Accuracy on clean masters was excellent and the errors clustered where the money is: amendments, order forms that override the master, and terms defined by reference
A summary tells you what an agreement roughly says. Extraction tells you exactly what it says, in fields you can search, match and monitor.
Prepared by Redress Compliance · August 19, 2026 · Contract extraction projects advised, 2024 and 2025.
Executive summary
Errors clustered in three predictable places: amendments, overriding order forms, and terms defined by reference. Those are exactly where the operative price lives.
Projects that trusted extraction with no confirmation step shipped a wrong renewal date into a live calendar. Capability is not the same as verified.
Page anchors were the difference between a confirmation taking seconds and meaning a reread. Extraction without anchors asks for trust it has not earned.
Teams that measured accuracy on their own messy contracts trusted the output. Teams that measured on the demo did not, and were right not to.
What is extraction, against summarization?
The conversion of an unstructured document into structured records: named fields with values, and clause positions with locations. The output is data rather than prose.
A summary cannot be queried
Summarization paraphrases. Extraction produces a renewal date field, an uplift cap field, a liability cap field, each tied back to the page and clause it came from.
That difference is what makes the downstream work possible
You can run a renewal calendar off the first field and an invoice check off the second. You cannot run either off a paragraph of prose, which is why the distinction matters before you trust any of them.
How does the pipeline actually work?
Four stages, and the third is the one most buyers underestimate because it is the only one that is not automated.
| Stage | What happens | Why it matters | Where it fails |
|---|---|---|---|
| Read | Parsing turns the document into machine text | Scan quality decides everything downstream | Poor scans, unusual layouts |
| Propose | The model proposes fields and positions with confidence | Where the intelligence lives | Amendments and overriding order forms |
| Confirm | A human reviews proposals against anchored source text | Turns plausible output into trustworthy data | Skipped entirely |
| Index | Clause level indexing makes the estate searchable | Enables one question across every contract | Indexing unconfirmed fields |
Page anchors are not negotiable
Every extracted field should link to the page and clause it came from. That is what makes confirmation a seconds long check rather than a reread, and what lets a negotiator or auditor verify the claim later.
Confidence should route the review
Good extraction attaches a confidence signal to each field so review can be weighted: high value contracts and low confidence extractions go to a human first. The model layer documented by makers such as one provider and another is capable, and capable is not the same as verified.
Read one contract before you trust a pipeline
Upload a single agreement and see what anchored, confirmed extraction actually produces.
Open the decoder →What the extraction projects showed
In the extraction projects Fredrik Filipsson advised on in 2024 and 2025, accuracy on clean masters was excellent and the errors clustered where the money is. Three findings recur.
- Page anchors were the difference between a confirmation step that took seconds and one that meant rereading the contract.
- Confidence weighted routing, high value and low confidence first, focused scarce human review where it changed outcomes.
- The teams that measured accuracy on their own messy contracts trusted the output; the teams that measured on the demo did not, and were right not to.
Projects that trusted extraction with no confirmation step shipped a wrong renewal date into a live calendar. The failure was not the model; it was the missing stage.
- Every risky clause flagged with the verbatim quote and page anchor
- Your agreements decoded into plain English before the auditor interprets them for you
- Entitlements, caps and protections verified across your whole contract portfolio
Where does it reliably fail?
In three predictable places, and they are exactly the places the commercial terms actually live.
Amendments that modify an old master
An amendment changing a master signed years earlier is where extraction is weakest and where a repriced term often sits. The model has to reconcile two documents written under different assumptions.
Order forms that override the master
An order form overriding the master for one purchase can carry the operative price, and extraction may attribute it to the wrong document entirely.
The negotiation those documents come out of is worked through in the AI provider negotiation series.
Terms defined by reference
When the real term lives in a linked policy or an appendix, extraction can miss it completely. Nothing in the document being read says the number is somewhere else.
Watch the briefing · 3:50Signing the AgreementWhich terms carry the money at signature, and where the operative language actually sits in the document set.
What does a trustworthy design look like?
Confirmation weighted by risk and value, with page anchors so review is fast. Not more accuracy, but a stage that assumes imperfect accuracy and catches it.
Three design rules
- Anchor every field, so confirmation is a glance rather than a reread.
- Route by confidence and value, so scarce review lands where it changes outcomes.
- Never index unconfirmed fields, because a searchable wrong answer is worse than no answer.
Measure on your own contracts, not the demo
The teams that measured accuracy on their own messy estate trusted the output, and were right to. The teams that measured on a clean demonstration set did not, and were also right. The governance expectations are set out in the risk management framework and the European regulatory framework.
Where the common advice on extraction is wrong
The common advice is to benchmark extraction vendors on headline accuracy and pick the highest number. We disagree.
Accuracy on clean masters is the wrong measurement
It was excellent everywhere. The errors clustered on amendments, overriding order forms and terms defined by reference, and a headline accuracy figure measured on clean documents tells you nothing about any of those.
The buyer side move is to test on your own messiest contracts, require page anchors on every field, insist on a confirmation stage weighted by value, and never index anything unconfirmed. The downstream uses sit in contract management, invoice reconciliation and deal diligence.
What the projects established, 2024 and 2025
Two design findings rather than measured ones, and both changed what the output was worth.
Amendments modifying an older master, order forms overriding it for one purchase, and terms defined by reference to a linked document.
Read, propose, confirm and index, where confirmation is the stage most often skipped and the one that makes the rest trustworthy.
Neither is a vendor comparison. Both are the test a buyer applies before trusting any calendar built on the output.
Your first five moves
- Test on your own messiest contracts rather than a demonstration set, because accuracy on clean masters was excellent everywhere and told nobody anything.
- Require a page anchor on every extracted field, since anchors are what make confirmation seconds long instead of a reread.
- Insist on a confirmation stage weighted by value and confidence, because projects that skipped it shipped a wrong renewal date into a live calendar.
- Watch the three failure clusters specifically, which are amendments, overriding order forms, and terms defined by reference to another document.
- Never index an unconfirmed field. The advisory practice runs the confirmation stage as part of the pipeline rather than as an optional review.
Frequently asked questions
What is contract data extraction?
The conversion of an unstructured document into structured records: named fields with values and clause positions with locations, tied back to their source.
How is it different from a summary?
A summary paraphrases and cannot be queried. Extraction produces a renewal date field and an uplift cap field you can run a calendar and an invoice check off.
Where does extraction fail?
In three predictable places: amendments modifying an older master, order forms overriding it for one purchase, and terms defined by reference to a linked document.
Why are those the dangerous ones?
Because they are exactly where the operative commercial terms live. The repriced term usually sits in the amendment rather than in the master.
What are page anchors for?
They link every field to the page and clause it came from, which makes confirmation a glance rather than a reread and lets anyone verify the claim later.
Can the confirmation stage be skipped?
Not safely. Projects that trusted extraction with no confirmation shipped a wrong renewal date into a live calendar. The failure was the missing stage, not the model.
How should review be prioritized?
By confidence and value. High value contracts and low confidence extractions go to a human first, which focuses scarce review where it changes outcomes.
How should accuracy be measured?
On your own messiest contracts. Accuracy on clean masters was excellent everywhere, so a headline figure from a demonstration set tells you nothing useful.
Should unconfirmed fields be indexed?
No. A searchable wrong answer is worse than no answer, because it will be trusted by somebody downstream who never saw the confidence signal.
Is the model the limiting factor?
No. The model layer is capable, and capable is not the same as verified. The design around it is what makes the output trustworthy.