Automation, Artificial Intelligence and Emerging Technologies

Introduction

Accounting information systems (AIS) are increasingly shaped by technologies that can capture data, execute procedures, identify patterns, generate explanations and record transactions with limited manual intervention. Robotic process automation (RPA), process mining, machine learning (ML), generative artificial intelligence (GenAI), blockchain and smart contracts are no longer experimental ideas confined to technology departments. They influence transaction processing, reconciliations, audit evidence, fraud detection, financial reporting, compliance and management decision-making.

The central professional issue is not whether a technology is impressive. It is whether it produces reliable information, operates within effective controls, can be explained and reviewed, complies with applicable requirements, and improves organisational performance without creating unacceptable risks. Accountants therefore need both technological literacy and professional scepticism.

Automation should be understood as a redesign of work, not merely the replacement of people with software. A useful implementation normally combines technology, redesigned processes, data governance, internal controls and continuing human accountability.

Learning Objectives

By the end of this chapter, the learner should be able to:

1. Explain the purposes and limitations of RPA, process mining, ML, GenAI, blockchain and smart contracts in accounting.
2. Distinguish deterministic automation from probabilistic and generative systems.
3. Evaluate reliability, explainability, data quality, human oversight and control risks.
4. Design appropriate control responses for automated accounting processes.
5. Assess workforce, ethical, regulatory and governance implications.
6. Recommend an implementation approach for an accounting information system using emerging technologies.

1. Automation in the Accounting Information System

An AIS can be viewed as a combination of people, procedures, data, software, infrastructure and controls used to collect, process, store and communicate accounting information. Traditional accounting systems already automate many calculations and postings. Emerging technologies extend automation into activities that previously required interpretation, document reading, prediction or coordination across systems.

Three categories are useful:

The categories overlap. An invoice process may use RPA to retrieve a document, ML to classify it, a rules engine to approve it and a blockchain network to record a settlement. The control environment must therefore address the whole process rather than one tool in isolation.

A fundamental distinction is between automation of execution and automation of judgement. Posting a recurring depreciation entry according to an approved schedule is mainly execution. Determining whether an unusual transaction indicates fraud involves judgement and uncertainty. Technology may support the latter, but management and professional accountants remain responsible for conclusions.

2. Robotic Process Automation

2.1 Meaning and characteristics

RPA uses software agents, often called bots, to perform structured, repetitive activities by interacting with applications through user interfaces, APIs or files. A bot can log into a supplier portal, download invoices, copy fields into an enterprise resource planning system, perform validation checks and produce an exception report.

RPA is particularly suitable when a process is:

• repetitive and high volume;
• based on stable, explicit rules;
• performed using structured or consistently formatted data;
• conducted across systems that do not integrate easily; and
• subject to clearly defined exceptions.

RPA does not automatically make a poor process good. If a process contains unnecessary approvals, duplicate data entry or unclear ownership, a bot may simply perform the inefficiency faster.

2.2 Accounting applications

Common applications include:

• importing bank statements and matching receipts;
• preparing standard account reconciliations;
• transferring approved journal data between systems;
• extracting balances for management reports;
• checking supplier master data for missing fields;
• sending reminders for overdue receivables;
• assembling audit schedules and evidence indexes; and
• monitoring whether routine controls have been completed.

2.3 Controls over RPA

A bot is an information-processing user and should be governed accordingly. Important controls include bot identity management, least-privilege access, segregation of duties, change approval, controlled deployment, logging, exception handling, backup procedures and periodic review of bot performance.

A bot should not silently discard an item it cannot process. It should route the item to a named person, record the reason for failure and preserve the original evidence. Bot accounts should not use shared passwords, and credentials should be stored in an approved secure mechanism.

A useful control design distinguishes:

1. Input controls — completeness, format, authorisation and validity of source data.
2. Processing controls — approved rules, version control, duplicate prevention and error handling.
3. Output controls — reconciliation, review, exception reporting and authorised posting.
4. Continuity controls — recovery procedures if the bot, interface or source system fails.

