Case Study: A Saudi manufacturer coordinating 200+ active purchase orders with 60+ global suppliers
A supplier can change a shipment date with a simple 12-word WhatsApp message, and that small update can affect production schedules, inventory levels, and cash requirements for weeks.
That’s why supplier ETD tracking is about much more than maintaining a spreadsheet. The bigger issue is visibility. Information is often scattered across different systems and communication channels. The purchase order may sit in the ERP, the latest supplier update may come through email or chat, the shipment plan may be maintained in a spreadsheet, and the production team may still be working with yesterday’s information. The issue may seem minor because each update is small. But as those updates add up, they can quickly become an operational control problem.
Agentic AI can add value here when it is designed as a controlled coordination layer. It can monitor approved communication channels, identify the purchase order or batch mentioned by the supplier, extract and standardize departure details, and compare them with the latest confirmed plan. When the context is unclear, the agent can route the information for human review rather than making assumptions. Once approved, it can update the system of record and escalate any changes that could affect production or inventory. This creates a more reliable process without removing human oversight from important decisions.
This article explains how a supplier ETD agent works in practice, what problems it can solve, where human oversight remains important, and what we learned from implementing one for a Saudi manufacturer with a large global supplier base. The case-study figures are based on outcomes the client reported from the implemented workflow. They reflect results from a specific operating environment and should be understood as project-specific outcomes, not as universal performance guarantees.
An AI agent for supplier ETD tracking continuously turns scattered supplier communications into structured and reviewable supply-chain updates. Instead of asking procurement coordinators to dig through email threads, WhatsApp chats, WeChat messages, and spreadsheets to find the latest shipping information, the agent identifies relevant messages and converts them into usable purchase-order line or batch-level data.
A production-ready agent needs to do more than just extract information. It should understand the context of supplier conversations, distinguish between an old date and a newly revised one, and make sense of approximate phrases such as “mid-April” or “after the holiday.” It should also link each update to the correct purchase order and batch, keep the original message as supporting evidence, assign a confidence score, send uncertain cases to a human for review, and sync approved information with the ERP or planning system.
The key word here is approved. In most manufacturing environments, the first practical version is not a completely autonomous procurement agent. Instead, it is an exception-management system that handles repetitive monitoring and reconciliation while people remain responsible for ambiguous situations and decisions that could have a significant financial impact.
Enterprise systems are usually good at storing a confirmed date once someone enters it. The real challenge is getting the latest update from the supplier and making sure it belongs to the correct order, item, and shipment.
Modern supply chains still rely on a mix of structured and unstructured communication. Large suppliers may share advance shipping notice (ASN) data or structured order confirmations, while others rely on spreadsheets. Many simply send updates by email. In the Middle East and Asia, suppliers and buyers often use WhatsApp or WeChat for operational updates because these are the communication channels they already use every day.
That creates a gap between supplier communication and the system where official data is stored. SAP identifies poor visibility, disconnected processes, weak communication, and manual data handling as common barriers to supply chain visibility. DCSA also notes that shipping operational data is often manual, inconsistent, and delayed. Although the technology landscape is improving, most manufacturers still need to bridge the gap between clean system data and unstructured supplier communication.
Industry Context: Industry standards are also moving toward better supply-chain visibility. SAP highlights the importance of connecting systems, tracking shipments in real time, and improving collaboration with trading partners. At the same time, the DCSA Track & Trace standard aims to make shipping events more consistent, easier to share, and interoperable across carriers and technology platforms.
A supplier might say, “first batch next week,” because the conversation already gives the buyer enough context to understand the reference. But a system looking at that message alone may not know which purchase order (PO), line item, or batch it refers to. To interpret it correctly, the agent needs to consider the supplier’s identity, conversation history, open Purchase Orders, and, when needed, previous quantities or delivery dates.
A single email can include three Purchase Orders (POs), two batches, and several delivery dates. For example, one Purchase Order might ship around March 15, another at the end of April, while a third could be delayed because a required component is still pending. An ETD agent needs to separate these details and connect each date to the correct Purchase Order rather than treating the entire email as a single shipment record.
Supply-chain communication often includes phrases such as “around the 15th,” “second week of May,” “after Chinese New Year,” “production complete by month-end,” or “vessel booking expected next week.” These phrases do not all mean the same thing. Some refer to a planned departure date, others to a production milestone, while some simply express an expectation rather than a firm commitment.
A good system should preserve that uncertainty instead of turning an approximate date into a definite commitment. For example, it might interpret “end of April” as a planning range of April 25–30, but it should not treat April 28 as the supplier’s confirmed date. The original wording, interpreted date range, and confidence level should remain visible to the reviewer.
