If you plan to introduce AI anomaly detection for purchase order unit prices in manufacturing, inspect the actual PO line before approval. A supplier quotation may have been compared correctly while the PO still carries a different pack size, price unit, quantity break, currency, freight treatment, tax treatment, or contract version. Comparing raw numbers with an average historical price creates false alarms and misses errors. This guide gives purchasing, finance, and IT teams at Thai factories requirements they can use in an RFP, pilot, FAT, and SAT.
The decision: AI assists review; people retain purchase authority
A useful alert explains which comparable price was used, how units and currency were converted, the applicable agreement, the price difference, and the business impact. Hard controls such as a contractual ceiling and delegation of authority belong in deterministic rules. AI can prioritize less obvious changes, but a low anomaly score must never itself approve a PO. The buyer and authorized approver decide after seeing evidence.
This is a different control point from AI comparison of supplier quotations, which supports supplier selection, and the purchase requisition approval workflow, which validates an internal need. Even when a quote and requisition were approved, the issued PO line may be wrong. Keep a distinct check at PO creation and amendment.

Define the anomaly at PO line level
The unit of review is a PO number, revision, line, delivery location and pricing date. Keep both percentage variance and total financial exposure. A large order may have a small unit variance but material exposure; a small order may have a severe unit error. Separate at least four outcomes: definite rule violation, price needing human review, informational change, and insufficient data. Never show missing contract or conversion data as a clean pass.
Choose a baseline hierarchy. An effective agreement for the same legal supplier, item, specification revision and receiving site is usually stronger evidence than an approved price list, which is stronger than a comparable old PO. Historical POs are not contracts. Mark canceled lines, corrections, emergency buys and one-off charges so that the model does not learn earlier mistakes as normal. Cross-supplier comparisons need checks for qualification, lead time, MOQ, terms and specification.
Normalize units, packs and quantities before comparing
SAP’s purchasing documentation distinguishes order unit from order price unit. A buyer might order by box and pay by kilogram. Store order quantity unit, price unit, inventory unit, condition unit, conversion factor, its source and effective date. A changed pack size invalidates a stale conversion. For variable-weight materials, distinguish planned from actual conversion.
For example, a PO of THB 2,000 per box of 20 pieces equals THB 100 per piece. A contract at THB 100 per piece is not a 20-fold price difference. Check quantity and extended value too: a mistaken box quantity can be invisible if the unit price is correct. Quantity discounts need the applicable tier and whether it is based on this release or cumulative purchases. Oracle explicitly distinguishes cumulative and noncumulative price breaks. Do not flag a legitimate discount simply because it differs from the previous PO.
Currency needs equal care. Preserve quotation, contract, PO and reporting currency separately, as well as the rate date, rate source, fixed-rate flag and rounding rule. SAP explains that a foreign-currency PO can use a fixed exchange rate. The Bank of Thailand publishes daily rates, but these are not automatically the accounting rate for every company’s PO. Follow the company’s approved policy. Compare net unit price and landed cost separately; do not silently mix tax, freight, insurance, duty or tooling amortization. The same charge may appear in a line condition, header or separate service line.
Respect agreement validity and specification
A baseline lookup should include agreement number and line, legal supplier, plant, item and drawing revision, start and end dates, currency, quantity tier and cumulative period. SAP documents time-dependent price conditions, and Oracle documents price-break start and end dates. These are software capabilities, not proof that a factory’s masters are current. Decide which record wins if an attached contract and ERP price master conflict, and who fixes the discrepancy.
Save the versions used at approval: agreement, price list, conversion master, FX rule and detection logic. If a later price update changes the reference value, the old decision must remain reconstructable. Re-evaluate an amended PO as a new revision and show the difference. MOQ and order-multiple control is also relevant: reaching a lower unit price by buying unnecessary stock is not automatically a good decision.

Build rules first, then measure what AI adds
Minimum inputs include PO item, supplier, site, quantity, units, pack size, currency, unit price, condition costs, delivery date, agreements, price lists, approved exceptions and PO revision history. Link requisition and quotation IDs where possible. Do not let a model silently invent missing attributes. Route missing conversions or agreements to a named data owner.
Rules can handle agreement ceiling, effective dates, unit mismatch, tier boundaries, delegated amount, and changes after approval. Oracle’s procurement documentation includes agreement price limits and price-change tolerances. Use values from your own policy, not vendor examples. AI may then highlight sudden shifts for the same item, unusual relationships within a genuinely comparable specification group, or changes unusual for a supplier. Look for suspiciously low as well as high prices. Show the number and age of comparable POs and variation in the group. For a new item with little history, prefer agreement, quotation and technical review over an overconfident score.
True erroneous POs are often scarce labels. Start by comparing a transparent statistical baseline plus rules against more complex models. Split training and evaluation by time to prevent future contract changes or corrections from leaking into past decisions. A language model can help extract clauses from a PDF or summarize evidence if it cites the exact source, but it must not create a final price or approval decision from prose alone.
Display four states: rule-based stop with a documented exception path, reviewer attention, information, and cannot assess. A score alone should not stop production-related purchasing. The alert should show current price, normalized reference, variance amount, unit and FX calculations, source agreement, and a short cause hypothesis. Route a typing error to the creator, negotiation to purchasing, and policy exception to the authorized approver. Record what was changed and how the decision changed.
Measure false alarms and missed cases
The following figures are hypothetical requirements examples, not reported TOMAS TECH outcomes. Suppose 10,000 PO lines in a month include 100 genuinely problematic lines. A system flags 300; reviewers confirm 80 true problems. Precision is 80/300, about 26.7%, and recall is 80/100, 80%. Reviewers spent time on 220 false alarms, while 20 problems were missed. At three minutes per review, 300 alerts require 15 hours. A stricter threshold might flag 150 with 70 true problems: precision about 46.7% and recall 70%. The preferred balance depends on exposure, item criticality and available review capacity.
Do not call the sum of flagged differences savings. Price rises may be justified. Track separately the overpayment prevented by actual correction, renegotiated amounts, verified avoided losses, review time and data maintenance cost. Evaluate by category and high-risk minority cases rather than one overall average.

