Why Klarefi exists

A commercial loan costs €11,800 to originate. An adjuster carries 130 open claims. A single KYC review costs €30 to €70 to clear, and 95% of the alerts your team chases are false positives. That is the work intake software is supposed to fix. It does not.

I started Klarefi because the intake stack regulated teams actually use is three vendors stapled together, and none of them know about each other.

What you actually have today

You have a form builder. It collects fields. It does not read documents.

You have an IDP tool. It reads documents. It does not chase the applicant for the page they forgot.

You have a workflow platform. It routes tasks. It does not know whether the packet in front of your reviewer is sufficient to decide.

The applicant submits. The packet is incomplete. Your team opens it anyway. Then they spend three days emailing the applicant for the schedule, the endorsement, the police report, the UBO declaration. The first review cycle is chase, not judgment. That is the work I want to cut.

The three failures we are designing against

Forms that do not know what they collected. A field marked "uploaded" tells you nothing about whether the document contained the fact you needed.
Extraction without evidence. A confidence score is not a citation. A reviewer cannot defend "92%" in an audit. They can defend page 4, line 18.
Silent fallback. When the model is unsure, the system should say so and route the case. It should not guess and hope.

Every product decision we make collapses back to those three.

The product model: Form. Fact. Verify.

One abstraction. Three pieces.

Form is what you ask the applicant. Structured questions, hosted on your domain, conditional on what they have already given you.
Fact is what you need to decide. Revenue, EBITDA, policy number, UBO, cause of loss. Each fact has a required shape and a sufficiency rule.
Verify is the contract between the two. Every fact resolves to one of three states: resolved with cited evidence, needs input from the applicant, or failed and escalated to a human.

That is the whole platform. No black box. No "AI magic". Just a state machine that knows whether you have enough to decide.

What an operator opens on Monday morning

Not 200 PDFs. Not a form export. A case file:

Required facts on the left, each with the quote and page that supports it.
Missing or contradictory evidence separated from accepted facts.
A chat history showing what the applicant was asked and what they sent.
An audit log every regulator can read without translation.

Your underwriter, adjuster, or analyst spends their day on judgment. Not on retyping numbers from a tax return into the origination system.

What we will not build

We will not ship a confidence score with no citation. We will not auto-decide adverse outcomes. We will not pretend extraction works on documents we have never seen. The customer pilots its own loan files, claim files, or onboarding packets in two weeks. If the accuracy is not there, we tell you.

Who this is built for

If you are running KYC at a bank, claims at an insurer, credit at a lender, or onboarding at a payments company, you already know the numbers. 70% of financial institutions lost clients to slow onboarding last year. 83% of dissatisfied claimants switch carriers, and the number one complaint is speed. A single corporate KYC review takes 100 days and 100 documents. Adjusters now carry 130 open claims and spend 30% of their time on admin work instead of adjudication.

The pattern is the same across every regulated vertical. Document chase eats your team. Manual data entry causes 42% of claim rejections. 11.5% of mortgage file data is wrong or missing. Half of digital applicants abandon onboarding that runs longer than three minutes. None of that is fixed by a faster form or a smarter OCR. It is fixed by a system that decides whether the packet on your reviewer's screen is sufficient to act.

What this blog is for

I will use this space for three things:

Numbers. Real cost per file, real cycle times, real error rates from regulated workflows, cited.
Mechanics. How we resolve a fact, how we cite a quote, how we route a case that the model could not answer.
Cuts. What we built, what we ripped out, what we wish we had not shipped.

No vendor blog posts about "transformation". No abstract future-of-work takes. If a piece does not give you a number you can take to your CFO, I will not publish it.

You need a front door that delivers verified cases. Build that, or watch your competition do it.

I started Klarefi because the intake stack regulated teams actually use is three vendors stapled together, and none of them know about each other.

What you actually have today

You have a form builder. It collects fields. It does not read documents.

You have an IDP tool. It reads documents. It does not chase the applicant for the page they forgot.

You have a workflow platform. It routes tasks. It does not know whether the packet in front of your reviewer is sufficient to decide.

