Agentic AI & ERP Systems

What Is Agentic AI?

Most companies have met AI as something you ask: type a question about last quarter's sales and get a summary, upload an invoice and get the fields extracted. Agentic AI goes a step further. An agent is given a goal and a defined set of permissions. It watches for relevant events, works out what needs doing, carries out the steps through the systems it can access, and hands back to a person when a decision falls outside its authority.

A simple example shows the difference. A conventional AI assistant tells you an item will run out in ten days. An agent notices that on its own, checks the last prices and lead times from three approved suppliers, drafts a purchase order, approves it if it falls under a set limit or routes it for approval if it does not, and then chases the supplier's confirmation.

Because an agent needs trustworthy data and a place to act, the ERP system is its natural home. This article explains how agents work inside an ERP, where they are genuinely useful, what has to be in place first, and how to set boundaries that fit Saudi data protection and AI governance frameworks.

Conventional AI vs. Agentic AI

The last row matters most. Once AI moves from talking to doing, governance, permissions and audit trails become part of the design, not an afterthought.

How an Agent Works Inside an ERP

The architecture can be described in five layers:

• Monitoring: the agent watches for defined events, such as an item hitting its reorder point, an invoice going overdue, or a supplier invoice that does not match the goods receipt.
• Reasoning: when an event fires, the agent gathers context, including movement history, alternative suppliers, company policy and available budget, and compares options using defined rules and estimation models.
• Action: it executes through the ERP's APIs, creating a purchase order, sending a reminder or reserving stock, using only the permissions it has been granted.
• Guardrails: monetary limits, permitted document types, mandatory human approval points and an immediate off switch. Every action is logged with its reason and data source.
• Review: outcomes are compared with expectations and the configuration or rules are adjusted, under human supervision.

The implication is simple. An agent cannot be better than the data it reads or the rules it follows. If inventory balances are wrong, it will order the wrong things faster than any employee could.

Practical Use Cases in ERP

1. Purchasing and replenishment

The clearest and lowest-risk use case is reordering repeat items. The agent compares stock against reorder points, accounts for open purchase orders and reserved quantities, then proposes a quantity and supplier.

Hypothetical setup: a distributor lets its agent raise purchase orders only for items tagged "repeat", only from approved suppliers, and only up to SAR 15,000 per order. Anything larger goes to the purchasing manager with a comparison summary. People still choose the approved suppliers and set the limits; the agent handles the daily routine. The underlying calculation is covered in our guide to the inventory reorder point.

2. Month-end close

Much of the close is definable work: matching bank lines, finding goods received but not yet invoiced so accruals can be booked, comparing expenses with budget, and listing variances that need explanation. An agent can prepare these drafts and raise questions. Posting final entries and approving the statements should stay with the accountable accountant.

3. Collections

The agent tracks due invoices, sends graduated reminders based on age and the customer's payment history, proposes how incoming transfers should be applied to open invoices, and warns the account manager when a customer nears its credit limit. Conversations with key accounts and decisions to stop supplying a customer remain human calls.

4. Workforce scheduling

For shift-based operations, an agent can propose rosters that reflect expected workload and legal limits. Saudi labor rules cap normal hours at 8 a day or 48 a week, reduced during Ramadan for Muslim workers to 6 a day or 36 a week, and overtime is paid at the hourly wage plus 50% of basic wage. An agent that does not know these limits will propose non-compliant rosters, so they must be built into its rules.

5. Customer service connected to the ERP

When a customer asks where an order is, the agent can read the sales order, picking and shipping status and invoice, and answer directly. It can take a return request, check it against policy and create a return for review. Exceptions and complaints go to a person with a full summary.

A Worked Example: One Day of a Purchasing Agent

Take a hypothetical dry-food distributor in Riyadh with 1,200 items, 300 of them tagged as stable repeat items:

1. 6:00 a.m.: after yesterday's transactions post, the agent reads stock levels and finds 14 items at or near their reorder point.
2. Context check: it drops three items that already have open, undelivered purchase orders and one that the category manager has flagged as on hold. Ten remain.
3. Quantity: for each item it calculates a quantity from average daily usage over 90 days, the supplier's recorded lead time and the defined safety stock, rounded to full cartons or pallets.
4. Supplier: it compares approved suppliers on last price, on-time delivery rate and rejection rate at receipt. It never adds a new supplier; that is a human decision.
5. Limits: eight orders fall under SAR 15,000 and are sent. Two exceed the limit and go to the purchasing manager with the comparison and the reason for the choice.
6. Follow-up: if a supplier has not confirmed within 24 hours, the agent sends a reminder, then alerts the responsible buyer if the delay continues.
7. Daily summary: the purchasing team receives a list of orders created, orders pending and items skipped, with reasons.

Every step can be traced and explained, and every strategic decision (approving suppliers, setting limits, classifying items) stayed with people.