Estimated Time of Departure (ETD) is the expected time a shipment leaves its origin or shipping point. Estimated Time of Arrival (ETA) is the expected time it reaches its destination or another key milestone. A manufacturer may also use an internal planning date that accounts for expected transit time, port handling, customs clearance, and inland transportation.
These dates should never be merged into a single value without making the distinction clear. If a supplier provides an ETD and the system calculates an ETA using historical or agreed transit times, the user interface and audit trail should clearly show that the ETA was calculated. Microsoft’s Dynamics 365 Procurement Agent follows a similar approach by recognizing the supplier-provided ETD, calculating an ETA based on agreed transit-time assumptions, and presenting the result to the buyer for validation.
A new supplier message may conflict with a formal confirmation, an ERP record, or information shared during a later phone call. In some cases, a supplier may repeat an older date in a new conversation thread. In others, different employees from the same supplier may provide different estimates. A reliable workflow should consider source, timestamp, and message context, apply clear precedence rules, and escalate conflicting information for human review rather than simply replacing the ERP date with the latest date mentioned.
A five-day delay for a non-critical consumable may require no immediate action. But a two-day delay for a constrained raw material could put an entire production line at risk. An agent becomes much more useful when it can connect the supplier update with business context, including material criticality, days of inventory cover, production requirements, available stock, alternate supply options, and customer commitments.
Email rules can route messages, while RPA can copy data when the input follows a predictable format. A traditional parser can extract a Purchase Order (PO) number when the supplier consistently places it in the same location. These tools are still useful, but they do not address the most challenging part of supplier ETD tracking: understanding the context and determining what the message actually means for operations.
Agentic AI becomes useful when a workflow involves multiple steps that require both reasoning and action. The agent can identify the supplier, retrieve the relevant open orders, interpret the message, and link each statement to the correct order line or batch. It can then compare the new date with the previous commitment, determine whether the confidence level is sufficient, and ask a person to review the case when it is not. Once approved, the agent can update the relevant business system and trigger an escalation if the change exceeds a defined planning threshold.
This shift from content generation to governed execution is also reflected in how current enterprise research views agentic AI. Deloitte describes agentic supply chains as systems that can sense changing conditions, reason across different data sources, and take action within defined governance rules. This approach allows people to spend less time on repetitive execution and more time on coordination, decision-making, and oversight.
Current enterprise direction: Deloitte’s guidance on agentic AI in supply chain management highlights supplier evaluation, constraint management, fulfillment, and logistics as areas where AI agents can monitor conditions, make decisions, and take action within controlled workflows.
The system should connect only to communication channels that the business has explicitly approved and can manage, such as a shared procurement mailbox, approved WhatsApp Business accounts, supplier portals, structured feeds, or other authorized interfaces. Personal messaging accounts should not be accessed or scraped without proper authorization. Before implementation, the business should clearly define rules for channel access, data retention, and employee privacy.
The agent should not rely on the language model to guess the meaning from the message alone. It should first retrieve relevant supplier details, open purchase orders, line items, batch schedules, quantities, previous ETDs, and recent conversation history. This context helps narrow down the possibilities and reduces the risk of linking a supplier update to the wrong order.
Supplier communication can include quotations, technical questions, payment discussions, quality issues, production updates, and even casual messages. The agent needs to distinguish an actual ETD update from other types of communication. For example, a message saying “raw material should be ready by Friday” should not automatically be treated as a shipment departure date.
The useful output is not just a date. The record needs enough context to show exactly what that date refers to. At a minimum, it should include the supplier, purchase order, relevant line item or batch when available, quantity or shipment reference if provided, the original date phrase, its normalized interpretation, and the source message.
The agent can convert natural-language date references into a structured planning range, but the normalized record should preserve the original wording and clearly indicate the level of precision. For example, “15 March” represents a specific date, while “Around 15 March” indicates an approximate date. “Mid-March” should be treated as a range rather than a fixed date. A phrase such as “After Chinese New Year” needs the relevant calendar and supplier-specific operating context before it can even be converted into a tentative planning range.
The review screen should make the key operational questions clear right away: Is the latest update earlier, unchanged, or later than the last approved ETD? If it is later, how much has it shifted? Has the quantity changed too? And does the revised date still meet the production requirement? This is much more useful than making a coordinator manually compare two supplier messages and then check the ERP system to understand what has changed.
Before accepting an update, the system can run several validation checks to make sure the information is accurate and consistent. It can confirm that the PO exists and belongs to the supplier, check whether the reported batch quantity is reasonable, and verify that the dates follow a logical sequence. It can also compare the update with recently confirmed shipment events to identify any conflicts. These validation layers help turn the language model’s output into a safer and more reliable decision that can move into the operational workflow.
