figure confidence score — there is a confidence number on the response and it is a fixed one, which means it tells you nothing, and we would rather say that than let you read it as a quality signal.
- The decision is not the software's to make. It drafts the next step. An underwriter or a claims lead approves it, rewrites it or throws it out, and their name stays on the record beside what they decided.
The Gathering Is The Job. The Judgement Should Be.
An adjuster's morning goes on finding the wording, the loss report and last quarter's reserve movement. The part that actually needed their training takes ten minutes at the end of it.
Where It Goes Wrong
A claim comes in. The policy wording sits in a document system. The loss report is a PDF somebody emailed on Friday. The reserve history is in the policy administration system. The authority limit that applies to a claim of this size is in a procedure note, and the current version of that note is not obvious. Joining the four is most of the work and none of the skill.
So the large files get the full treatment and the rest get handled on a partial reading of the record by somebody experienced enough to usually be right. “Usually” is carrying a great deal of weight in that sentence, and nobody wants to be the one who writes down how much.
The bottleneck is not judgement. It is assembly.
Where underwriting is delegated, the same problem arrives as paper. The schedules come back monthly. Reading them against what the delegation agreement actually permits is a rules-against-records job at a volume no team reads all of, and the accountability does not move just because the decision did.
Who This Is For
Three desks that answer for a file they did not have time to read in full.
- Claims operations — the adjuster and the claims lead. What lands today is a file in pieces: the wording in a document system, the loss report in Friday's mail, the reserve history in the policy administration system, and an authority limit written in a procedure note whose current version is not obvious. What changes is that those four arrive joined to the claim, with one drafted next step and the reading behind it, so the ten minutes that needed their training are not the last ten minutes of the morning.
- The underwriting manager who delegated authority. What lands today is a monthly schedule from the agent, and reading it against what the delegation agreement actually permits is a rules-against-records job at a volume no team gets through. What changes is that the terms you wrote are held against the schedule you were sent, and each place the two part company is named with the clause and the record both cited.
- Compliance and risk, and internal audit. What lands today is a question about a decision taken months ago, answered by reconstructing it across several systems. What changes is that approvals and refusals are both kept as they happen, carrying who decided and what they changed — and that the authority limit as documented can be set beside the limit as configured, which is the comparison that finds the threshold raised during a backlog and never put back.
What Happens Instead
Insurance records are recognised as insurance records. A claims extract out of a policy administration system is typed by its vocabulary — claim, reserve, adjuster, premium, deductible, settlement, payout — rather than being filed wherever its most generic column happened to point. This is a real failure we had to fix: a whole claims dataset once landed in operations because one of its columns was called “status”.
Documents are read where they already are: a PDF on an SFTP drop, a Word file in a shared drive, the spreadsheet whose formulas quietly implement a rate rule nobody has written down — and the formulas are read, cell by cell, not just the values they happen to be showing. What comes back carries the file it was read from and the method that read it, so a figure in a draft leads back to a page.
What it does not carry is a usable confidence on the reading, and that is the kind of thing we would rather you heard here than discovered. The extraction returns one score for the batch and that score is a constant: it is the same whether every page came out cleanly or every page failed. So it is not a quality signal, it should not be shown to a handler as one, and a blurred scan is still a document somebody has to open. Treat the extraction as having found the page for you, not as having verified what is on it.
Then the written rule is held against the record, and every finding takes the same shape: here is what the document says, here is what the file shows, here is where the two part company. Does the wording cover this. Were the documents the procedure requires actually in the file. Was the handler inside the authority that applies at that value. The comparison is only as good as the rule you gave it to compare against — it is reading your clauses, not a library of insurance law — and where it finds a disagreement it cites the clause and the record both, so the first question in review is about the case rather than about where the numbers came from.
What it does not do is underwrite. It does not price a risk. It does not accept or decline. It does not set a reserve, and it does not settle anything. It assembles the file, drafts one proposed next step, and shows what it read to get there. An underwriter or a claims lead makes the call. If you are shopping for software that makes the call instead, this is not it — and you should ask hard questions of anything that says it is.
Approving the drafted step is what sends it. Declining it is recorded too, which is the part most systems lose. And where the step has a leg with nothing implemented behind it — a write into a policy administration system, say — the response names that leg as not executed rather than reporting the action as done, so the file never shows a step as taken that was only approved. Months later, who decided this, and on what basis is answered out of the file rather than out of somebody's memory of a Tuesday.
How You Will Know It Worked
Every figure below is yours, not ours. Write down where you stand today, because the baseline is gone for good the moment things improve.
- How much of a file is gathering. Take a sample of claims closed last quarter. Ask the handler to split their time on each one into finding things and judging things. Most teams have never asked, and the answer tends to surprise the people who manage them.
- How many files were decided on a partial record. In the same sample, count the ones where the wording was not actually re-read, or the authority limit was assumed rather than checked. This is an uncomfortable number. It is also the one that matters.
- The gap between the written rule and the enforced rule. Put your authority limits as documented beside your authority limits as configured. If a threshold was raised during a backlog, find out whether anybody ever lowered it again.
- Days from a delegated schedule arriving to somebody reading it. Where you delegate underwriting. Measure it as a gap, not as a total, and count the schedules nobody opened at all.
Bring one closed claim file and your delegated authority schedule.
Drawn not measured
Runink's published work is in logistics and operations. The claim, the adjuster's morning and the delegated schedule above are drawn to show the shape of the arrangement — they are illustrations of how it works, not accounts of things that happened. There are no recovery rates on this page, no settlement amounts, no cycle times and no customer names, because we have not measured them.
Nothing here is an approval, an authorisation or a certification of anything. The software reads records and drafts; it holds no delegated authority, it is not a regulated actor, and using it does not satisfy an obligation on your behalf. Where a decision must be made by a person with authority to make it, that person makes it and the record says who they were.
Questions An Insurer Asks Before Buying
What it reads, who decides, and what it does not claim to be.