The process
Extraction is the easy part.
Document AI sells you an API and an LLM that reads a page. It works. The return only arrives once you’ve built the whole process around it.
The problem
You bought an API. You still don’t have a process.
Every vendor now ships document-in, data-out. The model reads the page, returns JSON, and that part genuinely works.
But JSON isn’t an outcome. Somebody still has to decide what the document is, whether the values are right, what they mean against your business, and what happens when they’re wrong.
That gap is where implementations stall, and where the promised ROI goes missing.
The claim
Capture is one step of seven.
The step everyone sells is the one that has commoditized. The other six are where the work lives, and the value with it.
What you’re sold
An endpoint. Send a document, receive fields. Priced per page, and increasingly hard to tell apart from the next vendor’s.
What you need
A process that classifies, extracts, links to your knowledge, reasons, validates, escalates to a person when it should, and learns from what they change.
The whole process
Where document AI stops, and where ROI starts.
Steps one to five are capability, and the frontier keeps improving them. Six and seven are where the asset accumulates, and it accumulates to you.
Stage 1 of 7 · Sorting the document
Know what you’re holding before you read it.
Routing is a real decision. The same page means different things depending on which document it belongs to, and everything downstream inherits this call.
One intake batch → three routes
Low confidence routes to a person before anything else runs.
Stage 2 of 7 · Reading the page
The part everyone sells.
It’s good, and it gets cheaper every year. Worth showing we do it well. Worth being honest that on its own it solves nothing.
Every value carries its source
The value, and where it came from. Provenance is what makes the rest auditable.
Stage 3 of 7 · Matching your records
Raw values mean nothing until they meet your business.
A vendor name is a string until it resolves to your vendor, with your terms, your history and your rules attached. This is where your knowledge enters the run.
resolved to your vendor record
Structured, versioned, and portable across any model.
Stage 4 of 7 · Automatic follow-up
Some documents need work before they’re usable.
Look something up, reconcile against another system, chase a missing reference, decide what else this triggers. An agent picks its own tools within the plan’s bounds.
Agent step · tool trace
Agent tooling improves every year. The plan it runs inside, and the record it leaves, stay with you.
Stage 5 of 7 · Running the checks
Three places to be wrong.
On the document (does it add up?), against your systems (does it match the record?), and through agent steps (is it reasonable?).
Line items sum to 411,200. The stated total is 412,000.
Vendor bank account does not match the record in your system.
Amount is 4x this vendor's twelve-month average.
Each catches a failure the other two cannot see.
Stage 6 of 7 · Human review
Judgment, where it’s worth paying for.
Not every document, and not every field. Only the ones where the system isn’t confident, or where the stakes justify it. The reviewer opens a case that’s already been assembled.
Review task · assembled for you
Why it reached you
Stage 7 of 7 · Remembering the fix
The correction is the asset.
When a reviewer changes something, they’re telling you what ‘correct’ means here. Captured, that becomes knowledge the next run starts with. Discarded, you pay for the same correction forever.
One correction, captured
Into knowledge that outlives the model that produced it.
Why it compounds
The seventh step is what makes the other six worth building.
A pipeline that runs the same way forever is a cost. One that gets a little more right every time someone corrects it is an asset. The question is what form that asset takes: weights you cannot inspect and cannot move, or structured knowledge you can read, version, and carry to whichever model you use next.
1.5 million activities a month in production, SOC 2 Type II. During a Test Drive we measure yours, on your documents, and you keep the measurement either way.
Share of documents needing a human
Same work, fewer exceptions each cycle. It improves as it runs, and the improvement is yours.
Implement the process, not the endpoint.
Kodexa runs all seven steps as one auditable workflow. Every correction your team makes becomes knowledge you can read, version, and take to any model.