When the agent is uncertain, it should take the safer route instead of making assumptions. If two POs could match the message, it should show both options and let the coordinator make the final choice. If the system cannot interpret a date with confidence, it should retain the supplier’s original wording and ask for confirmation. Similarly, if a supplier’s latest message conflicts with a previously confirmed shipment detail, the agent should flag the discrepancy for review rather than automatically overwriting the existing information.
The real business value comes when verified information reaches the systems that planners rely on every day. Once an ETD update is approved, the agent should securely write it to Oracle, SAP, Dynamics, a custom ERP, or the company’s planning database through controlled APIs. The system should also keep a clear audit trail, including who approved the update, when it was recorded, and which supplier message was used as the source.
A significant part of procurement coordination is not about interpreting information; it is about following up. The agent can keep track of POs and batches that are due for an update and draft or send a polite reminder through the supplier’s approved communication channel. Follow-up rules should consider supplier preferences, time zones, holidays, and escalation limits. This keeps the process proactive without turning the agent into an automated nuisance.
The real value is not in generating more alerts but in delivering fewer and more relevant ones. An agent should assess ETD changes alongside the broader planning context and flag only risks that require attention. These could include a material that may fall below safety stock, a batch that could affect a production date, a supplier that repeatedly changes its delivery commitment, or a shipment that is not progressing as expected.
Most ERP systems focus on the latest delivery date, but looking at the full history can provide a much clearer picture of supplier performance. By tracking every commitment and date change, businesses can measure how often suppliers revise their ETDs, calculate average delays, monitor response times and confirmation reliability, and compare promised departure dates with actual shipment movements. Over time, this historical data can highlight recurring patterns and help businesses make better supplier decisions.
Procurement coordinators spend less time searching across different channels and manually updating trackers with delivery dates. Instead, they can focus on reviewing exceptions, following up with high-risk suppliers, and resolving material issues before they impact operations.
When supplier updates reach the ERP within minutes of being reviewed, planners can rely less on outdated spreadsheets. The organization gets a more accurate, shared view of current supplier commitments, helping teams make decisions based on the latest information.
A delayed ETD is most valuable when the business knows about it before it turns into a shortage. By connecting supplier communications with inventory levels and production plans, the system can identify potential supply issues early. Instead of simply recording a date change, it can show how that change could affect materials, production, and overall operations.
Automated reminders help ensure that suppliers who have not responded and overdue confirmations are followed up on consistently. The procurement team still manages the supplier relationship, but they no longer have to depend on memory or individual to-do lists to keep track of every status request.
A manufacturer can keep the supplier’s original ETD while also using an internal, risk-adjusted date for planning based on the supplier’s past performance. The important part is to keep the distinction clear. An internal forecast should be presented as a forecast, not treated or recorded as a new supplier commitment.
Historical ETD changes give supplier reviews a clear, factual foundation. Instead of relying on memory to say that a supplier is often late, teams can look at the actual data. They can see how frequently delivery dates changed, how quickly those changes were communicated, how responsive the supplier was, and whether actual departure dates consistently differed from the ETDs provided.
Supplier messages often include business context, relationship details, and incomplete information. A responsible system should involve people whenever it cannot confidently determine a safe and reversible action. This ensures uncertain situations are reviewed before the system makes an important update or decision.
· Ambiguous PO or batch references: when the message could belong to more than one open order.
· Material changes with production or customer impact: when a date change may trigger expediting, alternate sourcing, or schedule changes.
· Conflicting supplier commitments: when two messages or two supplier contacts provide incompatible information.
· Forecast overrides: when the agent proposes a risk-adjusted planning date that differs materially from the supplier’s stated ETD.
· Commercial or relationship-sensitive follow-ups: when escalation wording could affect an important supplier relationship.
· High-impact ERP changes: when an update could trigger downstream planning, purchasing, or financial actions beyond simply recording status.
The mature state is not about having “no humans.” It is about having humans review a much smaller queue of decisions that genuinely require their judgment, while the system handles routine tasks on its own.
· You manage dozens or hundreds of active international purchase orders at the same time.
· Supplier status updates arrive through several channels and are not consistently structured.
· Procurement staff maintain manual Excel trackers because the ERP is not updated quickly enough.
· ETD changes directly affect production schedules, inventory, or customer commitments.
· Suppliers are unlikely to adopt a new portal or rigid form simply for your convenience.
· You already have a reliable PO master in Oracle, SAP, Dynamics, or another system that the agent can use as context.
· You can define clear approval and escalation rules.
· Your suppliers already provide high-quality structured ASN/EDI/API updates, and the ERP remains current automatically.
· There are very few active POs, and manual status tracking is not a meaningful cost or risk.
· The organization has no reliable supplier/PO master data to validate messages against.
· Procurement communication happens mainly through uncontrolled personal accounts that cannot be governed appropriately.
