A practical guide to using AI in finance, HR, manufacturing, and customer operations without creating a compliance black box.
“The question is no longer whether your business will use AI. The question is whether it will use AI with enough control to trust the result.”
Artificial intelligence is now part of everyday business software.
Companies use AI to read supplier invoices, forecast demand, write customer emails, summarise meetings, suggest purchasing actions, detect unusual payments, support recruitment, and help managers understand large amounts of operational data.
The potential is real. AI can reduce repetitive work, speed up decisions, and help teams find patterns they might otherwise miss.
But AI also creates new responsibilities.
If an AI system suggests a price change, recommends a supplier, ranks job candidates, predicts employee performance, or prepares a financial posting, the business must be able to answer important questions:
- What data did the system use?
- Who approved the action?
- Was a human able to review the recommendation?
- Could the AI create unfair, unsafe, or incorrect outcomes?
- Can the company explain what happened later?
- Is sensitive data protected?
- Does the user have permission to access the information shown?
Good AI does not remove control. It makes control faster, clearer, and more useful.
The AI Act Is Now Part of the Business Landscape
The EU AI Act entered into force on 1 August 2024. Its rules are being introduced in stages.
Since 2 February 2025, rules on prohibited AI practices and AI literacy obligations have applied. From 2 August 2026, the AI Act is broadly applicable, and the EU AI Office and national authorities have enforcement responsibilities. Some requirements for high-risk AI systems apply later. European Commission: AI Act framework
The current timeline includes:
| Date | What it means |
|---|---|
| 2 February 2025 | Prohibited AI practices and AI-literacy obligations began to apply. |
| 2 August 2025 | Governance rules and obligations for general-purpose AI models became applicable. |
| 2 August 2026 | The AI Act became broadly applicable, subject to specific later deadlines. |
| 2 December 2027 | Rules for certain high-risk AI systems in Annex III are scheduled to apply. |
| 2 August 2028 | Rules for high-risk AI systems embedded in regulated products are scheduled to apply. |
The AI Act does not mean that every company using AI must build a legal department.
It means companies need to understand what type of AI they use, what decisions it influences, what risks it creates, and what controls are appropriate.
Not Every AI Feature Has the Same Risk
An AI assistant that drafts a customer follow-up email is not the same as an AI system that ranks job applicants.
An AI tool that summarises a meeting is not the same as a system that determines whether an employee should receive a promotion, whether a customer should receive credit, or whether a worker should be monitored.
This distinction matters.
In an ERP environment, AI use cases can broadly be grouped into three practical levels.
Low-Risk Assistance
These are tools that help people work faster while leaving the decision fully with the user.
Examples include:
- Drafting customer emails.
- Summarising long notes or documents.
- Translating internal content.
- Suggesting descriptions for products.
- Explaining a dashboard.
- Helping users find an ERP function.
- Preparing a first draft of a proposal.
- Categorising documents for review.
These tools are still important, but the risk is usually lower because the user remains responsible for reviewing the output.
Operational Recommendations
These AI features analyse business data and suggest an action.
Examples include:
- Reorder recommendations.
- Demand forecasts.
- Cash-flow forecasts.
- Supplier-delay alerts.
- Duplicate-invoice detection.
- Suggested payment matching.
- Production scheduling suggestions.
- Anomaly detection in stock movement.
- Identification of late-paying customers.
- Suggested next-best sales actions.
These systems can create substantial value. But companies should ensure that the recommendation is understandable, measurable, and reviewable.
The AI may recommend. The business should remain responsible for deciding.
High-Impact or Sensitive Decisions
Some AI use cases deserve much stronger controls because they can affect people’s rights, safety, finances, or access to important services.
Examples may include AI systems used in:
- Recruitment and candidate screening.
- Employee evaluation or performance scoring.
- Promotion, dismissal, or work allocation decisions.
- Biometric identification.
- Creditworthiness or financial-risk decisions.
- Safety-critical manufacturing or product systems.
- Systems connected to regulated products or critical infrastructure.
These cases may fall into higher-risk categories depending on the exact purpose, design, and legal classification. Companies should seek specialist legal advice before deploying AI in sensitive areas.
The closer AI moves to decisions about people, money, safety, or legal rights, the stronger the controls must become.
AI Literacy Is Not Optional
One of the most practical obligations under the AI Act is AI literacy.
AI literacy does not mean every employee must learn machine learning, programming, or advanced mathematics.
It means people who use AI at work should understand enough to use it responsibly.
For example, employees should know:
- AI can make mistakes.
- AI outputs must be reviewed before use.
- Sensitive information should not be copied into unapproved tools.
- AI can reflect weak, incomplete, or biased source data.
- A confident answer is not automatically a correct answer.
- Important decisions need defined human accountability.
- Users should know when they are interacting with AI.
- The company’s internal AI rules must be followed.
A finance employee using AI to extract invoice data should understand how to check VAT, totals, supplier identity, and duplicate risk.
