What is Agentic ERP
Agentic ERP is an ERP system where AI agents take on back-office work, such as adjusting stock, receiving goods, balancing inventory across channels, summarizing tax, analyzing receivables, and preparing expense, invoice, and PP30 drafts, within approval levels that humans set. This page defines the term, explains the five BAL levels, shows real Nong Pao examples, and publishes the live status of every job.
Seller Pao is the agentic ERP for online-selling SMEs in Thailand: a back office you can ask Nong Pao anything, where AI prepares accounting and tax work in full and you confirm. Free to start.
Seller Pao defines this term and publishes the criteria openly. First definition: September 2, 2026. If you use this definition, please cite this URL so everyone means the same thing.
AI prepares the work; humans review and decide. This principle governs every job in the system.
How is Agentic ERP different from a regular ERP and an OMS
The difference is not whether AI exists, but how much the AI can act, and how much a human still clicks.
- A traditional ERP stores data from many departments, then waits for a person to search, compare, and record every step.
- An ERP with a chatbot or copilot: AI answers questions and summarizes data, but comparing numbers and recording stays fully manual.
- An agentic ERP: AI takes repetitive jobs and prepares them end to end within assigned boundaries, such as calling several tools to answer one question or preparing document drafts for confirmation. The saving stays with the human.
Ask every vendor: does the AI suggest or does it act, and where does a human approve? The phrase has AI does not tell you anything.
What are the Back-office Autonomy Levels (BAL 0 to 4)?
BAL is a five-level framework that says how much back-office work AI can do on its own, from level 0 that only reports data to level 4 that takes a goal and finishes within a narrow scope. Every level states what the human does. It is adapted from McKinsey (L0 to L3 in finance) and Gartner (Observe, Act with Approval, Act Autonomously), adjusted for Thai online sellers' back offices.
| Level | Name | What AI does | What the human does |
|---|---|---|---|
| BAL 0 | Report | Summarize data, answer questions, no actions proposed | Judge and do every step yourself |
| BAL 1 | Suggest | Propose actions with reasons | Read, decide, and click yourself |
| BAL 2 | Act on approval | Prepare the work fully, waits for per-item human confirmation | Review and confirm each item |
| BAL 3 | Act within written rules | Repeat tasks automatically inside rules set by humans, escalate exceptions | Write the rules and review escalated cases routinely |
| BAL 4 | Full automation in a narrow scope | Take a goal, plan, finish, report | Set goals and audit results afterwards |
The principle that anchors this framework: AI prepares the work, humans review and decide. No level is designed to run without human review. Tax and filing work is locked at BAL 2 or below by policy, and a feature's published level must match its real behavior: if a person must confirm every item, that is BAL 2, not BAL 3.
What Nong Pao's real work looks like
Nong Pao is the in-app AI assistant in the bottom right of every screen. Every example below is real, showing what AI does, where the human sits, and the BAL level.
- 1Open the expenses page and ask where to start. Nong Pao explains the sections, main buttons, and first step from the real page structure, without guessing (BAL 0).
- 2Type: prepare an expense for 3,500 baht of Facebook ads today. Nong Pao finds the vendor even with different spelling, fills the category and date, and shows a preview. A human reviews and confirms before saving (BAL 2).
- 3Ask what drove this month's expenses. Nong Pao reads totals, splits them by account and vendor, compares with last month, and merges several tools into one answer (BAL 0).
- 4Ask who is overdue beyond 90 days and get an aging table per vendor from due dates (BAL 0).
- 5Type: prepare an invoice for ABC company, service fee 80,000. Nong Pao prepares the line items and total; if ABC is missing, it proposes creating the customer for approval first (BAL 2).
- 6At month end, type: prepare the latest PP30. Nong Pao checks the period is closed, reconciled, and has no existing filing, then prepares a draft with points to review first. A human reviews and files with the Revenue Department, always (BAL 2).
The point: no step saves anything by itself. That is design, not a limitation. The more a job affects money or tax, the tighter the human confirmation.