· The business expects the AI to make consequential sourcing or production decisions without human accountability from day one.
The client was a large manufacturer in Saudi Arabia, importing raw materials and components from more than 60 international suppliers. At any given time, the procurement team was managing over 200 active purchase orders, with several orders often split into partial batches scheduled across multiple months.
The client is not named in this article because the implementation included confidential procurement and supplier information. The operational figures presented below are based on project records and outcomes reported by the client after the workflow was deployed.
The client already had a purchase order system in place. The challenge was that the most important day-to-day planning updates were coming through channels outside the system, making them harder to track and use effectively.
Supplier ETD updates were spread across email, WhatsApp, WeChat, spreadsheets, and phone conversations. Based on the client’s workflow analysis, around 40% of updates came through email, 35% through WhatsApp, 15% through WeChat, and the remaining 10% through phone calls or manually summarized communications. These figures were specific to this client’s workflow and should not be considered an industry benchmark.
The format of these updates varied widely. One supplier might send a structured spreadsheet with the PO number, line item, batch size, and departure date. Another might write, “PO-4521 first batch 500K around March 15, second batch end of April.” A supplier in China might send a mixed Chinese-English message through WeChat, while a WhatsApp update could include a photo of a schedule instead of written details. Some messages even referred to earlier conversations without mentioning the PO number again.
Three procurement coordinators maintained a master tracker with more than 2,000 rows. Each day, the team spent roughly four to five combined hours reviewing supplier messages, finding the correct PO and batch, interpreting date information, updating the tracker, and keeping the developing Oracle APEX environment aligned.
The biggest problem was not the manual work itself, but the delay and lack of confidence in the data. By the time an ETD update reached the planning system, another supplier message could have already changed it. Before the new workflow was introduced, the client estimated that around 15% of ETD records were outdated or inaccurate at any given time.
The client had already tried using standard email folders and filtering rules. These helped organize the inbox, but they could not understand the actual content of each message. A filter could identify a supplier’s name, but it could not reliably tell that a sentence changed the first batch of PO-4521 while leaving the second batch unchanged.
The team also tried encouraging suppliers to use a structured update form, but adoption remained below 20%. The key lesson was simple: an internal visibility project cannot assume that every supplier will change the way they communicate. Instead, the system needs to work with the supplier channels that are already part of the daily workflow.
Traditional RPA was also not a good fit for the unstructured parts of the workflow. The process involved multilingual messages, unclear references, different date formats, attachments, and previous conversation history. These inputs needed to be understood in context before the system could apply rules or trigger the right actions.
· Monitor approved supplier communication channels without forcing suppliers onto a new portal.
· Detect messages that contained ETD or shipment-timing information.
· Match each update to the correct supplier, PO, line item, or batch.
· Interpret English, Chinese, and mixed-language communication.
· Normalize approximate date phrases while preserving the original language and uncertainty.
· Show whether the new ETD was earlier, unchanged, or delayed versus the previous approved ETD.
· Allow procurement coordinators to approve or correct the interpretation quickly.
· Write approved updates into Oracle APEX and synchronize manual corrections back to the agent layer.
· Follow up when expected supplier updates were missing.
· Escalate changes that could affect production or inventory.
· Keep the history of ETD revisions for supplier-performance analysis.
Our company started with the shared procurement mailbox because email was the largest communication channel and offered the most controlled environment for testing. Using Microsoft Graph, the system monitored incoming supplier emails, grouped related messages into threads, and identified whether each message contained relevant ETD information.
Before interpreting each message, the agent first retrieved the relevant open-order details. It identified PO numbers, line items, and batches where possible, then extracted the supplier’s timing information and converted vague date phrases into practical planning ranges. During this first phase, the agent did not update the ERP directly. Instead, its outputs were saved in a structured review file, allowing the procurement team to compare the AI-generated interpretation with the existing manual tracker.
This was intentionally designed as a trust-building phase. The goal was not to automate the entire process within four weeks. Instead, the focus was on proving that the system could understand real supplier communication, highlight uncertainty, and make it easy for the team to review and correct its decisions.
Once email performance had stabilized, the system was expanded to approved WhatsApp Business messaging and the client’s WeChat integration. This moved the solution beyond email extraction, turning it into a broader agent for handling supplier communications across multiple channels.
Each message was stored with key source details, including the supplier, timestamp, communication channel, original message, detected language, attachment status, translation when needed, and any identified PO references. Attachments were linked directly to the related message instead of being stored separately, preserving the conversation context that helped explain them.
A React-based review dashboard became the procurement team’s main workspace. Coordinators could view the original message, translated text when needed, the extracted or interpreted ETD, previous ETD, confidence level, change direction, and suggested action. The dashboard was designed to support quick review without overwhelming users with unnecessary model details.