2.4 Limitations

RPA is fragile when screen layouts, file formats, passwords, business rules or application interfaces change. It may also magnify an error: an incorrect rule can cause thousands of incorrect transactions. RPA is therefore strongest when paired with process ownership and monitoring rather than treated as a “set and forget” solution.

3. Process Mining and Process Intelligence

Process mining analyses event logs generated by information systems. An event record commonly contains a case identifier, activity name, timestamp and user or system responsible. For example, a purchase-to-pay case may include purchase requisition, purchase order, goods receipt, invoice receipt, approval and payment.

Process mining normally performs three related tasks:

• Discovery: constructing a model of how a process actually occurred from event data.
• Conformance checking: comparing actual events with a prescribed policy or reference model.
• Enhancement: adding performance, cost, risk or resource information to understand the process more deeply.

Process mining can reveal repeated invoice rework, payments made before goods receipt, approval bypasses, excessive cycle times and unusual user behaviour. It is valuable because documented procedures describe intended processes, whereas event logs reveal operational reality.

Data requirements and cautions

The quality of conclusions depends on the quality of event logs. If timestamps are missing, systems use inconsistent case identifiers or manual activities are not recorded, the discovered process may be incomplete. A process-mining model can also mistake technical events for business events. Accountants must therefore validate the meaning of fields and understand system architecture before interpreting results.

Process mining supports control testing but does not itself prove that a transaction is valid. A transaction may follow the expected route and still involve a fictitious supplier or incorrect amount. Conversely, a legitimate emergency purchase may correctly deviate from the standard route if an approved exception exists.

4. Machine Learning in Accounting

4.1 Core idea

Machine learning systems identify relationships in data and use them to classify, rank, predict or detect anomalies. In supervised learning, a model is trained using examples with known outcomes. In unsupervised learning, the system searches for structure without predetermined labels. In accounting, the output may be a probability, score, category or forecast—not a guaranteed fact.

A simplified supervised-learning relationship can be represented as:

Here, X is the input data, θ represents model parameters learned during training, f is the model, and ŷ is the predicted output. The notation is general; the particular function may be a decision tree, regression model, neural network or another method. Predictions are applicable only when the training data and operating conditions reasonably represent the environment in which the model is used.

4.2 Applications

ML can assist with:

• invoice and expense classification;
• detection of duplicate or suspicious payments;
• prediction of late customer payments;
• estimation of expected cash collections;
• identification of unusual journal entries;
• prioritisation of audit samples; and
• extraction of information from semi-structured documents.

Anomaly detection should not be confused with fraud detection. An anomaly is an observation that differs from a pattern; it may be caused by fraud, error, seasonality, a new business event or a data problem. Investigation remains necessary.

4.3 Reliability and model evaluation

A model can be accurate on historical data and perform poorly after a change in suppliers, prices, systems, regulations or economic conditions. This is known as model drift. Performance should therefore be assessed using data not used to train the model and monitored after deployment.

For binary classification, a confusion matrix distinguishes true positives, false positives, true negatives and false negatives. Two common measures are:

TP is the number of correctly identified positive cases, FP is the number of negative cases incorrectly identified as positive, TN is the number of correctly identified negative cases, and FN is the number of positive cases missed by the model. Precision is meaningful when the system has identified positive cases; recall measures the proportion of all actual positive cases detected. These measures assume that the positive class and classification threshold have been defined clearly.

For example, suppose an investigation tool identifies 80 transactions as suspicious. Of these, 60 are genuinely suspicious, while 20 are not. Investigation later shows that 75 genuinely suspicious transactions existed in the full population. Precision is 60 ÷ 80 = 75%, and recall is 60 ÷ 75 = 80%. Neither measure alone establishes whether the system is suitable. The cost of investigating false alerts and the cost of missing genuine cases must also be considered.

4.4 Bias, explainability and validation

Training data may reflect historical discrimination, incomplete records or past control failures. A model can reproduce these patterns while appearing objective. Explainability is especially important when a model affects credit limits, employee investigations, supplier approval or audit attention.