The three failures we are designing against

Forms that do not know what they collected. A field marked "uploaded" tells you nothing about whether the document contained the fact you needed.
Extraction without evidence. A confidence score is not a citation. A reviewer cannot defend "92%" in an audit. They can defend page 4, line 18.
Silent fallback. When the model is unsure, the system should say so and route the case. It should not guess and hope.

Every product decision we make collapses back to those three.

The product model: Form. Fact. Verify.

One abstraction. Three pieces.

Form is what you ask the applicant. Structured questions, hosted on your domain, conditional on what they have already given you.
Fact is what you need to decide. Revenue, EBITDA, policy number, UBO, cause of loss. Each fact has a required shape and a sufficiency rule.
Verify is the contract between the two. Every fact resolves to one of three states: resolved with cited evidence, needs input from the applicant, or failed and escalated to a human.

That is the whole platform. No black box. No "AI magic". Just a state machine that knows whether you have enough to decide.

What an operator opens on Monday morning

Not 200 PDFs. Not a form export. A case file:

Required facts on the left, each with the quote and page that supports it.
Missing or contradictory evidence separated from accepted facts.
A chat history showing what the applicant was asked and what they sent.
An audit log every regulator can read without translation.

Your underwriter, adjuster, or analyst spends their day on judgment. Not on retyping numbers from a tax return into the origination system.

What we will not build

Who this is built for

What this blog is for

I will use this space for three things:

Numbers. Real cost per file, real cycle times, real error rates from regulated workflows, cited.
Mechanics. How we resolve a fact, how we cite a quote, how we route a case that the model could not answer.
Cuts. What we built, what we ripped out, what we wish we had not shipped.

No vendor blog posts about "transformation". No abstract future-of-work takes. If a piece does not give you a number you can take to your CFO, I will not publish it.

You need a front door that delivers verified cases. Build that, or watch your competition do it.

I started Klarefi because the intake stack regulated teams actually use is three vendors stapled together, and none of them know about each other.

What you actually have today

You have a form builder. It collects fields. It does not read documents.

You have an IDP tool. It reads documents. It does not chase the applicant for the page they forgot.

You have a workflow platform. It routes tasks. It does not know whether the packet in front of your reviewer is sufficient to decide.

The three failures we are designing against

Forms that do not know what they collected. A field marked "uploaded" tells you nothing about whether the document contained the fact you needed.
Extraction without evidence. A confidence score is not a citation. A reviewer cannot defend "92%" in an audit. They can defend page 4, line 18.
Silent fallback. When the model is unsure, the system should say so and route the case. It should not guess and hope.

Every product decision we make collapses back to those three.

The product model: Form. Fact. Verify.

One abstraction. Three pieces.

Form is what you ask the applicant. Structured questions, hosted on your domain, conditional on what they have already given you.
Fact is what you need to decide. Revenue, EBITDA, policy number, UBO, cause of loss. Each fact has a required shape and a sufficiency rule.
Verify is the contract between the two. Every fact resolves to one of three states: resolved with cited evidence, needs input from the applicant, or failed and escalated to a human.

That is the whole platform. No black box. No "AI magic". Just a state machine that knows whether you have enough to decide.

What an operator opens on Monday morning

Not 200 PDFs. Not a form export. A case file:

Required facts on the left, each with the quote and page that supports it.
Missing or contradictory evidence separated from accepted facts.
A chat history showing what the applicant was asked and what they sent.
An audit log every regulator can read without translation.

Your underwriter, adjuster, or analyst spends their day on judgment. Not on retyping numbers from a tax return into the origination system.

Numbers. Real cost per file, real cycle times, real error rates from regulated workflows, cited.
Mechanics. How we resolve a fact, how we cite a quote, how we route a case that the model could not answer.
Cuts. What we built, what we ripped out, what we wish we had not shipped.

No vendor blog posts about "transformation". No abstract future-of-work takes. If a piece does not give you a number you can take to your CFO, I will not publish it.

You need a front door that delivers verified cases. Build that, or watch your competition do it.