ISC · Information Systems and Data Management · 6 practice questions
Auditing ERP Automated Controls: ITGCs and Data Conversion
ERP automated controls support audit reliance only when their supporting IT general controls are effective. Below: follow a new ERP's three-way match from access testing and data conversion to the reliance decision.
Try one first
Hint
Which underlying controls must an auditor evaluate before relying on an application's automated processing?
Answer C. Automated application controls depend on effective IT general controls (e.g., change management and access controls). The stem identifies ITGC weaknesses (informal change management) and an SOD gap, so the auditor should expand ITGC testing and defer reliance on automated controls; if ITGCs are not effective, increase substantive procedures over vendor transactions and balances.
Why not A: Tempting because embedded matching and a centralized vendor master can reduce manual error. However, it's incorrect: those automated controls are not reliably usable until the auditor verifies the supporting ITGCs (change management and access controls) and addresses the SOD weakness.
Why not B: Plausible because the SOD issue directly affects vendor‑master integrity. But it's wrong: testing user SOD and targeted substantive tests are necessary yet insufficient when application configuration and change‑management controls are informal, auditors must also assess ITGCs that support automation.
Why not D: Tempting because centralization and automation can suggest more consistent processing and lower error rates. However, it's incorrect: centralization increases dependence on system controls and ITGCs; given the SOD gap and informal change management, auditors should not reduce substantive testing or sample sizes.
Step by step
- Identify the control and its dependencies
Company R's new ERP matches purchase orders, receiving reports, and vendor invoices before posting payables; opening payables and vendor records were migrated from its legacy system. Staff retained elevated production access after conversion, and configuration changes lack documented approval.
- Test access before trusting the match
Identify who can change vendor records, matching tolerances, and user permissions, and test authorization, segregation of duties, and independent monitoring. Determine whether temporary conversion privileges were removed and whether privileged activity was reviewed throughout the intended reliance period.
- Test changes and IT operations
Inspect evidence that production configuration changes were authorized, tested, and approved before implementation. Test relevant IT operations controls, including processing and exception monitoring, for the period of intended reliance.
- Verify the conversion
Reconcile legacy opening payables to ERP cutover balances and control totals. Reperform selected vendor mappings and trace migrated invoices to source records, emphasizing manual overrides and ambiguous matches to test completeness and accuracy.
- Test the automated match
Test whether the configured quantity and price tolerances accept permitted matches and block exceptions, with evidence covering the intended reliance period. Reduce substantive testing only to the extent supported by effective relevant ITGCs, reliable conversion results, and effective application controls.
- Respond to unresolved weaknesses
If Company R's elevated access or unapproved changes undermine the match, test compensating controls or expand substantive procedures over affected vendor balances and transactions. Communicate identified deficiencies promptly to appropriate recipients and do not treat planned remediation as evidence of current effectiveness.
Key points
- Remember: someone who can change the matching rules can undermine the match without changing an invoice.
- Matching total payables does not prove that invoices reached the correct vendor records.
- A promised access fix does not provide evidence that controls operated effectively before the fix.
How the exam traps you
- Reduce substantive testing because the ERP successfully matches invoices. Test supporting ITGCs and the match itself before relying on automation.
- Accept conversion because total payables agrees with the legacy system. Also trace individual invoices and vendor mappings; allocation errors can leave totals unchanged.
- Wait for planned remediation without changing the audit approach. Respond now with tested compensating controls or expanded substantive procedures, and communicate deficiencies promptly.
Question 2
Hint
AI = umbrella (human-intelligence-mimicking systems). ML = SUBSET of AI (systems that learn from data). RPA = SEPARATE category (rule-based task automation via bots). Intelligent automation = RPA + AI/ML.
Answer A. AI, ML, and RPA are related but DISTINCT technology categories with different scope and purpose. ARTIFICIAL INTELLIGENCE (AI) is the BROADEST term: any computer system performing tasks that typically require human intelligence (perception, language understanding, reasoning, problem-solving, learning). Includes both rule-based expert systems and learning-based systems. MACHINE LEARNING (ML) is a SUBSET of AI: systems that LEARN PATTERNS from data, improving performance with experience without being explicitly programmed for each scenario. Examples: image recognition trained on labeled photos; demand forecasting trained on historical sales; fraud detection trained on transaction patterns. ML includes supervised, unsupervised, reinforcement, and deep learning approaches. ROBOTIC PROCESS AUTOMATION (RPA) is a SEPARATE technology category: software 'bots' that AUTOMATE REPETITIVE TASKS by mimicking human user actions (clicking, typing, copying data between systems). Traditional RPA is RULE-BASED, it follows pre-programmed scripts and does not learn or adapt. Examples: RPA bots reconciling invoices between ERP and bank statements; updating customer records across multiple systems; processing form-based data entry. RELATIONSHIPS: AI ⊃ ML (ML is a subset of AI). RPA is INDEPENDENT of AI/ML, pure RPA is rule-based automation, not AI. INTELLIGENT AUTOMATION = RPA + AI/ML, where bots can handle non-deterministic decisions (e.g., reading invoices with varied formats using OCR + ML). Halcyon's IT manager should distinguish: (1) AI/ML for prediction, classification, and learning tasks. (2) RPA for repetitive rule-based automation. (3) Intelligent automation when both are combined. Answer C is incorrect: the three terms are NOT interchangeable; they have distinct technical meanings and different use cases. Answer B is incorrect: ML is a SUBSET of AI, not the broader concept. Answer D is incorrect: RPA does NOT inherently include AI/ML; pure RPA is rule-based.
