Production costing has moved from periodic spreadsheet preparation toward continuous, digitally integrated cost management. Modern manufacturing organisations collect transactions from enterprise resource planning (ERP) systems, machine data from industrial Internet of Things (IoT) devices, production events from manufacturing execution systems (MES), and commercial information from purchasing, sales and supply-chain platforms. These data can support faster quotations, more accurate product costs, early detection of waste and better operational decisions.

Digitalisation, however, does not automatically create reliable costing. A digitally generated cost can still be wrong if material masters are incomplete, machine rates are outdated, scrap is misclassified, or data are manipulated. The central managerial issue is therefore not simply technology adoption but the design of an integrated socio-technical cost system: technology, accounting rules, operational processes, people, controls and ethical safeguards must work together.

This chapter examines ERP, MES, IoT, digital twins, robotic process automation (RPA), cloud costing and artificial intelligence (AI). It also develops governance, cybersecurity, auditability and human-oversight requirements for responsible digital costing.

Learning objectives

1. The digital cost-management architecture

A production-cost system converts business events into monetary information. Typical events include purchase receipts, material issues, labour time, machine running time, energy consumption, production quantities, quality failures, rework, inventory movements and customer orders. In a traditional system, these events may be recorded at different times and reconciled manually. In a digital system, they are linked through shared identifiers such as item number, work-centre code, operation number, batch number, order number and timestamp.

Layer	Main role	Typical information
Enterprise layer	Financial and commercial integration	General ledger, purchasing, inventory, sales, suppliers
Execution layer	Control and record production	Orders, labour, quantities, downtime, quality events
Physical-data layer	Capture real-world conditions	Sensors, machine controllers, energy meters
Analytical layer	Explain and predict cost behaviour	Drivers, forecasts, anomalies, simulations
Governance layer	Ensure reliability and responsible use	Access, definitions, approvals, audit trails, security

The architecture should preserve the distinction between actual, standard, estimated and forecast cost. An automated dashboard that combines these categories without clear labels can mislead managers even when the underlying arithmetic is correct.

ERP systems

An ERP system integrates core organisational processes around a common database and transaction structure. For production costing, relevant ERP modules commonly include materials management, production planning, inventory, purchasing, finance, controlling, sales and maintenance. ERP supports the financial consequences of production: goods receipts, material issues, work-in-progress, finished-goods receipts, supplier invoices, labour postings and overhead allocations.

ERP costing may use standard costs, actual costs, moving-average costs or other approved valuation methods. A standard cost generally represents an approved benchmark for a defined period. Variances are then calculated by comparing actual results with the standard. For example, a material price variance can be expressed as:

The result is expressed in currency. The formula assumes that actual and standard prices refer to the same material, unit of measure and purchasing conditions. A positive result normally indicates an unfavourable variance when actual price exceeds standard price, although organisational reporting conventions must be stated.

ERP is strong at transaction integrity and financial integration, but it may not capture machine-level events at sufficient speed or detail. It therefore requires integration with MES, sensors and specialist analytics.

Manufacturing execution systems

An MES manages and records production execution between planning systems and the shop floor. It may dispatch work orders, record start and completion times, capture quantities, trace batches, document quality results and identify downtime reasons. MES data improve costing by replacing broad averages with operational evidence.

Suppose a work centre produces 1,000 good units, consumes 42 machine-hours and records 100 minutes of planned setup. The MES can distinguish productive time, setup, waiting, breakdown and quality-related stoppage. Such distinctions support more meaningful cost drivers than simply allocating all overhead by units produced.

Industrial IoT

Industrial IoT refers to connected sensors, machines and devices that collect, transmit and sometimes act upon physical data. Examples include spindle speed, temperature, vibration, power consumption, pressure, cycle time and machine status. IoT data can reveal the resource consumption associated with individual jobs or operating conditions.

If a machine consumes 18 kWh during a job and the rate is 0.20 currency units per kWh, energy cost is 3.60 currency units. The calculation assumes that the meter is calibrated, the measured interval corresponds to the job, and the tariff is correctly applied. Demand charges, time-of-use rates and shared loads may require a more detailed model.

IoT creates new risks. Sensors may drift, clocks may be unsynchronised, network outages may produce missing data, and a connected device may become an entry point for cyberattack. Data should therefore include quality status, timestamp, source identifier and calibration information.

Digital twins

A digital twin is a continuously updated digital representation of a physical asset, process or production system. Unlike a static simulation model, a useful twin receives data from its physical counterpart and can compare predicted and observed performance. In costing, a digital twin can estimate the effects of cycle time, tool wear, maintenance policy, scrap, energy use and production mix.

