Can DeepSeek Help with ERP?

Testing Methodology — How We Did It

We used DeepSeek-R1 via API (not the web-based chatbot) and tested it against 5 simulated ERP scenarios commonly encountered by Thai organizations. Each task was scored on a 1-5 star scale using the following criteria:

Score	Meaning
1/5	Unusable — results are incorrect or irrelevant
2/5	Somewhat usable, but requires extensive corrections
3/5	Usable as a draft, but needs significant improvements
4/5	Very good — minor edits required before use
5/5	Excellent — can be used almost as-is with minimal changes

Testing conditions: All prompts were in Thai, data used was simulated scenarios (not real organizational data), testing was conducted in March 2026 using the latest version of DeepSeek-R1.

Use Case 1: Financial Report Summarization for Executives

The Task

We provided 3 months of Profit & Loss (P&L) data (October-December 2025) for a fictional company, including revenue, cost of goods sold, operating expenses, and net profit. We then asked DeepSeek to summarize the data in Thai for executive presentation, along with recommendations.

Result: 4/5

DeepSeek-R1 analyzed the numbers excellently. It correctly calculated growth rates, gross margins, and net margins, then distilled the findings into 5 key points with sensible recommendations. For example, it identified that production costs increased by 8% in December, potentially due to rising raw material prices, and recommended reviewing the budget for the following quarter.

Strengths: Accurate numerical analysis, strong trend identification, actionable recommendations.

Weaknesses: Thai language output contained some awkward phrasing — for instance, repeatedly using overly formal expressions and sentences that read as though they were translated from English. When the prompt was switched to English, the output quality improved noticeably.

Use Case 2: Writing SQL Queries from Natural Language

The Task

"Find total sales by customer for January 2026, sorted from highest to lowest, showing the number of invoices and average amount per invoice."

Result: 5/5

This is without a doubt DeepSeek's greatest strength. The SQL output was 100% correct, well-structured, readable, and included impressive optimizations:

SELECT
    c.name AS customer_name,
    COUNT(DISTINCT t.t_transaction_id) AS invoice_count,
    SUM(tl.linenetamt) AS total_sales,
    ROUND(
        SUM(tl.linenetamt) / NULLIF(COUNT(DISTINCT t.t_transaction_id), 0),
        2
    ) AS avg_per_invoice
FROM t_transaction t
JOIN t_transaction_line tl ON t.t_transaction_id = tl.t_transaction_id
JOIN t_customer c ON t.t_customer_id = c.t_customer_id
WHERE t.issotrx = 'Y'                -- Sales transactions only
    AND t.docstatus IN ('CO', 'CL')  -- Completed or Closed documents only
    AND t.datetrx BETWEEN '2026-01-01' AND '2026-01-31'
GROUP BY c.name
ORDER BY total_sales DESC;

Strengths: Syntactically correct SQL, uses NULLIF to prevent division by zero, uses DISTINCT to count unique invoices, filters by docstatus to include only completed documents — demonstrating that DeepSeek understands ERP database structures remarkably well. It also recommended creating an index on datetrx for improved query performance.

Note: Coding and SQL are genuinely DeepSeek's strongest capabilities, which is consistent with the fact that the DeepSeek Coder model was specifically designed for this purpose. IT teams can effectively use DeepSeek to write queries for extracting data from ERP systems.

Use Case 3: Stock Mismatch Analysis

The Task

We provided data for 20 inventory items including: item code, item name, ERP system quantity, and physical count quantity. We then asked DeepSeek to identify anomalies, suggest possible causes, and recommend corrective actions.

Result: 4/5

DeepSeek identified every anomalous item and categorized them into 3 severity levels: critical (variance greater than 10%), moderate (5-10%), and minor (less than 5%). It then proposed 5 possible causes:

1. FIFO mismatch — The system deducts stock in FIFO order, but warehouse staff pick items out of sequence, causing FIFO discrepancies between the system and actual inventory
2. Unreported withdrawals — Staff withdraw items without recording them in the system, resulting in system quantities being higher than physical stock
3. Unrecorded goods receipts — Items arrive at the warehouse but the goods receipt has not been entered in the system, causing physical stock to exceed system records
4. Damaged or expired goods — Items are damaged but have not been written off in the system
5. Counting errors — Items miscounted due to wrong units of measurement or skipped items during physical count

DeepSeek also recommended implementing weekly cycle counts for critically mismatched items and reviewing the transaction log for the past 30 days to pinpoint when the discrepancies first appeared.

Strengths: Identified all anomalies, categorized by severity, proposed comprehensive root causes.

Weaknesses: The root causes were "generic" rather than organization-specific. For instance, it did not know whether the organization uses barcode scanning or bin location management — information that would enable much more precise root cause analysis.

Use Case 4: Drafting a Purchase Order Approval Workflow

The Task

Design a 3-tier purchase order (PO) approval workflow for a mid-sized organization: Department Head, Division Manager, and Managing Director. Define spending limits, conditions, and SLAs for each approval level.

Result: 3/5

DeepSeek produced a workflow structure that looked good at the surface level:

Level	Approver	Amount (THB)	SLA
Level 1	Department Head	Up to 50,000	1 business day
Level 2	Division Manager	50,001 - 500,000	2 business days
Level 3	Managing Director	Over 500,000	3 business days

Strengths: Solid basic structure, reasonable spending limits, clear SLAs for each level.

