Insurance
The rule is written down. The question is whether it is the one being applied.
Policy wording is a rule. So is a regulatory obligation, a delegated authority agreement, an internal limit. Each one is enforced in a system, performed by a person following a procedure, or neither — and which of the three tends to surface only when somebody goes looking. This page is Runink FACE, whose Finance domain models a claim as a reserve against a policy.
Where the decision happens
- Records you already keepSpread across the systems you already run.
- One finding, with its evidenceDrawn together into a single proposed action.
- A person approves itThe drafted action waits; approving is what sends it, under their name.
- The work goes outAnd the reason stays on the record.
Where it goes wrong
-
The threshold that drifted
The value at which a claim needs a second review is raised in the workflow to clear a backlog. It is meant to be temporary. The policy document is never updated, and months later nobody has decided either way.
-
Decisions explained to someone who was not in the room
Every step applied to claims, policy and payment data has to be reconstructable afterwards. When the answer is assembled by hand across several systems, you pay the assembly cost every time the question is asked.
-
Reconciliations done periodically because of volume
Reserve movements follow rules. Reinsurance recoveries follow treaty terms. Both compare a written instrument against a stream of transactions, and both get sampled because the stream is too long to read.
-
Authority you delegated and still answer for
Where underwriting is delegated, you remain accountable for what is decided under it. Reading the delegation agreement against the bordereaux the agent reports back is a rules-against-records comparison at volume.
Who owns this
The audit question is the same one in every function: not whether a control exists, but what it did.
-
Claims operations
A difference surfaces at the point of decision, not at period end — while the file is open and the payment has not gone out.
-
Compliance and risk
"Controls were operating effectively during the period" becomes a list: the exceptions found, when each was found, what was decided and by whom — and, named beside them, the controls that could not be checked at all.
-
Internal audit
Evidence is a by-product of the work, so demonstrating that a control operated stops being an assembly project.
-
Finance
Where the written terms and what was charged or accrued diverge is stated plainly, with the clause and the records behind it.
-
IT and security
The reasoning runs where FACE runs, never at an outside AI service, so the security review is a description, not a negotiation.
What Runink FACE changes
- The rules as written — policy wording, regulatory text, a delegation agreement — are set beside the logic your systems actually run, and every rule lands in one of four states: aligned, the policy says it and the systems do it; drift, the systems do something adjacent; shadow, logic is running that no policy describes; missing, the policy describes a control nothing performs. Shadow is the one that surprises people.
- Read that as a reconnaissance, not a ruling. Each rule comes back stated in plain English, naming the policy document and the implementation it was read out of, with a confidence attached — so a compliance officer is checking a specific claim against two named sources rather than accepting a score.
- Each finding carries the observation, the rule it invoked, the records it cited, a severity and one specific proposed action, in one queue.
- A finding that turns on risk appetite is not decided for you. It waits, with the ambiguity named and the write-up already done, which is the part that costs the morning.
- A drafted action waits in the queue until a named person approves, edits or rejects it, and that decision is recorded as an event carrying the actor. A connection to your claims data cannot be created without its credentials — they are written first, so a half-made connection does not exist to be used later.
- Where a check could not run, the result says so in those words: not a finding that the thing is compliant. A quantity nobody measured is held as unmeasured with a reason rather than rounded to zero, and a connection nobody has contacted is never reported as verified. For a function whose product is evidence, being told what was not checked is worth more than a clean-looking pass.
The jobs this covers
Each one has a page of its own: what it reads, what it hands to a person, and how you would know it worked.
-
Underwriting and claim files
The policy wording, the loss report, the reserve history and the authority limit for a claim that size arrive joined, with a drafted next step and the reading it rests on. The underwriter still decides.
-
Contract and obligation review
Which clause governs this, and who checked. The clause and the record arrive together, cited — and could not check is kept as an entry of its own rather than folded into a green tick.
-
Customer data privacy
Email addresses, telephone and card numbers, national identification numbers and network addresses stripped out of log and diagnostic output before it is written, so the trail a system leaves does not become a second copy of a claimant's file.
