Your tools catch contradictions inside a package. Nobody counts the ones that cross packages

Two buyer-side corporate funds just financed automated document review. At package scale. The cross-functional document falls outside every perimeter.

On 10 September 2026, Wyre AI announced a $5M round led by Ironspring Ventures with WND Ventures, the corporate arm of DPR Construction. Five days later, Fortune broke CADDi’s Series D: $114M at a $1.2B valuation, with Woven Capital, Toyota’s growth fund, on the cap table. In ten days, two buyer-side corporate funds financed the automated review of documents in their own industry.

If you are a CDO in a non-tech group of 5,000 people or more, the reflex is that this is a manufacturing and construction story. What transfers to your organisation is not the tool. It is two more uncomfortable facts: buyers now pay to have their own documents re-read, and they do it at the scale of a package. The document that arbitrates between packages — the internal standard, the cross-functional procedure, the reference technical note — falls inside none of those perimeters. It is precisely where your AI assistant draws its answers from.

On 21 September we wrote about arbitration itself: which of two documents prevails when they disagree. The question here comes earlier. Across which set of documents is a contradiction even looked for? Two answers come to mind immediately. The first is that your internal document reviews already cover it. The second is that the tool you already bought, a data catalogue or an AI governance platform, has the right scope. Neither tells you which set is actually being counted today.

The exercise, before you read on. List the mechanisms that already re-read documents automatically in your organisation looking for inconsistencies. Depending on your business lines, that may be contract review, client-file completeness checks, model-based clash detection, or nothing at all. Write down, for each, the exact scope it sweeps. Half a day is enough. If those scopes cover your reference documents, you will have spent half a day and gained an argument you can defend in committee. If they leave a gap, that gap belongs to no one yet. Those three or four lines are the first lines of the artefact described below.

What capital just certified, and where it stops

These two products have to be described at their documented best, or the argument does not hold. Wyre Check cross-checks drawings against specifications and surfaces gaps, contradictions and compliance risks before the bid, that is, before they resurface on site as requests for information, change orders or delays. More than 250 projects analysed, more than 250,000 scopes and issues identified, more than $3B of project value covered as of the announcement; DPR Construction reports a reduction in scope development effort of roughly 100 to 350 hours per project depending on complexity. CADDi gives semantic structure to fragmented data — 2D drawings, 3D CAD files, ERP records — so that human and AI workflows run on the same base without manual handoffs, and is using its Series D to build models that read those specific formats.

Both are serious products backed by published measurements. Their boundary is clear and deliberate: the unit of work is the package. A bid scope for one, a part family for the other. Inside that unit, a contradiction is detectable because the set is finite, dated and attached to a project. Outside it, there is no package. There is an estate: living documents, approved through different circuits, with no shared closing date, none of which knows that it contradicts another.

A group can therefore buy contradiction detection twice, in two verticals, and have instrumented nothing on the documents that cross them. That is not a flaw in these tools. It is a scope — and nobody in your organisation ever decided it. It is the residue of successive purchasing decisions, made by different departments, at different times, each for its own package. Saying so changes the conversation: this is not catching up on negligence, it is deciding for the first time something that had been deciding itself.

Detection scope is a governance decision, not a tooling detail

A Document Knowledge Platform (DKP) addresses exactly that scope: Govern (ownership, authority and lifecycle of the document estate), Clean (detecting and resolving anomalies, duplicates, obsolescence and contradictions), Activate (opening the corpus to AI systems only once the first two hold). It is not sold in the category of vertical package-review tools and has no business appearing in that RFP: it runs upstream, on the estate those tools consume — as do internal assistants, enterprise search engines and agentic knowledge layers. That position does not move with the funding calendar or the analyst calendar: the document layer precedes the engine, whatever artefact ships this quarter.

Counting across an enterprise-wide scope rather than package by package is catching up with what other functions settled long ago. A finance function does not reconcile invoice by invoice; it reconciles across entities, because the interesting error appears at the junction. An industrial risk analysis does not stop at the component; it covers the system, for the same reason. The document estate is the last large class of corporate assets still controlled one folder at a time.

What the field measurement establishes, and only that. On the technical corpus of a European energy and industrial group, a K-AI diagnostic detected 398 document conflicts, with the reliability of AI answers improved by more than 90% on that scope after remediation. That number states a prevalence on a given corpus at a given moment. It does not say how many you have, and it does not extrapolate: the only way to settle the question in your organisation is to count on your own scope.

This is where ingestion has to be spelled out, at the point where the method is described. A diagnostic of this kind covers document content, and nothing else: no conversation transcripts, no usage logs, no telemetry. The ingestion perimeter is contractual, and content is never reused to train models. Validation is joint between the business Document Owner and the CISO or DPO, never IT alone.