Explainability may involve feature importance, decision paths, counterfactual explanations or a simpler surrogate model. An explanation should be understandable to the responsible accountant and sufficient to support review. “The algorithm decided” is not an adequate professional explanation.

Validation should address conceptual soundness, data quality, coding, access, performance, bias, stress conditions and change management. High-impact decisions should normally require a human review and an audit trail showing the model version, inputs, output, reviewer and final decision.

5. Generative Artificial Intelligence

Generative AI produces text, images, code, summaries or other content in response to prompts. Large language models generate likely sequences based on patterns learned from training data. They do not possess human understanding in the professional or legal sense, and fluent output may be incorrect.

Accounting uses

Appropriate lower-risk uses may include:

• summarising an internal policy for a preliminary discussion;
• drafting a reconciliation commentary from verified figures;
• converting plain-language requirements into a checklist;
• proposing questions for a control walkthrough;
• assisting with spreadsheet formulas or code subject to testing; and
• searching an approved internal knowledge base using natural language.

Higher-risk uses include generating final financial statements, making unsupported accounting judgements, preparing unreviewed tax advice, approving payments or entering confidential information into an uncontrolled public service.

Hallucination and verification

A hallucination is plausible but unsupported content, such as an invented citation, incorrect accounting treatment or fabricated transaction explanation. Verification must occur against authoritative records and applicable standards. Generated narrative should never be treated as evidence merely because it is well written.

A controlled GenAI workflow should identify the approved model, permitted data, retention arrangements, responsible reviewer, source documents and escalation criteria. Prompts and outputs may need to be retained where they influence a material accounting conclusion.

Human oversight

Human oversight is not simply clicking “approve.” The reviewer must have sufficient competence, time, access to source evidence and authority to reject the output. A useful principle is that the level of review should increase with the financial materiality, legal impact, irreversibility and uncertainty of the decision.

6. Blockchain and Distributed Ledgers

A blockchain is a type of distributed ledger in which records are grouped into blocks and linked using cryptographic techniques. Participants maintain replicated records according to a consensus mechanism. The precise governance, privacy and validation properties depend on whether the network is public, private or permissioned.

Potential accounting benefits include shared visibility, reduced reconciliation between participating organisations, tamper-evident history and automated settlement. A permissioned supply-chain network, for example, could record the sequence of purchase order, shipment, receipt and invoice events for authorised participants.

Blockchain does not guarantee that an entry is true. It can preserve a false or fraudulent input with great durability. This is the “oracle” or input problem: external facts must still be entered by trusted people, devices or systems. Blockchain also raises questions about privacy, scalability, governance, correction of errors and responsibility when participants disagree.

Accounting implications

An organisation must determine whether a digital asset is inventory, an intangible asset, cash equivalent, financial asset or another category under the applicable reporting framework. The accounting conclusion depends on the rights and obligations represented by the asset, not merely on the fact that a blockchain records it. Evidence of ownership, valuation, access keys, impairment and custody also requires attention.

7. Smart Contracts

A smart contract is code that executes specified actions when programmed conditions are met. It is not necessarily a legally complete contract. Legal enforceability depends on the agreement, jurisdiction, identity of parties, applicable law and whether the code accurately expresses the intended obligations.

For example, a smart contract might release payment when a permissioned logistics system records an accepted delivery. This can reduce delay and manual intervention. However, the system must address damaged goods, partial delivery, disputed quantities, cyber compromise and incorrect sensor data. A technically automatic payment may be operationally wrong.

Controls should include code review, independent testing, access control, version management, emergency pause procedures, clearly defined dispute resolution and reconciliation between the ledger and the AIS. Immutability can make correction difficult, so organisations need an approved method for compensating entries or superseding records while preserving the audit trail.

8. A Framework for Evaluating Technologies

A professional evaluation should consider at least six dimensions.

The evaluation should begin with the business problem. A technology should be selected only if it improves a defined outcome such as shorter close time, fewer duplicate payments, better exception detection or stronger auditability. A pilot should use representative data and include failure cases, not only successful demonstrations.