For example, a twin of a machining cell may simulate two maintenance policies. Preventive maintenance may increase planned downtime but reduce breakdowns, emergency repairs and scrap. The costing decision should compare total relevant cost, not merely maintenance expenditure. A simple expected-cost model is:

This is a simplified decision model. It assumes that probability and consequence estimates are meaningful and that the alternatives are comparable over the same time horizon. A digital twin does not eliminate uncertainty; it makes assumptions and scenarios more visible.

Robotic process automation

RPA uses software bots to execute structured, repetitive, rule-based tasks through applications or interfaces. In costing, RPA can import supplier price lists, reconcile purchase orders with invoices, extract approved rates, generate variance reports and route exceptions for review. RPA is particularly suitable where rules are stable and inputs are digital.

RPA should not be confused with AI. A bot following a rule such as “copy the approved monthly machine rate into the costing template” is automation, not necessarily learning. RPA also reproduces errors if the source data or rule is wrong. Controls should include bot identity, version control, exception queues, execution logs, segregation of duties and periodic review.

Cloud costing

Cloud costing uses remotely hosted computing, databases and applications accessed through networks. It can provide scalable processing, shared dashboards, rapid software updates and access across plants. Cloud systems may be delivered as software-as-a-service, platform-as-a-service or infrastructure-as-a-service.

Advantages include lower local infrastructure requirements and easier multi-site standardisation. Risks include vendor dependence, data-location issues, connectivity failure, access misconfiguration, subscription-cost escalation and uncertainty about data ownership. A cloud contract should define availability, backup, recovery time, incident notification, portability, subcontractors, confidentiality and deletion of data after termination.

2. From data to cost drivers

A cost driver is a measurable factor that causes or explains a change in cost. Traditional drivers include direct labour-hours, machine-hours, material quantity and production volume. Digital systems enable more granular drivers such as number of setups, tool changes, inspection events, waiting minutes, energy consumed, distance travelled by automated vehicles, changeover complexity and number of engineering revisions.

Activity-based costing assigns resource costs to activities and then assigns activity costs to products using cost drivers. If an inspection department costs 60,000 currency units and performs 1,200 inspection events, the activity rate is:

The rate is 50 currency units per inspection event. The denominator should preferably reflect practical capacity rather than an unusually low actual volume, because using actual volume can make unused-capacity cost appear as a high product cost.

AI can assist cost-driver discovery by examining large datasets. Candidate drivers may be identified through correlation analysis, regression, decision trees or clustering. However, association is not proof of causation. A model may find that product family is strongly associated with cost because product family indirectly represents complexity, material grade or routing. Subject-matter experts must interpret the relationship and check whether it is stable and operationally meaningful.

3. AI applications in production costing

Cost-driver discovery

Supervised learning uses labelled outcomes, such as actual conversion cost per order, to estimate relationships between explanatory variables and cost. Unsupervised learning searches for structure without a predefined outcome, for example grouping orders with similar setup and quality patterns.

Here, C is estimated cost, β₀ is the intercept, βᵢ are estimated coefficients, xᵢ are cost-driver values, and ε is the unexplained error. The model is useful when relationships are reasonably stable and additive. Nonlinear models may be more accurate but are often less transparent.

Managers should assess out-of-sample performance, not merely how well the model fits historical data. Important checks include data leakage, multicollinearity, seasonality, structural change, missing values and whether the model remains valid when prices, product designs or production technology change.

Anomaly detection

An anomaly is an observation that differs materially from expected behaviour. Examples include an unusually high scrap rate, duplicate material issue, impossible machine runtime, sudden energy increase or labour posting outside a shift.

A basic statistical rule identifies an observation as unusual when it lies beyond selected limits around a reference mean. For a normally distributed measure, a commonly used screening rule is:

where x is the observation, μ is the reference mean and σ is the standard deviation. A large absolute z-value may indicate an anomaly. The rule assumes that the reference period is representative and that the measure is sufficiently stable. Manufacturing data are often skewed or autocorrelated, so robust methods, control charts or process-specific limits may be preferable.

An anomaly is not automatically fraud or error. It may represent a genuine new product, an approved overtime shift, a sensor fault or an unusual but valid customer order. The correct response is investigation and classification.

Forecasting

AI forecasting can estimate future material prices, demand, labour requirements, machine availability, energy cost and production cost. A simple forecasting model may use historical observations and explanatory variables such as product mix, order volume, supplier lead time and commodity prices.