How you will know it worked
These figures are yours, not ours. Write down where each stands today, and over what period, before anything changes — once it moves, the baseline is gone.
| The measure | Your figure today | What moves it |
|---|---|---|
| Claims cycle time | Claims system, last four completed quarters: median days from first notification of loss to settlement, by product. | Down. The wait is mostly evidence being gathered; here the supporting documents are read against each other as the file is built. |
| Leakage | Your last file-review cycle: amounts paid beyond what the wording, the authority limit or the procedure required. Note the sample size beside it. | Down, and the sample stops being the limit — every file FACE can reach is read against the wording and the authority for its value, rather than the hundred somebody had time for. Reaching them is the part to check first: the claims and policy systems FACE connects to are named ones — Guidewire, SAP, Dynamics 365, Salesforce, ServiceNow, SharePoint, your databases and warehouses, your spreadsheets and object storage — and every one of those reads is read-only by construction. |
| Loss adjustment expense | Ledger, last twelve months: allocated and unallocated loss adjustment expense per closed claim, by class. | Down, because the assembly is the cost — finding the report, matching it to the schedule, checking the invoice against both. |
| Straight-through handling, read with reopen rate | Claims workflow, same period for both: the share of closed claims that needed no manual queue step, and the share reopened inside your standard window. | The first up, the second flat — but be clear where the mechanism sits. FACE does not settle claims and takes no settlement decision; it reads the file against the wording and the authority and drafts what it thinks should happen. What moves this pair is that the files whose documents disagree are the ones surfaced, with the disagreement named, so a handler's time goes to those instead of to the ones that were always going to be fine. Read alone the first figure is easy to game, which is why it is written here beside the second. |
| Subrogation recovery rate, and recoveries lost to time bar | Recovery ledger: amounts recovered against amounts identified as recoverable, plus the files where the right to recover lapsed before anything was filed. | Both by one mechanism. Small cases expire because assembling one costs more than it is worth, and that threshold follows the assembly cost. |
| Evidence-assembly time for an audit or a supervisory request | Your last internal audit and your last supervisory request: elapsed days from question to answer, and the person-days inside that. | Down. The observation, the rule it invoked, the records it cited and who approved it are kept as the work is done, so answering is retrieval. |
Two things that make the above possible
-
Claim files go to no outside AI service
The reasoning runs on the machines your plan names: your own servers or cloud account, or Runink's shared machines. Claimant details, medical evidence and adjuster notes are read there, and the model FACE reasons with runs there too: no third-party model dependency, one inference endpoint, the one set up for your plan.
- What you would notice
- The privacy assessment before any claims tooling goes live: no transfer out to a model provider to argue about. And when a supervisor asks who read a claimant's file, and under what authority, the answer is a record. Personal detail is masked in transit by a redaction pass covering email addresses, card numbers, telephone numbers, national identification numbers and network addresses, plus named credential fields, called at more than thirty places.
-
Outside checks with no account attached to them
When a file needs public sources — adverse media on a claimant, a supplier's history, a court listing — the search runs from the machines your plan names through a public search endpoint, and a browser on those same machines fetches and reads the pages behind the results.
- What you would notice
- The search engine sees the query, as it would from any browser. What is absent is the account. No API key, no vendor contract, no per-question bill — so no supplier is building a record of which claimants your firm has been asking about, filed under your firm's name and retained on their terms. That is the difference between a check you can run on a sensitive file and one your privacy officer stops.
Who you would be working with
Runink is founder-led. The person in the first meeting is the person who designed the thing being discussed — his track record, what he does here, and the means to check any of it are on the company page.
See whether it fits
Bring one control and the systems it is meant to live in — a second-review threshold, a delegation agreement, a treaty term. Half an hour is usually enough to see whether what is written and what is applied still agree.
Runink FACE is the product behind this page. The FACE paper is the long version of what is here: what gets read, what a finding contains, who approves it and where it runs. It carries no case studies, no customer names and no return-on-investment figures.