The most common concern we hear from senior auditors evaluating Pramaana is that they do not want a system making substantive audit judgments for them. That concern is reasonable and well-founded. It is also based on a misunderstanding of what document intelligence in the audit context actually does.
I want to address this directly, because getting it wrong in either direction creates problems. If we understate what the technology does, we obscure its value and teams cannot make informed decisions about adoption. If we overstate it, we create expectations that lead to misuse and, more seriously, to auditors relying on automation in places where professional judgment is required.
What document intelligence does and does not do in audit contexts
Document intelligence for audit documentation performs one core function: it reads heterogeneous source documents, identifies numerical claims within them, and builds citation links between those claims and corresponding figures in working papers. That is it.
This function is document retrieval, not audit judgment. The system does not determine whether a figure is materially correct. It does not assess whether the accounting treatment is appropriate. It does not evaluate the quality of the source document or the reliability of the underlying data. Those are professional judgments, and they remain entirely with the auditor.
What the system eliminates is the mechanical retrieval work that precedes judgment: locating the source document, finding the specific passage, confirming the numerical match, building the citation link. This work consumes significant time in every audit cycle, and it does not require professional judgment. It requires patience, familiarity with the document set, and the ability to navigate heterogeneous file formats. Automation handles those requirements more consistently and much faster than human review.
Why experienced auditors should want this
The objection to automation in audit work often comes from a correct intuition: the audit function exists because human judgment over evidence matters, and anything that removes humans from that loop is a threat to audit quality. This intuition is right. The error is in applying it to document retrieval, which is not a judgment function.
When an experienced auditor spends 40 minutes locating a board package and identifying the table that contains a revenue figure, that 40 minutes is not audit work in any substantive sense. It is document retrieval. The auditor is using skills that overlap with good file management, not the professional expertise that makes them valuable to the audit process.
Reducing time on document retrieval means the same amount of time is available for the work that does require professional judgment: evaluating whether the accounting treatment is appropriate, assessing the reasonableness of estimates, identifying anomalies that merit investigation, talking with client management about what the figures actually represent. Those activities benefit from having more experienced time allocated to them, not less.
The teams using Pramaana who report the highest satisfaction are not the teams who have reduced headcount. They are the teams where senior auditors have reclaimed time from document retrieval and are using it for substantive analysis. The quality improvement is in what the audit actually produces, not in how quickly it gets done.
Where the technology is genuinely not trustworthy
Document intelligence produces confidence scores for every citation link it creates. A match between a working paper figure and a source document passage might have a confidence of 96%, meaning the system is highly confident that the specific passage produced the specific figure. Or it might have a confidence of 72%, meaning the figure is plausibly derived from that passage, but there is meaningful uncertainty.
Low-confidence links are flagged for human review, not presented as established citations. This is a design choice, not a product limitation. We believe that presenting uncertain matches as established citations would be a misuse of the technology in an audit context, where the difference between a 72% match and a 96% match has material implications for the defensibility of the evidence chain.
There is also a category of figures for which automated citation is genuinely not possible: estimates, reserves, fair values, and other figures that rest on management judgment rather than on a direct citation from a document. For these figures, Pramaana captures the relevant source documents but does not generate an automated citation. The auditor reviews the documents and makes the judgment. The system supports that judgment by having the documents ready and organized, but it does not short-circuit it.
The preparation time reduction: where it comes from
In practice, the preparation time reduction from document intelligence comes from three specific activities that are eliminated or drastically shortened.
First, document retrieval. Finding the right version of a source document, navigating to the specific passage, and confirming the numerical match account for a large share of pre-fieldwork preparation time. Automated citation eliminates this activity for high-confidence matches, which represent the majority of straightforward financial figures.
Second, evidence package assembly. Building the evidence package for a regulatory examination or external audit review currently requires manually pulling supporting documents for each figure in the working papers, organizing them against the working paper structure, and confirming completeness. When citation links are maintained continuously in a provenance graph, evidence package assembly becomes a navigation exercise through an existing structure rather than a fresh assembly from scratch.
Third, responding to follow-up requests during fieldwork. When external auditors ask for substantiation of specific figures, the internal team typically needs to locate and produce supporting documentation within hours. With pre-built citation links, the response is nearly immediate: here is the figure, here is the source document, here is the specific passage, here is the confidence level. Follow-up requests that previously consumed half a day are resolved in minutes.
The skeptical case for trying it on a limited scope first
Experienced auditors are rightly skeptical of technology claims in their field. Audit automation has been oversold before, and the consequences of relying on automation where judgment was required have been significant. That history earns skepticism.
The most credible way to evaluate document intelligence for audit use is to apply it to a bounded scope: one project, one set of working papers, one set of source documents. Compare the citation links the system produces against what the audit team would have produced manually. Check the confidence scores against the actual quality of the matches. Identify the cases where the system flagged uncertainty and verify whether that uncertainty was warranted.
That evaluation will tell you more than any description of what the technology does. It will show you specifically where automation is reliable enough to use directly, where it requires human review before use, and where it does not apply at all. Those boundaries are different for every organization's document environment, and understanding them before relying on them in a live audit cycle is the right approach.
The Pramaana demo process is structured around exactly this evaluation: we onboard a sample of your real working papers and source documents, run the citation process, and walk through the results with your team. The goal is not to persuade. It is to show you specifically what the technology does on your documents, with your figures, against your evidence base. That is the only basis for a judgment about whether it belongs in your audit workflow.
Document intelligence does not replace auditor judgment. It removes the preparation work that prevents auditors from spending their time on the work that judgment actually requires. The distinction matters for how you evaluate the technology, how you deploy it, and what you should expect it to produce. Getting that distinction right is the first step toward using it well.