A delivery lead we work with, at a large engineering firm, reviews five project management plans a week. Each is around twenty sections. Each has to be defensible against the bid, the contract, the client's original conditions, and the firm's method statement library. Historically each plan took a senior project manager two weeks to assemble by hand and then another week for the delivery lead to check clause by clause.
The firm ran an experiment with a commercial document-drafting AI. The AI wrote the first draft in an hour. The delivery lead's review time went up. Not down.
The reason came out on the second review. About a third of the paragraphs the AI wrote were correct. Another third were plausible but not actually in the source documents. The final third were correct but sourced from documents the AI should not have used, and would not have used if the delivery lead had known to constrain it.
The delivery lead's job had moved from checking whether the plan was complete to checking whether every sentence was traceable. She was now doing forensic work on her own document. That is the opposite of the productivity story the AI was supposed to tell.
Generation is commoditized. Provenance is not.
Every large language model can produce a fluent 20-section project management plan. The base rate for surface-level competence is now high enough that no product wins on the drafting alone. The differentiator is not whether the sentences read well. It is whether the sentences are traceable to a source clause the delivery team can cite in a room full of reviewers.
The industry has a name for the failure mode when they are not: hallucination. It is a bad name, because it makes the failure sound like an occasional rounding error. In practice on regulated deliverables, the failure rate is closer to a third of the paragraphs. Any of those paragraphs that a downstream reviewer stamps without catching becomes the firm's professional-indemnity exposure.
Provenance is not a feature of good writing. It is the requirement. If a sentence in a deliverable is not traceable to a source clause, the sentence should not be in the deliverable.
What "cited" has to mean
There is a version of "cited" that many tools ship. The tool writes a paragraph. Alongside the paragraph, it puts a footnote number. The footnote number links to a source document. The reader who bothers to click the footnote finds a source document that broadly relates to the topic, but rarely says the specific thing the paragraph asserts.
That is not citation. That is decoration.
The version of "cited" that survives a governance review has four properties.
Verbatim. The citation resolves to the exact clause, sentence, or table cell the claim rests on. Not "the contract". Not "clause 12". "Master Services Agreement §12.3, second sentence: 'The Contractor's liability shall be unlimited and uncapped, surviving termination.'"
Stable. The citation carries an identifier that survives a re-run. If Provenant reads the contract twice, the clause is the same clause. If the contract is versioned, the citation carries the version. Reviewers can compare the diff between Rev 03 and Rev 04 and see which cited clauses moved.
Page-anchored. The citation resolves to the page and the paragraph in the source document, not just the document. When the reviewer wants to open the source, one click gets them to the exact location.
Behind the sentence, not next to it. Every sentence in the draft is a link. Click any word and the source clause opens next to it. There is no "reference section". There is no "citation appendix". Every sentence is its own footnote, resolving to its own source.
The delivery lead's forensic work goes away when citation has these four properties. She reads the draft the way she would read a submission from a senior engineer. The sentences she trusts stay. The sentences that do not add up, she opens the source clause and sees exactly where the AI is standing.
Gaps are surfaced, not filled
The other half of the story is what happens when the source documents do not answer the question.
A well-written bid will state the contract type, the payment terms, the escalation ladder, the risk allocation, the change-order thresholds. A less-well-written bid, or a bid with gaps the firm accepted at signing, will leave one or more of those fields silent. The 20-section project management plan needs a value for each of them. If the value is not in the source, the value has to come from somewhere.
The commercial AIs the firm tried came from the model's prior. The model had seen many project management plans. It had a plausible value for every field. It filled the fields. The plans read as if the source was complete.
This is a specific and dangerous failure. The delivery lead reads a plan that looks complete. She stamps it. The client reads a plan that looks complete. The client acts on it. Two months later, a change order comes up that the plan says has a $50,000 threshold. The contract does not actually say that. The AI generated the number.
Provenant does not generate the number. Where the sources are silent, Provenant flags the gap in the draft. The gap is marked at the exact position in the section it should occupy. The routing goes to a named human on the reviewer chain. The delivery lead sees the gap before she stamps.
There is a tempting objection to gap-flagging. It slows the pipeline. If the tool refuses to write half of section §04, the section is not done. The senior PM has to fill it in.
The counter is direct. If the tool writes half of section §04 confidently and half of that writing is wrong, the delivery lead's review takes longer than filling in the gap would have. And when a review misses one of those confident-but-wrong sentences, which happens because reviews are done by humans and humans miss things, the deliverable ships with a fabricated commitment inside it.
Refusal is the productive answer. The productivity gain has to come from correct sentences, not from more sentences.
The review chain can only downgrade
The final safeguard sits after the draft is complete. Every document type Provenant handles moves through an ordered sign-off. On a project management plan, the order is technical review, commercial review, legal review, then the delivery lead. On a contract review memo, the order is different. The document type carries its own workflow.
The final gate is enforced server-side. Nothing is released until its chain clears.
Each reviewer sees the draft, the citations behind each sentence, and the gaps still open. Each reviewer can downgrade a section from pass to hold. Only the last reviewer, the accountable professional, can convert a hold into a pass, and only by explicitly accepting the departure on the record. That acceptance becomes a signed entry in the audit trail alongside the source clauses. If the review is later challenged, the record shows who accepted what, when, and against which source.
What Provenant is actually doing
Provenant is Nuviax's grounded drafting engine. It reads a project's source documents (bid, contract, prior submittals, house templates, method statement library) and drafts the deliverable, section by section, with a verbatim citation behind every sentence. Where the source is silent, it flags a gap and routes it to a named human. When the draft is complete, an ordered review chain moves each section through sign-off with a server-enforced gate at the end.
The deliverable that comes out the other side is not longer than the manual version. It is not more readable. It is not more original. It is more defensible. Every sentence has a source. Every gap has an owner. Every acceptance has a signature.
AI Permitting and RFP AI Solution are Provenant tuned to one domain, permits and bids. Provenant is the same engine, available for every other deliverable, and for the ones you invent. Project management plans. Contract review memos. Incident investigation reports. Design and handover reports. Compliance narratives. Whatever your reviewers have to defend.
Tell us the document that eats your team's week. We will draft it, cited.