Audit fieldwork (test of controls if applicable, substantive audit procedures, including
test of details or substantive analytical procedures, evidence gathering, review of defi-
ciencies and determining whether the auditor needs additional audit evidence)
Artificial intelligence: Fieldwork opportunities (continued)
HR and entity policy data can also be layered into approval reviews, for example, an inappropri-
ate ordering of preparers and reviewers (e.g., a manager reviewing a vice president’s work), or
reviews that have taken place that do not meet the designated level of authority for review.
Extracting information from external and internal support
Building on the OCR tools, there are opportunities to both automate and use AI in substantive
testing (e.g., vouching of invoices). Through combinations of OCR, NLP and natural language
generation to apply AI-enabled RPA, the AI tool can read and cross validate sub-ledger records
with external confirmations (e.g., bank or debtor confirmations) or other relevant appropriate
audit evidence. The process is simpler where external confirmation formats are more standard-
ized and becomes more complex where an array of layouts is used.
Scanning documents and extracting information through some of the techniques listed above
are the simple first steps in this process.
Machine learning techniques can enable the tool to learn how to identify the right information.
Over time, a tool can be trained on how to identify the relevant information from the source doc-
ument. Gradually, the tool will begin to require fewer and fewer training instances and will begin
to identify the relevant information in documents with formats outside of the training examples.
Estimates
The assessment of management’s estimates is a key and complex area of any audit – one that
requires significant auditor judgment. However, in some cases, management may come up with
an estimate for which AI can be used as part of the audit process.
Traditional audit techniques used to audit estimates will typically fall into one of three categories
(or a combination of the three): reperformance of management’s process; retrospective testing;
or development of an independent estimate. An array of automation and AI techniques can be
used to perform variations of these techniques.
For example, in estimating the likelihood of non-repayment for a debtor or bad debt provision,
management has set a rate at which they believe the likelihood of default is expected. Using
machine learning, the audit team could build an independent model to predict this likelihood
based on historical bad debt write-os. Once the model is built, it could be retrained every year
based on actual loss data. This independent estimate could be made for each individual loan (or
by portfolio or type of loan), and then compared to the result of management’s estimate. The AI
tool could also be trained to incorporate other relevant observable factors, such as: interest rate
movements, customer credit ratings, share price, contractual terms, housing starts and unem-
ployment rates. Inclusion of these factors could also enable determination of an independent
expected loss estimate for comparison with the client’s estimate.
While the audit team would still need to understand the underlying data as well as manage-
ment’s methodology, a machine-learning model would provide a more comprehensive estimate
of the likelihood of default. Information gathered across industries and geographical locations
could also provide the auditor with industry information to come up with expected loss provision
by customer.
22
The Data-Driven Audit: How Automation and AI are Changing the Audit and the Role of the Auditor
APPENDIX A: AUDITING WITH AUTOMATION, ANALYTICS AND AI