Weaknesses: Missing several critical edge cases, including:

• No delegation rules: If the Department Head resigns or is on sick leave, who approves in their place?
• No rejection handling: If the Division Manager rejects the PO, which step does it return to? How is the resubmission process handled?
• No timeout/escalation: If an approver does not respond within the SLA, does the system automatically escalate to the next level?
• No emergency fast-track: Is there an expedited approval path for urgent procurement needs?
• No split PO detection: If a PO exceeding 500,000 THB is split into two POs of 300,000 THB each, will the system flag this?

For real-world approval workflow design, review by an ERP specialist who understands the organization's processes is essential. Every organization has different business rules. DeepSeek can provide the "skeleton," but the "substance" must be filled in by domain experts.

Use Case 5: Answering Employee ERP Questions

The Task

"How do I create a Purchase Order in the ERP system? What are the steps?"

Result: 3/5

DeepSeek provided an answer that was broad and correct in principle, offering 7 general steps: Log in, navigate to the procurement menu, create a new PO, select the vendor, choose items, specify quantities and prices, then save and submit for approval.

Strengths: Correct overall logic, complete step sequence, each step clearly explained.

Weaknesses — and this is the critical issue:

• The answer describes generic steps, not specific to any particular ERP system
• It does not know the user interface (UI) of Saeree ERP, SAP, Oracle, or any other system — it cannot tell the user "click this menu" or "fill in this field"
• It does not know organization-specific business rules, such as whether vendors must be pre-approved, or what documents need to be attached to a PO
• If an employee asks a question like this and receives a generic answer, it is no different from searching Google

Summary of DeepSeek ERP Test Results

Here are the consolidated results from all 5 use case tests:

Use Case	Score	Strengths	Weaknesses
Financial Report Summary	4/5	Excellent numerical analysis, strong trend detection	Thai language output not fully natural
SQL Query Writing	5/5	100% correct, well-optimized, includes index hints	-
Stock Mismatch Analysis	4/5	Identified all anomalies, comprehensive root causes	Lacks organization-specific context
Approval Workflow Draft	3/5	Good basic structure, reasonable spending limits	Missing several critical edge cases
Employee ERP Q&A	3/5	Broad coverage, correct principles	Does not know any specific ERP system; too generic

Average score: 3.8/5 — This qualifies as "good" for a low-cost AI, but it is not yet at the level where it can replace an ERP specialist.

DeepSeek vs ChatGPT vs Claude — For ERP Work Specifically

We ran the same test scenarios through ChatGPT (GPT-5.4) and Claude (Opus 4.6) for comparison. Here are the results:

Capability	DeepSeek R1	ChatGPT GPT-5.4	Claude Opus 4.6
SQL / Coding	5/5	5/5	5/5
Thai Language Analysis	3/5	5/5	4/5
Reasoning / Deep Analysis	4/5	5/5	5/5
Price	5/5 (cheapest)	2/5	2/5
Data Privacy	2/5 (China)	4/5	4/5
Self-hosting	5/5	-	-

Bottom line: If cost is your top priority or you need self-hosting to maintain full data control, DeepSeek is a strong option. However, if you need excellent Thai language output and enterprise-grade data privacy, ChatGPT or Claude still have the edge. For more details, see our article on The Missing Report — The Gap That AI Can Fill.

Conclusion: How Should DeepSeek Be Used for ERP?

Based on all 5 use case tests, the conclusion is clear:

DeepSeek Is Best Suited as an "Assistant," Not a "Decision-Maker"

No AI — whether DeepSeek, ChatGPT, or Claude — can replace an ERP specialist. However, AI can serve as a highly effective "assistant" that dramatically speeds up work, particularly in these areas:

• SQL and data extraction: DeepSeek writes excellent SQL queries, saving IT teams significant time
• Numerical analysis: Report summarization, trend detection, and anomaly identification are all strong — ideal for creating "first drafts" before executive presentations
• Workflow and process drafting: It can provide solid initial structures, but business rules must be added by domain experts

DeepSeek's Strengths for ERP

• SQL / Coding — On par with GPT-5.4 and Claude Opus
• Numerical data analysis — Accurate calculations and strong trend detection
• Low cost — 10-50x cheaper, making it accessible for SMEs
• Self-hosting capable — Data stays within your organization (but requires sufficient GPU resources)

DeepSeek's Weaknesses for ERP

• Thai language — Cannot match GPT or Claude; awkward phrasing and sentences that read like translations from English
• No knowledge of specific ERP systems — Can only answer with general principles, not specific UI navigation or business rules
• Data privacy — API usage sends data to China (mitigated by self-hosting, but requires infrastructure investment)
• Lacks contextual understanding — Provides "generic" answers rather than organization-specific insights

For organizations that need AI in their ERP operations at an enterprise level, what they truly require is AI fine-tuned on their own ERP data + secure data handling with no external transmission + understanding of organization-specific business rules. DeepSeek (and general-purpose AI in general) cannot do this out of the box — it requires a dedicated development team or specialist to customize.

DeepSeek Series — All 5 Episodes Complete

The best AI for ERP is not the most powerful AI — it is the AI that best understands your organization's context.

— Saeree ERP Team

References

• DeepSeek API Documentation
• DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning — Technical Report