Forecast accuracy should be measured using an appropriate metric. Mean absolute error is:

where yᵢ is the actual value, ŷᵢ is the forecast, and n is the number of observations. MAE has the same units as the forecasted quantity and is easy to interpret. No single metric is universally best; percentage measures can be problematic when actual values are zero or close to zero.

Forecasts should include uncertainty ranges where possible. A single point estimate can encourage false precision, particularly during supply disruption or major process change.

Human oversight

Human oversight is essential because costing decisions have financial, operational and employment consequences. A responsible system uses a risk-based model:

Users must be able to understand the purpose, inputs, limitations and confidence of an AI output. “The algorithm recommended it” is not an adequate managerial justification.

4. Governance, cybersecurity and auditability

Data governance

Data governance defines who may create, change, use and approve data. Core practices include a data dictionary, master-data ownership, validation rules, version control, retention schedules and stewardship responsibilities. A costing data dictionary should define terms such as standard cost, actual cost, scrap, rework, productive time, downtime and capacity.

Data quality can be assessed through dimensions such as accuracy, completeness, consistency, timeliness, uniqueness and validity. A dashboard should distinguish missing data from zero values; these have different managerial meanings.

Cybersecurity

Cybersecurity protects confidentiality, integrity and availability. In production costing, integrity is especially important: unauthorised changes to material prices, routing times or machine rates can distort decisions without immediately appearing as a system failure.

Controls include multi-factor authentication, least-privilege access, network segmentation between operational technology and enterprise systems, encryption, secure backups, patch management, vulnerability testing and incident-response procedures. Interfaces should use authenticated connections and reject malformed or unauthorised messages.

A practical control framework separates duties. The person who proposes a standard-cost change should not be the sole person who approves and implements it. Emergency access should be time-limited and reviewed after use.

Audit trails

An audit trail records who performed an action, what changed, when it changed, the previous value, the new value, the source and the approval. Immutable or tamper-evident logs are preferable for critical events. Auditability also requires reproducibility: an organisation should be able to reconstruct which data, model version, assumptions and exchange rates produced a reported cost.

For AI, the model registry should record training data period, features, algorithm, parameters, validation results, owner, deployment date, approval status and retirement conditions. Model drift—deterioration caused by changed data or processes—should be monitored.

Ethical controls

Ethical costing involves fairness, transparency, accountability, privacy and proportionality. For example, using worker-level data to estimate labour productivity may create surveillance or discrimination risks. Data collection should have a legitimate purpose, limited scope and appropriate access. Automated supplier scoring may disadvantage smaller suppliers if the model uses variables that reflect scale rather than reliability.

Ethical review should ask: Who benefits? Who bears the risk? Can affected parties challenge an outcome? Is the use consistent with law, policy and professional standards? Human review must be meaningful rather than a ceremonial approval step.

5. Real-time and autonomous costing

Real-time costing updates estimates as events occur rather than waiting for month-end closing. A production order may show current material consumption, labour, machine time, energy, scrap and expected completion cost. This improves responsiveness but increases the importance of timestamps, data latency, provisional classifications and reconciliation.

Real-time cost should be labelled clearly as actual-to-date, estimated-at-completion or forecast. A useful structure is:

The equation assumes that actual-to-date costs are complete and correctly classified and that the remaining-cost forecast is based on the current production plan and valid rates.

Autonomous costing refers to systems that initiate or adjust costing actions with limited human intervention. Examples include automatically recalculating a quotation when material prices change, triggering a review when scrap exceeds a threshold, or updating a simulation when a machine’s condition changes. Full autonomy is rarely appropriate for high-impact accounting decisions. Bounded autonomy is safer: the system operates within approved rules, thresholds and escalation paths.

Important performance measures include cost accuracy, forecast error, close-cycle time, percentage of transactions automatically matched, exception resolution time, data completeness, model drift and control failures. Technology should be assessed through total cost of ownership, including integration, cybersecurity, training, validation and change management.

Illustrative case: Meridian Components

Background

Meridian Components, a fictional mid-sized manufacturer, produces precision housings in three plants. Its monthly product costs were prepared from ERP material issues and broad machine-hour allocations. Managers suspected that setup complexity, energy consumption and rework differed substantially among product families.

Problem

Quotations were often revised after production began. A plant also experienced unexplained increases in conversion cost. Investigation showed inconsistent downtime codes, delayed material postings and a sensor that reported implausible power values.

Analysis

Meridian formed a cross-functional team of cost accountants, production engineers, IT specialists, maintenance personnel and data-security staff. The team created a common data dictionary and linked ERP order numbers with MES operations and machine identifiers. IoT energy data were accepted only when timestamp, calibration and quality status passed validation.