Question 3
Hint
AI/ML audit toolkit: data audit, reproducibility, holdout testing, fairness metrics, explainability TOOLS (SHAP/LIME, quantitative), governance docs, monitoring. NOT useful: 'interviewing the model' for reasoning, models don't think; outputs are post-hoc, unreliable.
Answer C. This question targets AI/ML AUDIT PROCEDURES, meaning what auditors should do (and not do) when evaluating ML models. The LEAST USEFUL procedure is interviewing the model directly to verify that it 'thinks' the same way a human loan officer would. ML models do not 'think' or 'reason' in a human sense; asking a model 'why did you approve this loan' yields outputs that may sound reasonable but are post-hoc explanations, not the actual computation. EFFECTIVE AI/ML AUDIT PROCEDURES include: (1) DATA AUDIT: check training data for representativeness (does it reflect the population the model will see?), completeness (missing data?), labeling quality (errors in ground truth?), and known biases (historical discrimination patterns?). Garbage in, garbage out. (2) MODEL REPRODUCIBILITY: can the model be retrained from the same data and code, yielding the same model? Reproducibility is foundational to model risk management. (3) TESTING: test on a holdout set; check accuracy, precision, recall, fairness metrics (disparate impact ratio), drift over time. (4) EXPLAINABILITY TOOLS: SHAP, LIME, and feature importance quantify which inputs drive predictions, supporting compliance with adverse action notice requirements (ECOA). These are MATHEMATICAL explanations, not natural-language interviews of the model. (5) GOVERNANCE: documented model approval, validation, monitoring procedures (SR 11-7 model risk management guidance). (6) MONITORING: performance and fairness metrics tracked over time; retraining triggers; champion/challenger model framework. WHY 'INTERVIEWING THE MODEL' IS PROBLEMATIC: Generative models can produce plausible-sounding rationales that don't reflect actual computation. For non-LLM ML models (like loan approval classifiers), 'interviewing' isn't even technically possible. Even for LLMs, 'asking why' produces unreliable explanations, known as the 'inner monologue problem' in interpretability research. Auditors should rely on quantitative explainability methods (SHAP, LIME, partial dependence plots), not narrative explanations. Re-performing the training pipeline to compare the resulting model to the production model is a USEFUL procedure and therefore not the answer. Reproducibility checks are foundational. Inspecting the training data for representativeness, completeness, labeling quality, and known biases is a USEFUL procedure and therefore not the answer. Data quality audits are essential; most ML failures trace to data issues. Testing the deployed model against a holdout test set and benchmarking accuracy, fairness metrics (disparate impact), and stability over time is a USEFUL procedure and therefore not the answer. Performance and fairness testing are the heart of model validation.
Question 4
Hint
Think about procedures that directly test the conversion logic and trace individual legacy records into the new ERP; include records with manual overrides or ambiguous matches in your sample.
Answer A. Re-running the conversion routine on a representative sample, especially those with manual overrides or ambiguous matches, directly tests the mapping logic and permits tracing of individual legacy records to new customer records, revealing split or duplicate master records and misallocated invoices. Reconciling control totals addresses completeness of migration, and duplicate-master detection (tax ID, address, etc.) identifies cases where a single legacy customer was split across multiple new customer numbers. Together these procedures provide direct evidence about both accuracy and completeness of the AR migration.
Why not B: Tempting because an aggregate balance comparison is a straightforward, high-level check that may reveal material completeness issues. However, totals can mask mapping or allocation mistakes (for example, invoices split among duplicate customer numbers or offsetting misallocations), and this approach does not test the conversion logic or trace individual legacy records.
Why not C: Tempting because UAT documentation and confirmations are common controls that support existence and client testing. However, UAT is management-controlled and may not exercise ambiguous-match cases; confirmations substantiate selected balances but are unlikely to detect systematic mapping errors or duplicate customer masters created during conversion.