A Permissions Matrix for Your Agents

A practical way to bound any agent is a matrix showing what it may do alone, what needs approval and what is off-limits. A hypothetical starting point:

Readiness Assessment

Before discussing any agent, score your organisation on five dimensions. A weak answer on any of them is your real starting point:

The good news is that working on these dimensions pays off even if you never deploy an agent. Clean data, documented processes and clear authority make any system work better.

The Saudi Regulatory Frame

Any agent that touches customer or employee data falls under the Personal Data Protection Law, in force since September 2023 with enforcement from September 2024 after a one-year grace period. Points that bear directly on agents:

• Processing personal data for direct marketing needs consent, and people must have an easy way to opt out at any time.
• Data breaches must be reported to SDAIA within 72 hours of becoming aware of them.
• Transfers outside the Kingdom are allowed under conditions and may require a risk assessment, which matters if your agent calls a model hosted abroad.

SDAIA also published version 2.0 of its AI Ethics Principles in 2023, with seven principles: fairness; privacy and security; humanity; social and environmental benefits; reliability and safety; transparency and explainability; and accountability and responsibility. They are risk-tiered: unacceptable-risk systems are prohibited, high and limited-risk systems must comply, and compliance is recommended for low-risk systems. For data protection inside business systems, see our article on PDPL and ERP data protection.

A Four-Stage Roadmap

1. Foundation: clean master data, standardise processes across branches, and document authority limits and approval routes. This can take months and matters most.
2. Rule-based automation: before AI, put in the simple automation: reorder alerts, three-way matching, automatic reminders. It shows you where processes actually break.
3. One pilot agent: pick a low-risk, measurable use case such as repeat-item reordering. Let the agent only propose for a month or two and compare its suggestions with the team's decisions before granting it execution rights.
4. Governed expansion: widen permissions step by step based on results, add use cases, and review the agent's decision log regularly.

Risks and How to Handle Them

• Confident mistakes: language models can produce wrong conclusions that sound convincing. Use hard monetary limits and human checkpoints, and do not rely on a model for calculations a fixed rule can do.
• Broad permissions: an agent with write access to every module is an attractive target. Apply least privilege; a purchasing agent has no business near payroll.
• No traceability: if you cannot tell why an agent acted, you cannot defend it to an auditor. Log the data it used and the rule it applied.
• Running cost: calling large models continuously costs compute. Use simple rules where they suffice and keep the model for tasks that need language or context.
• Staff anxiety: be clear that the aim is removing routine work, and involve the team in defining rules and reviewing results. Our piece on ERP change management covers this side.

Use Cases by Sector

• Trading and distribution: repeat-item replenishment, supplier confirmation chasing, order status queries, and spotting slow movers before they turn into dead stock.
• Contracting: comparing materials issued with bill-of-quantities budgets per project, warning when a line nears its budget, and drafting progress claims from recorded completed work.
• Manufacturing: comparing actual material usage with the standard bill of materials, proposing schedule changes when a raw material is late, and summarising quality results by supplier and batch.
• Services and property: tracking contract renewals before expiry, reminding tenants of payments, and turning unit complaints into maintenance requests.

The common thread: the agent works on data that already lives in the system. If the BOQs, bills of materials or contracts sit outside it, the agent has nothing to work with.

When You Do Not Need an Agent

Not every improvement needs AI. If the task follows a clear fixed rule, such as "remind five days before due date" or "hold the invoice if quantity differs from receipt", standard ERP automation is cheaper, easier to audit and more predictable. Agents add value when a task requires reading unstructured text like supplier emails, balancing several shifting factors, or handling cases nobody wrote a rule for.

Where Tranquil Fits

Agents need unified data and orderly processes underneath them. Tranquil is a cloud ERP that brings financial accounting, e-invoicing, inventory, sales, purchase management, HR, payroll, projects and asset management into one database, with multi-level approval workflows and API integrations built to your needs. That is the foundation any business needs before deeper automation, whether rule-based or AI-driven.

Frequently Asked Questions

Can an AI agent make large financial decisions on its own?

It should not. Good practice gives agents narrow permissions bounded by amount and document type, and escalates larger decisions to an accountable person with a summary of the reasoning.

How is an agent different from a chatbot?

A chatbot answers within scripted scenarios. An agent understands a goal and carries out several steps across systems, such as spotting low stock, comparing suppliers, raising a purchase order and following it up.

Do we need an in-house AI team?

Not at the start. Clean data, documented processes and an operations team that can define rules and review outcomes matter more. Technical expertise can be brought in for implementation.

Can agentic AI comply with PDPL?

Yes, if it is designed to: a lawful basis for processing, data minimisation, access controls, a log of every action, and checking transfer conditions if the model is hosted outside the Kingdom.

What happens when an agent gets it wrong?

With good design, the impact is limited to its permissions and its actions can be reversed, for example cancelling a purchase order not yet delivered. The decision log shows which data it relied on, so you fix the cause and review its limits before it resumes.

Where should we start?

Assess data quality and process clarity, then automate with rules. Talk to the Tranquil team about getting your processes and data in shape on the system.