What to write into an RFP
Specify the decision point, PO revision logic, sources of truth, unit conversions, contract validity, tier pricing, exception authority, audit evidence, API behavior and evaluation metrics. Vendors should price the same plant, languages, PO volumes and integration scope. A useful requirement table is:
| Requirement | Example question | Acceptance evidence |
|---|---|---|
| Decision unit | PO revision, line, site, pricing date? | Different conditions receive separate results |
| Normalization | Unit, pack, FX, freight and tax? | Calculation and master version reproduced |
| Agreement | Effective period, tier, ceiling? | Expired agreement not used as baseline |
| AI output | Score, peer count, period, reason? | Reviewer can inspect evidence |
| Exception | Reason, authority, expiry, attachment? | Override is traceable and reviewable |
| Integration | ERP API, amendment event, retries? | No duplicate alerts or stale revisions |
| Monitoring | Precision, recall, drift, workload? | Owner and review cadence defined |
Use PO ID, revision and line as an idempotency key so event retries do not open duplicate cases. Test out-of-order agreement and PO events. Define the fallback when AI is unavailable: a documented manual review route with a deadline and substitute approver, not unconditional passage. Specify data residency, retention, access, logging and permitted model training for confidential agreement prices. Check company and supplier obligations before sending contract text to an external model.
Pilot, FAT and SAT
The pilot should include anonymized real cases: piece versus box, pack-size change, FX date, tax or freight included twice, expired agreement, tier crossing, new item, specification revision, emergency order, approved exception and amended PO. Have buyers establish labels using source evidence, not the model’s own output. Compare vendors on the same time-separated dataset using precision, recall, alert volume, severe misses, response time and behavior when data are missing.
FAT verifies calculations, agreement dates, tier selection, authority, versioning, event retries and logs on agreed test data. A move from 20 to 25 pieces per box must invalidate the old conversion. Record inputs, expected and actual outputs and reference versions. SAT follows the actual Thai factory workflow on real terminals and ERP schedules: draft, approval, rejection, amendment, issue, outage and recovery. Verify local language, decimal formats, substitute authority and emergency buying. People must be able to understand the alert within their normal approval time.
Avoid five common failure modes
A supplier brand may represent different legal entities. Check the contract party rather than matching a display name. An item code may hide a drawing revision; compare the actual specification. A quantity discount may lower unit price while increasing inventory and obsolescence cost; inspect demand and MOQ. Emergency freight may justify a price exception, but give the exception a reason, approver and expiry instead of teaching it as normal. Tax or freight included in a contract and again on a PO must not be counted twice. These problems are best found with source-linked evidence, not an opaque anomaly score.
A practical rollout starts with one plant and category with usable agreement and unit masters. Run rule-only checks first, then measure incremental true issues found by AI. Use a shadow period before presenting alerts to approvers. Display normalized price, agreement validity, tier and source evidence on one screen. Expand only after alert volume and review time are manageable. Each new plant may need its own currency, terms and supplier-master mapping. Include master cleanup, integration, training and monthly review in total ownership cost.
Operate and audit the system
For each decision, retain original PO values, normalized values, conversion, FX source and date, agreement and price-list revision, peer-data period, model version, displayed alert and reviewer action. Retain the approver’s role and authority at that time. If a language model drafts an explanation, compute numbers from structured records and link the source clause; never accept an invented clause. Preserve old revisions after amendments.
Monitor alerts, review completion, time spent, true corrections, severe misses, unassessable cases and overrides by item category, supplier, currency, plant, contract coverage and emergency status. A small subgroup’s percentage can be unstable; show its denominator. NIST’s AI RMF emphasizes human oversight, evaluation and production monitoring, but it does not prescribe a PO pricing model or threshold. Purchasing, finance, IT and audit should review errors, drift, master-data gaps and exception misuse on a regular cadence. Test new model versions against the old on the same set before rollout, and retain a rollback path.
Summary
Manufacturing PO price anomaly detection works when it normalizes comparable PO lines before approval, applies explicit policy rules separately from AI signals, and gives an authorized human a verifiable reason to act. Include units, packs, currency, tax, freight, quantity tiers, agreement dates and revision history. Put these into the RFP, measure both false alarms and misses in a pilot, test calculations in FAT and real approval and outage paths in SAT.
If your Thai factory is still mapping PO data, unit conversions or agreement-price sources, contact TOMAS TECH to discuss the decision point and existing ERP workflow before selecting a product.
FAQ: AI for PO unit-price anomalies
How is this different from AI quotation comparison?
Quotation comparison supports supplier selection. PO anomaly detection checks the actual line to be issued or amended against applicable agreement, units, currency, quantity tier and approved terms.
Should a price-change alert automatically block a PO?
A confirmed policy or contract violation can follow the company’s stop and exception rules. A statistical outlier needs evidence-based human review; the AI score alone should not approve or block a purchase.
How much history is needed?
There is no universal count. It depends on item and supplier coverage, agreement quality and price volatility. Sparse new items need stronger agreement, quotation and expert review.
What should a procurement AI RFP include?
Include anonymized PO and agreement examples, conversion and FX rules, exception authority, ERP integration, amendment cases and measurable acceptance criteria.
What do FAT and SAT test?
FAT tests calculations, reference versions and interfaces using controlled data. SAT tests actual site roles, terminals, approval, amendment, emergency and fallback workflows.