Historical data were analysed to identify candidate drivers. Setup count, machine-hours, product complexity and rework events explained more variation than units produced alone. The team used a transparent regression model for initial analysis and compared its forecasts with a more complex model. The complex model offered only a small accuracy improvement and was harder to explain, so it was not used for approval decisions.

Approach and solution

Meridian introduced activity rates for setup, inspection and rework, while retaining machine-hours for resources with a demonstrable relationship to machine usage. An anomaly system flagged duplicate issues, abnormal energy intensity and extreme scrap observations. Alerts were routed to supervisors rather than automatically changing accounting records.

RPA reconciled approved supplier prices with purchase orders and invoices. Standard-cost changes required approval by the cost accountant and plant controller. A cloud dashboard displayed actual-to-date cost, estimated final cost and variance, with clear labels. Access used role-based permissions and multi-factor authentication. Model versions, data extracts and approvals were retained in an audit repository.

Outcome

After a pilot period, quotation revisions decreased and managers identified a recurring setup problem on one product family. The system did not eliminate variance; instead, it made variance earlier, more traceable and more actionable. Meridian also discovered that data-quality work consumed significant resources, leading it to include data stewardship in operational responsibilities.

Learning points

The case demonstrates that digital costing is an organisational redesign, not merely a software installation. Integration improves visibility, but governance determines whether the resulting information is trustworthy. Transparent models, escalation rules and human accountability are especially important when AI influences financial decisions.

Challenges and implementation guidance

Common implementation failures include attempting to automate poor processes, ignoring master data, treating sensors as inherently accurate, selecting opaque models without validation, and underestimating cybersecurity. Organisations should begin with a defined decision problem, establish a baseline, pilot a limited process, validate results against independent evidence and scale only after controls are effective.

Chapter Summary

Digital cost management integrates ERP financial transactions, MES production events, IoT measurements, digital-twin simulations, RPA workflows and cloud computing. These technologies can improve cost-driver visibility, transaction speed, forecasting and operational control, but they also introduce data-quality, cybersecurity and accountability risks.

AI can discover candidate cost drivers, detect anomalies and forecast future costs. Its outputs must be validated for causality, stability, bias, leakage and model drift. Human oversight should be proportionate to the risk and decision impact. Data governance requires clear definitions, ownership, validation, retention and access controls. Cybersecurity protects confidentiality, integrity and availability, while audit trails make transactions and model outputs reconstructable.

Real-time and autonomous costing can support rapid decisions when actual, estimated and forecast values are clearly distinguished. Bounded autonomy, escalation rules and approval controls are preferable to unrestricted automation. The future cost professional will combine accounting judgement, operational understanding, data literacy, technology governance and ethical responsibility.

Key Terms

Industrial IoT: Connected industrial devices that collect and exchange operational data.

Digital twin: Data-connected digital representation of a physical asset or process.

Cloud costing: Costing applications and data hosted on remotely accessed computing infrastructure.

Activity-based costing: Costing approach that assigns activity costs using relevant drivers.

Anomaly detection: Identification of observations that differ materially from expected behaviour.

Forecasting: Estimation of future values using historical and explanatory information.

Data governance: Policies and responsibilities controlling data quality, access and use.

Master data: Shared reference data such as items, suppliers, routings and work centres.

Cybersecurity: Protection of information systems and data from unauthorised harm.

Human oversight: Deliberate human review and accountability for automated outputs.

Bounded autonomy: Automation constrained by approved rules, thresholds and escalation.

Ethical AI: AI designed and operated with fairness, transparency, accountability and respect for rights.

Review Questions

Short Answer Questions

Long Answer Questions

Analytical / Application Questions

Multiple Choice Questions with Answers

Question 1

Which system primarily records and integrates financial, purchasing, inventory and production transactions?

Question 2

A machine consumes 25 kWh at a rate of 0.16 currency units per kWh. What is the energy cost?

Question 3

Correct Answer: B. A data-connected digital representation of a physical asset or process

Question 4

Why should an AI-identified anomaly not automatically be classified as an error?

Correct Answer: B. An anomaly may represent a valid but unusual operating condition

Question 5

Question 6

Question 7

Correct Answer: B. It may reproduce errors when source data or rules are incorrect

Question 8

Correct Answer: B. Recording the data period, model version, inputs and assumptions

Question 9

Correct Answer: B. The system is constrained by approved thresholds and escalation paths

Question 10

Which is an appropriate ethical question when using worker-level data for costing?

Chapter 15 Digital Cost Management, AI, Governance and Future Practice

Digital Cost Management, AI, Governance and Future Practice

Introduction and context