The gap between confidence and maturity has a documentary cause

The Cloud Security Alliance’s 2026 State of AI Governance report, covered on 16 September 2026 and based on a survey of more than 500 US enterprise leaders, measures a gap worth reading twice: 74% believe their organisation would pass an AI compliance audit today, while only 27% consider their AI governance programme fully mature. In the same sample, 90% have funded AI governance, 57% have a formal policy, 44% have documented incident procedures. Informatica’s CDO Insights 2026, run with 600 data leaders, points the same way: data and AI governance sits among the leading stated investment drivers for 2026.

What follows is a reasoned reading of that gap, not a finding of the studies cited. A leader who answers “we would pass the audit” is thinking of what has been instrumented. What is instrumented in most large groups are packages: contracts, models, bills of material, certification files. The cross-functional estate is governed by roles and committees, not by a mechanism that counts. The confidence is sincere, and it covers the visible scope.

Doing nothing remains an option, and its outcome is predictable rather than abstract. The contradictions in the cross-functional estate keep existing, assistants keep drawing on them, and they surface at the moment someone relies on the answer: a customer, an internal auditor, a committee. That is not a disaster scenario. It is the default discovery mode when nobody counts.

A word on where responsibility sits. A vendor such as K-AI answers for counting contradictions, for the reproducibility of that count, and for routing each case to the right expert. It does not answer for the authority decision: saying which of two documents prevails is a business call, and it must stay one.

The detection-scope register

The missing artefact fits on one page, and it adds itself to no existing repository. A detection-scope register lists, for each document source, what re-reads it and what that review does not look at. Five columns: document source, mechanism that reviews it, type of contradiction detected, what is not covered, named owner.

One line, spelled out, so the first one can be written on Monday morning:

Product technical notes (Quality document space) — reviewed by: no automated mechanism — contradictions detected: none — not covered: divergences between successive notes, and against the parent procedure — owner: Product Quality domain lead.

The register has no life of its own: it lives as an annex to the AI use-case register or the application map, whichever of the two is actually maintained in your organisation, and its owner is whoever already maintains that document. Its lines are written during the periodic review of that register, never during the incident that reveals the contradiction: a register filled under deadline pressure reproduces the exact problem it exists to prevent.

That leaves the serious objection, and it usually comes from legal or the CISO: this written, dated document records in black and white what the company does not control. To our knowledge no published decision indicates how such a record would be read in litigation or an inspection, and what follows is a reasoned reading to be validated by your own legal function. A named blind spot, with an owner and a due date, is a managed risk; the same blind spot left unwritten stays a risk discovered by a third party, at a moment of their choosing. The record only works against you if it is written and then abandoned — which is why the owner column is not decorative, and why the register belongs inside a document that is already reviewed periodically rather than in a shared drive.

That column creates no new position, incidentally. It is a written attribution added to an existing role: process owner, domain lead, quality manager. An article asking you to create a function would be closed before the end, and rightly so.

Conclusion

In ten days, two buyer-side corporate funds confirmed that re-reading your own documents is worth an investment. They confirmed it at package scale, where the set is finite and dated. At estate scale the sequence is unchanged: audit what you actually hold, clean what contradicts itself while leaving arbitration to the business, monitor continuously, because a living estate starts diverging again the day after the clean-up.

A one-hour conversation about your document estate is enough to fill the first lines of that register: we look together at two or three of your real document sources, what re-reads them today, and what that review leaves out.

Frequently Asked Questions

Our business tools already catch inconsistencies inside our files. What is left?

Everything that belongs to no file: cross-functional procedures, internal standards, reference technical notes, sales support material. These are the documents AI assistants draw on most and the ones least covered by any detection mechanism.

Does a DKP replace our data catalogue or our AI governance platform?

No. A catalogue describes and locates assets; an AI governance platform frames models and use cases. A DKP works on document content itself, upstream, and hands those layers a corpus whose contradictions have been counted and resolved.

Who should validate the scope of a document diagnostic?

The business Document Owner, who carries the pain, jointly with the CISO or DPO, who carries the framework. IT integrates; it does not validate alone. The ingestion perimeter is contractual, limited to document content, with no reuse for model training.

Is the detection-scope register one more deliverable to maintain?

It is only useful as an annex to a document that is already maintained — the AI use-case register or the application map. As a standalone file it will die the way standalone files die. On the question of the written record it creates, see the relevant section and have your legal function rule on it.

How many contradictions should we expect to find?

No external number answers that honestly. Published measurements cover specific corpora and do not extrapolate. The only answer defensible in committee is a count on your own scope.

Sources

Where to Go From Here

K-AI Corpus Diagnostic — 10 business days on your document estate, full report of the 20 most critical anomalies, money-back guarantee if no meaningful anomaly is found. To write the first lines of your detection-scope register, reach the K-AI team: contact@k-ai.ai. The scope of every diagnostic is validated jointly by the business Document Owner and the CISO/DPO, never by IT alone.

K-AI already works with CMA CGM, Veolia, PwC, BNP Paribas, TotalEnergies and CEVA Logistics. Partners: AWS, Snowflake, Microsoft, Wavestone, Devoteam.