The next stage integrated approved ETD updates with the client’s Oracle APEX purchase-order tracking application through REST APIs. However, the agent did not automatically send every model-generated update to Oracle. Only updates that met the required confidence threshold and passed the defined review process were eligible for synchronization.
The integration worked in both directions. If a procurement manager changed an ETD in Oracle after speaking with a supplier or making an internal decision, the updated value was synced back to the agent database. This kept both systems aligned and prevented the AI layer and operational system from maintaining different versions of the same information.
Every approved change also created a history record. This allowed the system to track the original commitment, each subsequent revision, who approved the change, and how the current ETD evolved.
Once the team was confident in the data flow, the agent also started handling overdue status follow-ups. Rules determined when a supplier or PO batch was due for a new confirmation. The system could then prepare or send a polite follow-up through the approved supplier channel, using the supplier’s preferred communication method and the correct order details.
This removed a repetitive part of the procurement workload without trying to automate supplier relationship management. Coordinators remained responsible for strategic discussions, negotiations, and unusual supplier issues that required human judgment.
Once enough ETD history had been collected, the client could compare supplier-stated dates with their actual delivery patterns. For suppliers that consistently departed later than their stated ETDs, the system could recommend an internal planning buffer to help teams plan more realistically.
A key design decision was to keep the supplier’s stated ETD separate from the predicted planning date. The forecast did not change or overwrite the supplier’s original commitment. Instead, it gave planners a second, clearly labelled signal based on actual supplier behavior. This made the system more useful for planning while keeping the data transparent and auditable.
The agent could flag delayed updates when they crossed predefined risk thresholds. For high-priority materials, the escalation included the latest supplier message, previous commitment, proposed new ETD, relevant PO and batch details, and recommended next steps. These could include confirming the update with the supplier, checking alternative stock, expediting a batch, or notifying the production planning team.
The agent did not make decisions such as switching suppliers or changing the production plan. Those decisions remained with the procurement and planning teams. Its role was to give the right people the relevant evidence early enough to take action.
The most difficult messages were not always the longest. Often, they were short updates that relied on shared context from earlier conversations.
One common example combined Chinese and English in the same message. It might include a PO reference, a Chinese date phrase, shorthand quantities, and an English note about a later batch. The system had to recognize that the message referred to more than one shipment event, interpret and standardize the Chinese date, and distinguish between an approximate date and a firm commitment.
For holiday-related phrases, the system used holiday calendars and the supplier’s location to provide context before suggesting a planning range. The original wording was still shown so the team could review it. We found that interpreting approximate dates is not just a language issue; it also depends on business context. A supplier’s past behavior can be an important factor.
When a message referred to multiple open POs and the correct match was still unclear after checking the conversation history, the system did not guess. Instead, it sent the message to a review queue with the possible matches and a clear reason for escalation.
The system used a service-based architecture that allowed inbound communication, language processing, approvals, and ERP synchronization to run asynchronously without blocking each other.
· Python and FastAPI for backend services and APIs.
· Celery and Redis for asynchronous message processing and job queues.
· Microsoft Graph API for approved shared mailbox monitoring.
· WhatsApp Business API and the client-approved WeChat integration path for messaging intake.
· A large language model for multilingual message understanding, thread-aware extraction, and approximate-date interpretation.
· PostgreSQL for structured supplier, message, ETD, approval, and history records.
· React for the human review and exception dashboard.
· Oracle APEX REST APIs for controlled two-way synchronization with the client’s internal application.
· AWS infrastructure for compute, database, storage, and event-driven processing in the deployed architecture.
The model was only one part of the solution. Most of the production value came from the controls around it, including context retrieval, deterministic validation, approval workflows, source evidence, identity checks, API integration, error handling, and observability.
The project was evaluated based on operational outcomes rather than model accuracy alone. The figures below are specific to this client and reflect the observation periods used during the project.
ETD record accuracy: Improved from approximately 85% before the project to 97.5% after the supervised workflow had stabilized. This measures the quality of approved ETD records used by the business after extraction, validation, and human confirmation. It does not mean the language model independently identified the correct dates 97.5% of the time without human review.
Manual tracking effort: Reduced from roughly 4–5 team-hours per day across three procurement coordinators to about 45 minutes of exception review per day, mainly handled by one coordinator once the workflow had matured.
Supplier-message-to-system latency: Reduced from potentially several hours to an average of about 12 minutes in the implemented workflow, including AI processing and the required human approval step.
Oracle APEX ETD completeness: Improved from roughly 60% to approximately 98%, as approved supplier updates were added to the system automatically without requiring separate manual data entry.
Supplier response within 24 hours: The client measured an improvement from around 50% with manual follow-ups to approximately 85% after introducing structured, channel-aware follow-up automation.
