Most audit documentation systems are physical filing systems with a digital overlay. The underlying logic has not changed in decades: gather supporting documents, attach them to the relevant working paper section, and retain everything in an organized folder structure. The medium shifted from binders to shared drives, but the model stayed the same. Documents are stored near their working papers, not linked to the specific figures those documents support.
This distinction matters more than it sounds. When documentation is stored near a working paper, retrieval requires a person who knows where to look. When documentation is linked to a specific figure, retrieval is a property of the figure itself. The second model is what we mean by a provenance chain, and the shift from one to the other is what document intelligence actually enables in audit contexts.
What a paper trail actually is
The phrase "paper trail" implies sequence: one document leads to the next, and following the trail gives you the full story. In practice, audit paper trails are rarely sequential. They are more like a pile of evidence that collectively supports a set of conclusions, organized by someone's judgment about what should sit next to what.
The working paper structure imposes order. Section 3.2 covers revenue recognition. Section 3.2's supporting folder contains the board package, the revenue schedule, and three emails from the CFO. An auditor who looks at section 3.2 can find those documents, provided they know to look there and the documents were filed correctly.
What the paper trail does not tell you, explicitly, is which document produced which number. That link exists in the auditor's head, or in a brief note scribbled in the margin of a printed spreadsheet. It is not a structured property of the documentation system. It is implicit knowledge that degrades over time.
What a provenance chain is instead
A provenance chain treats each figure in a working paper as a node in a graph. Every node has at least one edge pointing to the source record that produced it: a page in a PDF, a cell in a spreadsheet, a paragraph in an email. When a figure passes through intermediate calculations, those intermediaries are also nodes, with their own edges pointing to their sources.
The result is a navigable structure. You can start at a working paper figure and follow the chain backward to the original source record. You can also start at a source document and follow the chain forward to every working paper figure that depends on it. When a source document is updated, you know immediately which figures are affected.
This is not a new concept in data engineering. Lineage graphs are common in analytics pipelines. What is new is applying the same structural logic to audit documentation, where the "records" are heterogeneous: PDFs, Excel files, email threads, SharePoint exports, each with different document structures and different ways of presenting numerical evidence.
Where document intelligence fits
The gap between a paper trail and a provenance chain has historically been too expensive to close manually. Building explicit citation links from every working paper figure to its source would require someone to annotate every cell at the time of entry, maintain those annotations as documents are revised, and propagate changes through calculation chains. The overhead would be prohibitive.
Document intelligence changes the cost calculus in two specific ways.
First, modern NLP models can read heterogeneous document formats and extract structured numerical claims with high confidence. A board package is a PDF with tables, prose, and footnotes. A document intelligence system can parse that structure, identify the table cells and prose passages that contain financial figures, and represent them as addressable records in a citation graph. This extraction step, which would take an analyst hours per document, runs in minutes.
Second, the matching problem between a working paper figure and a source document passage is tractable with the right combination of numerical matching and contextual embedding. A figure of $4.2M in a working paper cell can be matched to the passage in a board package that says "Q3 net revenue of $4.2 million" with high confidence, especially when the surrounding context (account labels, period references, document section) is used as additional signal.
The combination of these two capabilities is what makes automated provenance chain construction viable. The system reads the source documents, extracts the figures and their context, and builds the citation graph automatically as working paper cells are populated or imported. Human review flags the cases where confidence is below threshold, rather than reviewing every cell.
What "audit-ready" means in a provenance chain model
In a paper trail model, audit-ready means having the documents organized and accessible. An audit-ready team can respond to PBC requests within hours because they know where everything is filed.
In a provenance chain model, audit-ready means having the citation links intact. An audit-ready team can respond to any question about any figure by navigating the graph: here is the figure, here is the source it came from, here is the document version, here is the timestamp of when the link was established.
We are not arguing that one is strictly better than the other in every dimension. Paper trail discipline has its own value: it forces teams to organize supporting documentation systematically, which has independent benefits for review efficiency. The point is that paper trail organization is a necessary condition for audit readiness, not a sufficient one. You can have well-organized files and still be unable to answer the specific question: which part of which document produced this figure?
Provenance chain documentation answers that question directly. It is the layer that sits between organized filing and defensible substantiation.
The practical difference during fieldwork
Consider what happens when an external auditor requests substantiation for a revenue figure during fieldwork. In a paper trail model, the internal team locates the supporting folder, pulls the board package and revenue schedule, and hands them over with a verbal explanation of how the figure was derived. The external auditor reviews both documents, asks follow-up questions, and the process takes half a day.
In a provenance chain model, the internal team exports a trace report for that figure: a structured document showing the working paper cell, the source document and specific location, the confidence level of the match, and any intermediate calculation steps. The external auditor can verify the chain mechanically rather than inductively. Follow-up questions focus on judgment calls and edge cases, not on basic substantiation. The process takes an hour.
The time difference compounds across an entire audit cycle. The early-access teams using Pramaana report that pre-traced evidence packages reduce fieldwork substantiation requests by roughly half, because the most common questions are already answered in the trace output.
Where the model has limits
Automated provenance chain construction works well for numerical figures with clear source matches. It works less well for figures derived from judgment: estimates, reserves, fair values, or figures that require narrative reconciliation across multiple documents rather than a single traceable source.
For those figures, the provenance chain approach records what it can: the source documents that are relevant, the confidence level of the association, and any manual annotation the auditor provides. The chain is partial rather than complete. A partial chain with explicit confidence levels is still more defensible than an implicit link that exists only in someone's memory, but it is not the same as a fully automated trace.
We think it is important to be direct about this boundary. Document intelligence does not eliminate auditor judgment from the substantiation process. It eliminates the mechanical portion of substantiation: finding and linking the sources. The judgment about whether those sources adequately support a figure still belongs to the auditor. That is where it should belong.
The shift from paper trail to provenance chain is a shift in infrastructure, not in professional responsibility. Internal audit teams that adopt it do not change what they are accountable for. They change how much time they spend on mechanical retrieval versus substantive analysis. In our experience, that shift is significant enough to change the character of what audit fieldwork looks like in practice.