Risk-based implementation stages

1. Define the process and objective. Identify the owner, inputs, outputs, risks, manual judgements and performance measures.
2. Assess process readiness. Remove unnecessary variation, document policies and improve data quality before automating.
3. Select the technology. Match deterministic tasks with rules or RPA, pattern-recognition tasks with ML and language tasks with GenAI, subject to controls.
4. Design controls. Specify access, approvals, monitoring, exception handling, records, security and fallback procedures.
5. Test and validate. Test normal, unusual, incomplete, adversarial and changed conditions.
6. Deploy gradually. Begin with a controlled pilot, parallel operation or limited business unit.
7. Monitor and improve. Review errors, overrides, drift, incidents, user feedback and control effectiveness.
8. Retire responsibly. Decommission obsolete bots, models, credentials and data stores.

9. Detailed Illustrative Case: Automated Accounts Payable

This is a fictional illustrative case designed for learning purposes.

Background

Northstar Manufacturing processes approximately 45,000 supplier invoices annually. Invoices arrive by email and portal upload. Staff manually read invoice fields, check purchase orders, obtain approvals and enter postings into the ERP system. The finance manager wants to reduce processing time and duplicate payments.

Problem

A proposed vendor solution combines optical character recognition, ML classification, RPA and a GenAI assistant. The supplier claims that 95% of invoices can be processed without human intervention. The finance team must decide whether the proposal is appropriate.

Analysis

The project team maps the process and finds that 70% of invoices have purchase orders, 20% are recurring non-purchase-order invoices and 10% contain unusual formats or disputed amounts. Historical records contain duplicate supplier names and inconsistent tax codes. The claimed 95% rate is therefore not directly transferable to Northstar.

The team identifies major risks: incorrect extraction, duplicate payment, unauthorised supplier changes, inappropriate tax treatment, overreliance on generated explanations and lack of evidence for exceptions. It also recognises that a high straight-through-processing rate could be a poor objective if it increases payment errors.

Approach and solution

Northstar first standardises supplier master data and defines tolerance rules. The ML component extracts invoice fields but does not post directly when confidence is below the approved threshold or when the invoice conflicts with purchase-order data. RPA retrieves invoices and routes them through the workflow. A deterministic matching rule checks supplier, purchase order, quantity and amount. GenAI is restricted to drafting an explanation of an exception from verified fields; it cannot approve or release payment.

Controls include dual approval for supplier-bank changes, a daily duplicate-payment report, role-based bot access, model-version records, exception queues, monthly sample review and reconciliation between invoice registers and the ERP. During the pilot, 10% of invoices are processed in parallel with the old procedure.

Outcome

The pilot reduces manual keying and shortens processing time, but only 68% of invoices qualify for straight-through processing. Several apparent ML errors are traced to poor supplier data rather than the model. Northstar therefore expands data cleansing instead of lowering review standards. Human reviewers remain responsible for disputed invoices and unusual tax treatments.

Learning points

• Vendor performance claims require testing with the organisation’s own data.
• Automation should be measured by control-adjusted value, not speed alone.
• Data governance may be more important than sophisticated algorithms.
• GenAI can support communication without being given authority to approve transactions.
• Exceptions are a normal part of a controlled process, not evidence that the project has failed.

10. Governance, Ethics and Regulation

Technology governance should assign responsibility to identifiable individuals and committees. Relevant roles may include process owner, data owner, model owner, information-security officer, internal audit and senior management. The board or audit committee should receive information about material technology risks, incidents and control weaknesses.

Important ethical issues include privacy, surveillance, discrimination, opacity, unequal access to technology and displacement of workers. Confidential payroll, customer and supplier information should not be entered into an external AI service unless contractual, security and legal requirements have been assessed. Data minimisation and purpose limitation reduce unnecessary exposure.

Financial reporting remains a management responsibility. Automation does not transfer accountability to a vendor, bot or model. External auditors also need appropriate evidence about system controls, data lineage, model changes and the reliability of automated outputs. Applicable legislation and professional standards vary by jurisdiction, so organisations should obtain current legal and regulatory advice rather than relying on general technology guidance.