Material-shortage production stoppages: The client had previously recorded around three to four shortage-related stoppages per quarter. During the six-month observation period after the integrated Phase 3 workflow was implemented, the client reported no such events. This is a significant operational result, but it should not be taken as proof that the ETD agent alone caused the improvement. Procurement actions, inventory decisions, and supplier conditions also played a role.
Working capital: The client estimated that better ETD visibility and confidence allowed them to reduce some uncertainty-driven safety stock, releasing approximately USD 2 million in working capital. This was the client’s finance estimate and should not be treated as a guaranteed return directly attributable to the software.
Accuracy figures in AI case studies can be easy to misinterpret. For this project, the key measure was not simply whether the LLM extracted a date correctly. The more important operational question was whether the ETD record used by the planning team was current, correctly mapped to the right order context, and supported by the latest approved supplier communication.
The 97.5% figure reflects the quality of the supervised workflow after context retrieval, extraction, validation, confidence checks, and human correction. This is a more meaningful business metric than raw extraction accuracy, but it also shows that the result came from the complete system and review process, not the model alone.
The client’s working-capital benefit came from greater confidence in supplier ETDs. When planners cannot rely on incoming dates, they may protect production by holding extra safety stock. Better ETD visibility reduces some of that uncertainty, allowing the business to review and potentially reduce inventory buffers for selected materials.
The client estimated that this contributed to the release of roughly USD 2 million in working capital. The agent did not directly “generate USD 2 million.” The financial impact depended on factors such as purchasing volumes, material values, safety-stock policies, lead times, supplier reliability, and management decisions. Another manufacturer with different operating conditions could see a much smaller or larger impact.
· The agent did not autonomously change suppliers.
· It did not cancel or reissue purchase orders because a supplier date changed.
· It did not silently convert every vague date phrase into a hard commitment.
· It did not overwrite a manually confirmed ETD without preserving source and history.
· It did not change production schedules on its own.
· It did not treat a predicted internal planning date as if the supplier had confirmed it.
· It did not guess when the message could reasonably map to more than one PO or batch.
Those boundaries were important. They kept the system useful while ensuring that people remained accountable for decisions with real operational or commercial consequences.
Many companies treat email and chat as separate from their core ERP data. In practice, these channels often contain the latest operational updates. A visibility program that ignores them can end up with a clean dashboard built on outdated information.
We started with email because it had the highest volume and was easier to govern from a technical perspective. During the project, however, a time-sensitive WhatsApp update showed why volume alone is not enough to determine channel priority. A lower-volume channel can still carry highly urgent operational information. Future implementations should therefore assess communication channels based on both volume and business impact.
Users are more likely to trust a system when it recognizes that “mid-April” is a range rather than forcing it into an exact date. False precision may make AI outputs look cleaner, but it can make planning less reliable. The original phrase, its interpretation, and the confidence level should therefore be stored separately so users can see what the supplier actually said and how the system interpreted it.
A dashboard can accurately summarize supplier messages, but it still creates a manual handoff if planners do their work in Oracle, SAP, or another business system. For the solution to improve day-to-day operations, it needs reliable synchronization, clear change history, and proper error recovery. These are core parts of the product, not technical details that should be added later.
Procurement coordinators adopted the workflow more easily when they could see the original message, extracted interpretation, previous ETD, date change, and confidence level in one place. A well-designed human review process turns AI uncertainty into a clear, manageable queue that teams can work through efficiently.
Historical supplier behavior can help improve planning, but a predicted date should never be treated as a confirmed supplier commitment. Keeping these values separate preserves the evidence behind each decision and allows planners to use forecasting without confusing it with supplier accountability.
A well-planned implementation should establish baseline metrics before deployment. Without a clear starting point, the organization may end up tracking model activity instead of measuring whether the project is actually improving business performance.
· ETD record accuracy: percentage of current ETD records that match the latest approved source information.
· ETD completeness: percentage of active PO lines or batches with a usable current ETD.
· Update latency: time from supplier message receipt to approved system-of-record update.
· Manual coordination effort: team-hours spent finding, interpreting, copying, and following up on ETD information.
· Exception rate: percentage of supplier messages that require manual interpretation.
· False mapping rate: updates incorrectly associated with the wrong PO, line, or batch.
· Supplier response time: percentage of follow-up requests answered within the defined SLA or target window.
· Date volatility by supplier: frequency and magnitude of ETD revisions before actual departure.
· Shortage or expedite events: production-impacting shortages or premium-freight decisions linked to late visibility.
· ERP data completeness and freshness: how much of the planning system reflects the latest approved supplier information.
· User correction rate: how frequently coordinators need to correct the agent and what categories cause those corrections.