Where Seller Pao stands today
This table reports the real status, updated with every release.
| Job Nong Pao does | BAL | Status | What the human does |
|---|---|---|---|
| Explain the current page and navigate to the right one | 0 to 1 | Live | Read and click buttons |
| Summarize revenue, expenses, profit, compare periods | 0 | Live | Read and judge |
| Summarize output VAT, input VAT, net VAT, and WHT | 0 | Live | Read and judge |
| Receivables, payables, outstanding balances, and aging | 0 | Live | Read and judge |
| Rank customers and products, split expenses by account | 0 | Live | Read and judge |
| Search contacts, expenses, products, transactions by meaning | 0 | Live | Read |
| Prepare expense drafts from one sentence | 2 | Live | Review preview, then confirm |
| Prepare sales invoice drafts | 2 | Live | Review, then confirm; new customers need approval to create |
| Prepare general journal entries (full accounting) | 2 | Live | Check balance, then post |
| Prepare PP30 drafts after a period readiness check | 2 | Live | Review and create; filing is always human |
| Prepare income drafts from messages or receipt photos, LINE only | 2 | Live | Review, then confirm |
The highest level of back-office work today is BAL 2. Every posting passes through a human, by design. WebMCP is live for external AI agents (developer tooling), see /webmcp. Stock-side and order-side work via Nong Pao, such as adjusting stock, receiving goods, balancing inventory across channels, and COD reconciliation, is on the roadmap, see /features/erp
What are the risks of Agentic ERP and how are they controlled
A faster system fails faster too, so the controls are built in: Nong Pao reads only data inside your account; when context is unclear it says it does not know; when semantic search is unavailable it states the limitation instead of inventing an answer; every number is computed server side; every draft shows a full preview before confirmation; journal entries must balance before being offered; documents have multiple statuses you can look back on; tax work is locked at BAL 2 with no exceptions. Security and PDPA details at /trust
Who Agentic ERP is not for
- Shops with a few orders a day and one person: basic accounting software or a simple OMS still finishes the job; this system costs more than needed to set up and maintain. Overview at /features/erp
- Businesses whose processes have no written rules: agents run on rules written by humans; if the way you work is not agreed on paper yet, the system follows vague rules and gets faster in the wrong direction.
- Organizations that want to switch on and walk away: we do not design systems without human review. If you want a system that does everything including filing taxes, we are not the right fit.
Frequently asked questions about Agentic ERP
How is Agentic ERP different from a regular ERP and an OMS?
An OMS manages orders from every channel in one place. An ERP consolidates all back-office data and waits for a human to command every step. Agentic ERP continues from there: AI agents take repetitive jobs and prepare them end to end within human-set boundaries, such as summarizing the period's tax, analyzing who is overdue beyond 90 days, or preparing expense and PP30 drafts, then hand them back for human review and confirmation.
How is agentic AI different from RPA?
RPA follows scripts written in advance; when a screen or file changes slightly, the script breaks. AI agents read the situation, decide within rules, and refuse to guess when data is missing. For identical daily screens RPA is still faster and cheaper; for non-deterministic input like Thai chat messages, agents win. They do not replace each other.
How does AI-native ERP differ from bolting external AI onto an existing system?
Built-in AI sees data directly under the system's own permissions, works through the same paths users do, and is auditable from one place. External AI via connectors or MCP is easy to start and keeps your current system, but data scope and call logging live with the middleman. If you want agents to finish accounting work, a system with AI inside is more deeply controllable.
Can AI do tax and filing by itself?
Not at Seller Pao, and that is intentional. Tax work is locked at BAL 2: AI prepares and proposes, a human reviews and clicks every time. Filing with the Thai Revenue Department is always a human act, because tax mistakes carry legal penalties that land on the business owner.
If AI prepares the work, who is responsible when it is wrong?
In Seller Pao today, AI saves nothing by itself. Every draft needs a human confirmation, so there is always a human approver, and documents have multiple statuses you can look back on. Final responsibility sits with people and the company, not the model.
Is Seller Pao's Agentic ERP available now?
The list of jobs Nong Pao does and their per-row status is in the table above. /features/erp summarizes features and the roadmap. Free to start; packages at /pricing
Sources
- Infor, Agentic ERP product page, accessed Sep 2, 2026, https://www.infor.com/solutions/erp/agentic-erp
- Microsoft Dynamics 365 blog, Reinventing source to pay with agentic ERP, June 18, 2026, https://www.microsoft.com/en-us/dynamics-365/blog/business-leader/2026/06/18/reinventing-source-to-pay-with-agentic-erp/
- ASUG, Agentic ERP: The Future of the Enterprise, accessed Sep 2, 2026, https://www.asug.com/insights/agentic-erp-the-future-of-the-enterprise
- NIST, AI Risk Management Framework, accessed Sep 2, 2026, https://www.nist.gov/itl/ai-risk-management-framework
- EU AI Act, Regulation (EU) 2024/1689, Article 14 human oversight, https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX:32024R1689
- Wikipedia, AI agent, accessed Sep 2, 2026, https://en.wikipedia.org/wiki/AI_agent
Last updated: September 2, 2026