A sales employee using AI to draft a proposal should understand that pricing, legal terms, and delivery promises must be approved.
A manager using AI analytics should understand that a forecast is a decision-support tool—not a guaranteed future.
AI literacy protects both the company and the employee.
The Biggest ERP Risk: AI Acting on Bad Data
AI is only as useful as the data it receives.
If product data is incomplete, supplier lead times are outdated, VAT codes are inconsistent, inventory records are wrong, or customer information is duplicated, AI can create very confident but very poor recommendations.
For example:
- A demand forecast may be distorted by incorrectly posted sales.
- A purchasing suggestion may create excess stock if customer orders were duplicated.
- A cash-flow forecast may be unreliable if payment terms are inconsistent.
- A production recommendation may ignore a machine constraint that was never recorded.
- A customer summary may expose incorrect or outdated personal data.
Before using advanced AI, companies should improve the business data already inside their ERP.
This is why AI and ERP belong together. ERP provides the structured operational data. AI helps interpret it, identify patterns, and prepare decisions. But the ERP must remain the controlled source of truth.
Human-in-the-Loop: The Most Important Principle
Human-in-the-loop means that a person remains involved in important decisions.
This does not mean every AI suggestion requires a committee meeting.
It means the level of approval should match the level of risk.
For example:
- An AI-generated email draft can be reviewed and sent by the user.
- A supplier invoice can be extracted automatically but require finance approval before posting.
- A purchase-order draft can be prepared automatically but require purchasing approval before sending.
- A production rescheduling proposal can be reviewed by the planner before changing work orders.
- A price change can require commercial approval.
- A hiring recommendation should never become an automatic employment decision.
The goal is to remove repetitive manual work while keeping business accountability where it belongs.
AI should prepare better decisions. People should own the decisions that matter.
Explainability: Can Your Team Understand the Recommendation?
A business should not accept a recommendation simply because AI produced it.
If the system suggests increasing a purchase order, the buyer should be able to see why.
For example:
“Recommended increase: 1,200 units. Reason: open sales orders increased by 18%, forecast demand is above plan, supplier lead time increased from 12 to 18 days, and available stock will fall below the safety threshold in nine days.”
This is useful because the recommendation includes:
- The proposed action.
- The data considered.
- The expected risk.
- The expected benefit.
- The confidence level.
- The responsible approver.
The same principle applies across ERP processes.
An AI system should be able to explain why it flagged an invoice as suspicious, why it predicted a stock shortage, why it suggested a production change, or why it prioritised a customer follow-up.
If a recommendation cannot be explained clearly, it should not be allowed to make an important decision on its own.
AI in Finance: Fast Processing With Strong Controls
Finance is one of the best places to start with practical AI because many tasks are repetitive but still require review.
AI can support:
- Invoice reading and field extraction.
- Supplier identification.
- VAT and currency checks.
- Duplicate-invoice detection.
- Purchase-order and goods-receipt matching.
- Payment matching.
- Expense categorisation.
- Cash-flow forecasting.
- Late-payment risk analysis.
- Unusual transaction alerts.
- Management reporting summaries.
The important controls include:
- Defined approval thresholds.
- Clear separation of duties.
- Audit trails.
- Validation of tax and accounting rules.
- Role-based access.
- Exception queues.
- Human review for unusual or high-value items.
- Secure handling of bank, payment, employee, and customer information.
The objective is not touchless finance at any cost. It is faster finance with stronger evidence.
AI in Manufacturing: Recommendations Before Problems Escalate
Manufacturing is another high-value area for AI, especially when the ERP is connected to production, warehouse, quality, maintenance, and purchasing data.
AI can help identify:
- Demand changes.
- Material shortages.
- Supplier lead-time risk.
- Capacity bottlenecks.
- Increased scrap rates.
- Delayed work orders.
- Machine-performance changes.
- Unusual material consumption.
- Quality drift.
- Transport and delivery risk.
A useful manufacturing AI system should not simply say, “There is a problem.”
It should help the planner understand the options.
For example:
“Production order MO-1048 is at risk because the supplier ETA for a critical component moved by six days. Option one: use approved substitute material A. Option two: move the order to work centre B. Option three: prioritise customer order SO-2917 and delay lower-priority order SO-2904. Expected effect on cost, capacity, and delivery dates is shown below.”
This is where AI becomes practical.
It does not replace the planner. It gives the planner better options before the production line starts firefighting.
AI in HR: Use Extra Caution
HR is one of the areas where businesses must be especially careful.
AI can help with administrative HR tasks, such as drafting job descriptions, summarising training feedback, answering policy questions, organising documents, or helping employees find information.
But AI should be treated with far more caution when it influences:
- Recruitment.
- Candidate ranking.
- Employee monitoring.
- Performance scoring.
- Promotion.
- Compensation.
- Dismissal.
- Shift allocation.