Accuracy should always be broken down by message type and complexity. A 98% result on simple English emails does not show how well the system handles mixed-language WhatsApp images or messages containing multiple ambiguous POs. To get a more realistic picture of performance, these difficult categories should be evaluated separately.
Start by mapping every channel suppliers actually use, not just the ones listed in the procurement policy. Identify the systems that hold PO master data, the planning decisions that depend on ETD information, the current tracking process, the people responsible for approving changes, and the situations where an incorrect date could have a significant operational impact.
Start with one controlled communication channel and a representative group of suppliers. Let the agent extract and structure supplier updates without making any changes to the ERP. Compare its output with human-reviewed results and classify errors by type, such as PO mapping, date interpretation, language handling, attachment parsing, conversation context, or business-rule validation.
Once the system is consistently useful, give coordinators a review interface that clearly shows the source evidence and what has changed. Track metrics such as review time and correction rate to understand how well the workflow is performing. The goal is to make human confirmation much faster than manually reading through the entire inbox.
Only approved or high-confidence updates that meet the required policies should be written to the ERP. The integration should include duplicate-update protection, retry handling, authorization, change history, and a safe rollback process. If users can also edit the same ETD directly in the ERP, two-way synchronization should keep both systems aligned and prevent silent differences between them.
Add WhatsApp Business, supplier portals, and other relevant channels based on their operational value. Introduce overdue-update detection and supplier follow-ups gradually, with clear communication rules, defined escalation limits, and appropriate human oversight.
Only after a reliable history of ETD data has been established should the project introduce supplier-specific planning buffers or predictive signals. Keep forecasts separate from confirmed supplier commitments, and validate whether these signals actually lead to better decisions around shortages, expediting, or inventory levels.
· Approved channel access only, with explicit ownership of each mailbox, messaging account, and integration.
· Least-privilege permissions for ERP and procurement-system actions.
· Supplier and employee data retention rules appropriate to the jurisdictions involved.
· Encryption in transit and at rest for messages, attachments, and extracted business data.
· Clear separation between model inference, deterministic validation, and authorized system actions.
· Source evidence retained for each approved ETD change.
· Immutable or controlled audit history showing message, interpretation, approver, time, and downstream system update.
· Confidence thresholds and explicit fallback behavior when the model cannot safely resolve context.
· Versioning for prompts, models, date-normalization logic, and business rules.
· Monitoring for integration failures, delayed queues, duplicate messages, and stale credentials.
· A rollback procedure for a bad model or rule release.
· A documented human accountability model for high-impact procurement and planning decisions.
DHL’s current logistics research also highlights that successful AI adoption depends on data quality, system integration, security, governance, and responsible-use safeguards, not just model capability. This becomes even more important when AI agents move beyond summarizing information and start making changes to enterprise records.
A demo that only summarizes an email is not enough. Ask how the system handles an implicit PO reference, a message involving multiple POs, and a date update that conflicts with an existing commitment.
If the vendor stores all of these dates in a single field, the audit trail becomes weak. The system should keep separate records for what the supplier actually said and what the AI inferred. This makes the source of each date clear and keeps the data easier to review and audit.
The right approach should combine confidence scoring, candidate matching, and human review. Simply saying that “the model usually figures it out” is not a reliable enterprise control.
Real supplier messages often leave out important details because people assume the other person remembers the earlier conversation. The system therefore needs controlled context retrieval so it can connect the current message with relevant conversation history and ERP data, rather than treating each message as a standalone piece of information.
Ask for examples that reflect your actual supplier base. Test real or representative messages from China, India, Europe, and other major sourcing markets instead of relying on generic English templates. This provides a more realistic view of how the system handles different languages, formats, and communication styles.
The integration should use approved business interfaces or client-managed connector methods. Avoid approaches that rely on scraping personal devices or using uncontrolled credentials. This helps keep supplier communications secure, governed, and easier to audit.
Ask how the integration handles two-way updates, conflicting changes, retry attempts, duplicate updates, and temporary ERP downtime. The provider should explain how each situation is detected, handled, and recovered without losing or overwriting important data.
A reliable evaluation should separate extraction accuracy, PO mapping accuracy, date normalization, approved ETD record accuracy, and exception rates. A single overall “AI accuracy” percentage can hide the specific failure points that have the biggest impact on operations.
Permissions should be clearly defined and easy to configure. Reading messages, drafting follow-ups, updating a low-risk status field, and changing a production plan all involve different levels of risk. These differences should be reflected in the controls governing what the agent can do automatically.
Supplier disputes and performance reviews can arise months after the original message was sent. The system should retain the original communication, its interpretation, the approval decision, and the full change history so the team can clearly understand why an ETD was changed.
Ask how the provider handles versioning, testing, gradual rollouts, and rollbacks. A supplier communication agent is part of the operational system, so it should follow proper software development, testing, and release practices.