Why not D: Tempting because reconciliations can show whether subsidiary ledger totals tie to the GL after migration. However, if invoices were migrated but assigned across duplicate customer records, subsidiary ledger and GL totals may still reconcile; reconciliations do not test the conversion mapping logic or identify split/duplicate master records.
Question 5
Hint
RPA suitability check: stable structured inputs + rule-based logic + high volume = good fit. Variable inputs break RPA, fix with input standardization OR AI augmentation (Intelligent Automation = RPA + AI/ML).
Answer D. This is a classic RPA SUITABILITY MISMATCH. Traditional RPA works best for tasks with: (1) HIGH VOLUME (worth automating); (2) STABLE STRUCTURED INPUTS (predictable formats); (3) RULE-BASED LOGIC (no judgment); (4) DETERMINISTIC RULES (same input → same output). When inputs vary unpredictably, RPA fails, because the bot follows pre-programmed scripts and doesn't adapt. SOLUTION OPTIONS: (1) STANDARDIZE INPUTS: require vendors to submit via a portal with standardized fields, eliminating format variability. The cleanest fix when feasible. (2) AUGMENT WITH AI/ML (Intelligent Automation): add OCR to read scanned/blurry invoices, ML/NLP to parse varied email formats, exception routing for low-confidence cases (human-in-the-loop). Addresses variability but adds complexity, cost, and ongoing model maintenance. (3) RETHINK THE PROCESS: maybe direct ERP-to-bank API integration eliminates invoice-level reconciliation altogether. KEY LESSON: RPA is not a panacea. Selecting RPA without assessing input variability is a common implementation failure. Evaluate process suitability BEFORE deploying RPA: structured/predictable inputs = good RPA candidate; varied/unstructured inputs = needs AI augmentation or process redesign. Answer A is incorrect: traditional RPA does NOT learn. Bots follow pre-programmed rules and break when inputs deviate. Answer B is incorrect: RPA IS appropriate for stable, structured invoice reconciliation processes, many companies use it successfully. The issue is suitability for THIS specific process. Answer C is incorrect: different RPA vendors have similar core capabilities; switching won't fundamentally change RPA's brittleness to input variability.
Question 6
Hint
Focus on what integration does well and what it does not do by itself. Automation improves processing, but it does not automatically remove the need for controls.
Answer B. An integrated ERP system shares data across functions, which can improve consistency and reduce redundant data entry. However, automation does not eliminate control risks, errors, unauthorized access, or improper system changes can propagate through an integrated environment. Therefore effective input validation, access controls, and change-management controls remain necessary.
Why not A: This is tempting because automatic updates seem more reliable than manual processes. In reality, integration means a bad input or processing error can flow through several modules (inventory, receivables, general ledger), so automation can propagate errors if controls fail.
Why not C: Standardization and automation do not replace segregation of duties. Segregation remains a core control; in ERP systems it is typically enforced through roles, permissions, and workflow approvals rather than manual separation, but it is not eliminated.
Why not D: Keeping separate databases would negate the main advantage of an ERP's shared database. Auditability is maintained through audit trails, logs, and access controls rather than by separating operational and accounting data into different databases.
Common questions
Why do ITGCs matter when auditing an ERP three-way match?
The match depends on protected configuration and master data. Uncontrolled privileged access can change tolerances or vendor records, undermining apparently successful automated processing.
Is reconciling total payables enough to test ERP data conversion?
No. Equal totals can hide invoices assigned to the wrong vendor. Test individual mappings and migrated invoices as well as cutover control totals.
Can auditors rely on ERP controls after ITGC weaknesses are fixed?
Remediation alone is not enough. Test whether the relevant controls operated effectively during the period of intended reliance, and separately address the earlier affected period.
Watch it solved
A full CPA ISC task-based simulation on Enterprise and accounting information systems, worked step by step.
ISC Simulation: Mapping Controls Through Order-to-Cash on YouTube
Practice ISC like the real exam
The free ChatCPA simulator: real exam layout, timed testlets, starting with a question on this topic. No account needed to start.
Open the free simulator →More on Enterprise and accounting information systems
- ERP Cores, Hybrid Architectures, and Best-of-Breed
- AP-to-GL Batch Recovery: Fix, Reprocess, and Reconcile
- ERP Revenue Recognition: Shipment vs. Invoice Date
- ERP Vendor Master Data Governance: One Controlled Source
- ERP Data Residency: When Local Laws Govern Deployment
All Enterprise and accounting information systems practice →
Questions from the ChatCPA bank of 17,658 CPA exam questions, each with a written reason for every wrong answer. ChatCPA is built by Nicholas Miller, CPA (Oregon #14907). How these pages are made. Spot an error? Tell us.