11. Workforce and Professional Implications

Automation changes tasks before it eliminates entire occupations. Routine transcription and reconciliation may decline, while demand increases for data analysis, process design, systems assurance, cybersecurity, controls testing and communication with technology specialists.

Accountants should develop competence in process mapping, data quality, access controls, model limitations, analytical review and technology governance. They do not all need to become programmers, but they must be able to question system outputs and understand the evidence supporting them.

A responsible workforce transition includes consultation, training, redeployment, redesigned performance measures and clear accountability. If staff are judged solely on the number of transactions processed, they may be pressured to override safeguards. Measures should include accuracy, exception quality, control effectiveness and stakeholder outcomes.

Contemporary Developments and Best Practices

Current developments include cloud-based automation platforms, application programming interfaces, continuous close processes, privacy-preserving analytics, synthetic data for testing, foundation models connected to controlled enterprise data and increased regulatory attention to high-risk AI. These developments create opportunities but do not remove the need for basic controls.

Best practices include:

• automate only after understanding the process;
• maintain a complete inventory of bots, models, data sources and owners;
• keep human approval for material, ambiguous or irreversible decisions;
• validate outputs against independent evidence;
• monitor both performance and control effectiveness;
• document assumptions, versions, changes and exceptions;
• protect confidential data and restrict vendor access;
• test for bias, drift, cyberattack and unusual transactions;
• provide a manual fallback and business-continuity plan; and
• review whether the technology still creates value and remains compliant.

Chapter Summary

RPA automates repetitive, rule-based interaction with systems, while process mining reveals how processes actually operate through event logs. Machine learning classifies, predicts and detects patterns, but its outputs are probabilistic and require validation, monitoring and consideration of false positives, false negatives, bias and model drift. Generative AI can assist with drafting, summarisation and natural-language interaction, but fluent output may be incorrect and must be verified against authoritative evidence.

Blockchain can create shared, tamper-evident records, but it does not guarantee the truth of input data. Smart contracts execute programmed conditions and therefore require code review, legal analysis, exception procedures and controls over external data. Across all technologies, reliability, explainability, human oversight, data governance, security and accountability are essential.

The accountant’s role is evolving from manual processing toward process ownership, control design, data stewardship, analytical review and technology assurance. Successful adoption begins with a clearly defined business problem, reliable data, risk-based controls and a workforce prepared to use professional judgement alongside automated tools.

Key Terms

Automation: Use of technology to perform tasks with reduced human intervention.

Robotic process automation (RPA): Software-based automation of repetitive, rule-based activities across applications.

Process mining: Analysis of event logs to discover, compare and improve business processes.

Machine learning: Methods that learn patterns from data to classify, predict or detect observations.

Generative AI: AI that produces new text, code, images or other content from learned patterns.

Hallucination: Plausible but unsupported or incorrect output generated by an AI system.

Model drift: Decline in model performance as operating conditions or data change.

Explainability: The extent to which a system’s output and reasoning can be understood and reviewed.

Precision: The proportion of predicted positive cases that are actually positive.

Recall: The proportion of actual positive cases detected by a system.

Blockchain: A distributed ledger using linked records and a consensus process.

Distributed ledger: A shared record maintained across multiple participating systems or entities.

Smart contract: Code that executes specified actions when programmed conditions are satisfied.

Oracle problem: The challenge of obtaining trustworthy external facts for a digital system.

Event log: A record of activities, cases and timestamps generated by an information system.

Human oversight: Competent review and responsibility for automated outputs and decisions.

Data governance: Policies and responsibilities for data quality, security, use, retention and accountability.

Model validation: Assessment of whether a model is conceptually sound, accurately implemented and fit for use.

Straight-through processing: Processing completed without manual intervention for qualifying cases.

Exception handling: Identification, routing, investigation and resolution of items outside normal rules.