Your team should understand the integrations, business rules, confidence thresholds, model dependencies, monitoring setup, and support process. The goal is to build sustainable ownership rather than create long-term dependence on a single developer for every supplier-related change.
Supplier ETD tracking involves keeping the latest expected departure date for inbound purchase orders, batches, or shipments. Manufacturers rely on this information to plan material availability, inventory, production, and downstream logistics. The challenge is that ETD changes are often shared outside the ERP through email, chat, spreadsheets, or phone calls, making them difficult to capture and keep up to date.
An AI agent for ETD tracking is an AI-powered workflow that monitors approved supplier communications, identifies ETD updates, links them to the correct PO or batch, interprets date changes, and compares them with existing commitments. When needed, it can request human approval before updating the relevant enterprise system.
ETD means the Estimated Time of Departure from the origin or a relevant shipping point, while ETA means the Estimated Time of Arrival at the destination or a later milestone. An AI agent can use the ETD and expected transit time to estimate an ETA, but that calculated date should always be kept separate from the supplier’s stated ETD.
Yes. An AI agent can process supplier updates from WhatsApp when the organization uses an approved WhatsApp Business integration with appropriate permissions, data-retention policies, and security controls. It can analyze text and supported attachments, while ambiguous or low-confidence updates should be sent to a human for review.
Yes, provided the client has a lawful and reliable integration for the relevant WeChat business workflow. The main challenge is usually not translation alone. The system also needs to handle mixed-language shorthand, understand the conversation context, and accurately link each update to the correct order.
Yes. In the case study described here, approved ETD updates were connected to an Oracle APEX application using REST APIs. Changes made directly in Oracle were also sent back to the agent layer, helping keep both systems aligned and the data consistent.
Yes, provided the required purchasing and order data is available through supported APIs or integration services. The approach is the same: retrieve the approved PO context, interpret the supplier communication, validate the update, and write the approved result using controlled access.
Yes, for clearly defined, low-risk follow-up situations. The system can detect an overdue confirmation and send or prepare a message through the supplier’s approved communication channel. High-value escalations and sensitive supplier relationship issues should remain under human control.
Historical ETD changes, supplier response patterns, and actual departure records can help identify potential delay risks. These predictions should be used as planning signals, not treated as confirmed supplier commitments. Before planners use them to guide inventory or production decisions, the model should be tested against actual outcomes to confirm how reliably it forecasts delays.
The better use case is to reduce repetitive monitoring, data entry, and supplier follow-ups so procurement teams can spend more time on supplier relationships, negotiations, risk management, and commercial decisions. Human judgment remains important for unclear, high-impact, and relationship-sensitive situations.
A focused, read-only pilot using one communication channel can often be built within a few weeks, provided the PO data and channel access are already available. A production-ready system covering multiple messaging channels, human review, ERP synchronization, automated follow-ups, security controls, and monitoring usually needs a phased rollout over several months. In most cases, the timeline depends more on data and integration readiness than on the AI model itself.
Start with the communication channel and supplier group that creates the biggest coordination burden and has clear PO context. Email is often a practical starting point, but volume should not be the only consideration. Channel urgency matters too. If critical supplier updates mainly arrive through WhatsApp, delaying that channel for too long can reduce the business value of the automation.
Measure ROI by tracking saved coordination time, reduction in outdated or missing ETDs, faster update processing, fewer expedite or shortage events, improved supplier-response performance, better ERP data completeness, and any inventory benefits linked to better planning confidence. Avoid treating every prevented disruption as if it were caused by AI alone, since other operational factors may also contribute.
Supplier ETD tracking may seem like a simple administrative task, but it becomes much more complex when a manufacturer manages hundreds of open orders across dozens of suppliers. Every informal message can become a potential planning event, while the gap between supplier communication and the ERP can create a significant source of operational risk.
Agentic AI is useful because it can operate within this gap. It can monitor the communication channels suppliers already use, retrieve the relevant purchase-order context, understand what has changed, flag uncertainty, involve a person when needed, and update the system of record after approval. It can then bring only meaningful exceptions to the procurement and planning teams.
The Saudi implementation described in this article was not successful simply because an LLM could identify dates. Its success came from a complete workflow that combined language understanding with PO context, validation, human review, Oracle APEX integration, structured follow-ups, historical tracking, and managed escalation.
For manufacturers, importers, and distribution businesses facing the same challenge, the first question should not be, “Which AI model should we use?” A better question is, “Which supplier updates do our teams repeatedly search for, copy, reconcile, and follow up on—and which business decisions depend on getting that information right?” That is the workflow worth automating first.
Related aTeam capability: AI-driven solutions for logistics and supply chain – including supplier monitoring, procurement agents, document processing, customs workflows, and shipment exception automation.