- Workplace access.
These decisions affect people directly. They can create discrimination, privacy, employment-law, and fairness risks.
A responsible HR AI approach should include:
- Clear human ownership.
- No automatic adverse employment decisions.
- Transparency for affected people where required.
- Bias testing and regular review.
- Data-minimisation rules.
- Defined retention periods.
- Documented purpose and lawful basis for personal-data processing.
- Strong access control.
AI can support HR. It should not become an unaccountable judge of employees.
Agentic AI: Powerful, but Only With Boundaries
Agentic AI goes beyond answering questions. It can plan steps, use approved tools, gather data, compare options, and prepare actions.
In an ERP system, an agent could:
- Check sales demand.
- Review material availability.
- Identify supplier delays.
- Review production capacity.
- Simulate options.
- Create a draft purchase order.
- Prepare a revised work-order schedule.
- Send an approval request to the responsible manager.
This can save enormous time.
But the agent should operate through defined boundaries.
For example:
- It can read stock data but not export customer data without permission.
- It can create a purchase-order draft but not send it automatically.
- It can suggest a price adjustment but not change the price list without approval.
- It can recommend a production change but not override safety rules.
- It can explain an issue but not delete financial or HR records.
Agentic AI should be controlled autonomy—not uncontrolled automation.
A Practical AI Governance Framework
A strong AI governance framework does not need to be bureaucratic. It needs to be clear.
Every company using AI in ERP should define the following.
Approved Use Cases
Document where AI may be used and what problem it is intended to solve.
For example:
- Supplier invoice extraction.
- Proposal drafting.
- Demand forecasting.
- Production-risk alerts.
- Customer-service summaries.
- ERP search and user guidance.
Data Rules
Define what data AI may access, store, process, or share.
Special care is needed for:
- Personal data.
- Financial data.
- Bank details.
- Payroll information.
- Health information.
- Employee records.
- Customer contracts.
- Trade secrets.
- Technical drawings.
- Security information.
Access Rules
AI must follow the same access rights as the user.
A sales representative should not use AI to access payroll records. A warehouse worker should not receive confidential financial information. A consultant should not see data outside their authorised customer scope.
Approval Rules
Define which AI actions require human approval.
For example:
- Posting invoices.
- Changing prices.
- Creating supplier orders.
- Editing work orders.
- Sending customer communications.
- Changing payroll or HR data.
- Approving credit limits.
- Releasing regulated product information.
Audit Evidence
Keep a clear record of:
- The AI use case.
- Data accessed.
- Recommendation produced.
- User who reviewed it.
- Approval or rejection.
- Final action taken.
- Exception or correction.
- Model or prompt version where relevant.
Ongoing Monitoring
AI performance should be reviewed regularly.
Ask:
- Is the recommendation accurate?
- Are users correcting it often?
- Is it creating bias or unnecessary exceptions?
- Has the source data changed?
- Are there new legal or customer requirements?
- Does the use case still create business value?
How SIX ERP Supports Responsible AI
SIX ERP is designed to bring AI into the operational environment where people already work—without separating intelligence from business control.
A responsible AI approach inside SIX ERP can include:
- Role-based access to data and AI functions.
- Embedded AI assistance in finance, CRM, purchasin_g, manufacturing, and warehouse workflows.
- Approval processes before sensitive actions are applied.
- Audit trails for documents, decisions, and changes.
- Context-aware AI guidance based on the current record.
- Controlled document AI for invoices, receipts, and expenses.
- Predictive alerts for stock, cash flow, supplier risk, and production performance.
- Human review for exceptions and high-impact decisions.
- Clear separation between recommendations and final approvals.
The goal is straightforward:
Use AI to remove unnecessary work—not necessary responsibility.
A Sensible First Step
Do not begin with the most complex or sensitive AI use case.
Start with one process where the value is obvious and the risk is manageable.
Good first projects include:
- Supplier invoice extraction and validation.
- Expense processing.
- AI-assisted proposal drafting.
- Customer-service summaries.
- ERP search and user guidance.
- Demand and inventory alerts.
- Production-delay detection.
- Duplicate-invoice and anomaly detection.
Define the baseline before you start. Measure processing time, error rate, exception rate, user satisfaction, and financial impact.
Then improve the process step by step.
The Future Is Controlled Intelligence
The best AI systems will not be the ones that make the most decisions without people.
They will be the ones that help people make better decisions with less delay, less manual work, and stronger evidence.
AI can make ERP more helpful, more predictive, and easier to use. It can connect information across finance, sales, purchasing, warehouse, production, and service processes.
But it must remain understandable, controlled, secure, and accountable.
The strongest companies will not ask, “How much AI can we add?” They will ask, “Where can AI create real value while protecting people, data, and trust?”
SIX ERP helps businesses build this kind of AI foundation: connected data, clear workflows, role-based access, approval controls, and intelligence embedded where work happens.