Review Questions

Short Answer Questions

1. What characteristics make a process suitable for RPA?
2. Distinguish process discovery from conformance checking in process mining.
3. Why is an anomaly not necessarily evidence of fraud?
4. State one benefit and one limitation of blockchain in accounting.
5. Why must GenAI output be verified against authoritative evidence?

Long Answer Questions

1. Explain how an organisation should design controls over an RPA bot used in accounts payable.
2. Discuss reliability, explainability and model drift in machine-learning applications for accounting.
3. Compare RPA, machine learning and generative AI in terms of inputs, outputs, risks and suitable accounting uses.
4. Evaluate the accounting and governance implications of implementing smart contracts for supplier payments.
5. Explain how automation may change the role, skills and responsibilities of professional accountants.

Analytical / Application Questions

1. A fraud-scoring model identifies 120 transactions as suspicious. Investigation confirms 90 are genuinely suspicious, while the population contains 150 genuinely suspicious transactions. Calculate precision and recall, and explain whether the model should be deployed without further analysis.
2. Design a controlled workflow for using GenAI to draft monthly management commentary from verified financial data. Identify at least four controls and two situations requiring human escalation.
3. A company wants to automate invoice processing but has inconsistent supplier master data and frequent manual exceptions. Recommend an implementation sequence and justify why your sequence reduces risk.

Multiple Choice Questions with Answers

Question 1

Which process is most suitable for robotic process automation?

A. A repetitive task governed by stable, explicit rules

B. A negotiation requiring professional judgement

C. A dispute involving incomplete evidence

D. A strategic decision with changing objectives

Correct Answer: A. A repetitive task governed by stable, explicit rules

Question 2

What is the primary purpose of process-mining conformance checking?

A. To generate synthetic training data

B. To compare actual process execution with a prescribed model

C. To encrypt transactions on a blockchain

D. To write explanatory accounting narratives

Correct Answer: B. To compare actual process execution with a prescribed model

Question 3

A machine-learning system identifies 50 transactions as positive, of which 40 are genuinely positive. What is its precision?

A. 20%

B. 40%

C. 80%

D. 90%

Correct Answer: C. 80%

Question 4

Which statement best describes an AI hallucination?

A. A system outage caused by excessive demand

B. A plausible but unsupported or incorrect generated output

C. A legally approved smart contract

D. A duplicate entry detected by a reconciliation

Correct Answer: B. A plausible but unsupported or incorrect generated output

Question 5

Why does blockchain not guarantee that an accounting record is truthful?

A. It cannot store timestamps

B. It cannot use cryptography

C. Incorrect external information may be entered into the ledger

D. All participants can freely delete records

Correct Answer: C. Incorrect external information may be entered into the ledger

Question 6

Which control is most appropriate for a bot that posts accounting entries?

A. Allowing the bot to use a shared employee password

B. Disabling all exception reports

C. Using approved access, logging, reconciliation and exception handling

D. Permitting the bot to change its own code

Correct Answer: C. Using approved access, logging, reconciliation and exception handling

Question 7

What does recall measure in binary classification?

A. The proportion of predicted positives that are correct

B. The proportion of actual positives that are detected

C. The percentage of all records processed on time

D. The proportion of negative cases correctly rejected

Correct Answer: B. The proportion of actual positives that are detected

Question 8

Which use of generative AI presents the highest risk?

A. Drafting questions for a control walkthrough

B. Summarising an approved internal policy for discussion

C. Generating a preliminary checklist

D. Approving payments without human review

Correct Answer: D. Approving payments without human review

Question 9

What is model drift?

A. A change in accounting standards caused by inflation

B. A decline in model performance as data or conditions change

C. A method for linking blocks in a ledger

D. A process for assigning user passwords

Correct Answer: B. A decline in model performance as data or conditions change

Question 10

Which principle is most appropriate for high-impact automated accounting decisions?

A. Human oversight should be competent and capable of rejecting the output

B. The fastest output should always be accepted

C. The vendor should retain all accountability

D. Exceptions should be deleted to simplify reporting

Correct Answer: A. Human oversight should be competent and capable of rejecting the output