A big manufacturer in Saudi Arabia was juggling over 200 active purchase orders from more than 60 international suppliers. However, the key planning information they needed — the supplier’s actual Estimated Time of Departure (ETD) — was buried in emails, WhatsApp chats, WeChat messages, spreadsheets, and even handwritten notes from phone calls. Three procurement coordinators were spending around 4-5 hours each day sifting through supplier updates, trying to match them to the correct PO and batch, deciphering vague phrases like “around mid-March” or “after Chinese New Year,” and updating a 2,000-row Excel tracker that the production planning relied on.
The issue wasn’t a lack of effort but rather the way the work was structured. Updates came in various languages, through different channels, with inconsistent formats and frequent ambiguities. Missing or misinterpreting an ETD could either lead to a stockout that halted production or excess inventory that tied up capital.
At aTeam Soft Solutions, we created an AI agent for tracking supplier communications that observed incoming messages, pulled ETD updates at the PO-line level, standardized date expressions, highlighted changes, and gradually fed verified information into the client’s Oracle APEX environment. This transformation led to a switch from manually pursuing ETDs to a more organized exception management system: achieving 97.5% ETD accuracy, updates seamlessly entering the system in minutes rather than hours, preventing any production delays due to material shortages post-launch, and improving cash flow management across international purchasing.
When we first explored this workflow, we discovered that the client’s ETD issue wasn’t just a spreadsheet issue—it was actually a communication issue that was masquerading as a tracking problem.
This manufacturer imports bulk raw materials and components from suppliers spanning China, India, Germany, Turkey, and Southeast Asia. They place large orders that can range from 5 million to 10 million units for one item, but those units don’t arrive in a single, neat shipment. Instead, they come in batches over a span of three to six months. So, the crucial operational question shifts from “Was the PO issued?” to “When is each batch really scheduled to leave?”
Theoretically, this should be simple, but in practice, it turned out to be quite chaotic.
About 40% of supplier ETD updates came via email, while another 35% were shared through WhatsApp. Approximately 15% of updates were communicated through WeChat, particularly from suppliers in China. The last 10% were taken care of through phone calls, and often, a coordinator would jot down a quick note in Excel or send a summary message to the team afterward. Suppliers chose whatever communication method felt most comfortable for them, with some being quite structured and others less so.
For example, one supplier might send a detailed Excel sheet with the PO number, line item, batch size, and expected departure date. Another might simply write a brief email stating, “PO-4521 first batch 500,000 pieces will ship around March 15, second batch end of April.” A WhatsApp update could even include a photo of a handwritten production schedule, while a WeChat message might mix languages, saying half in Chinese and finishing with something like “second batch maybe end of April” in English. Some suppliers would refer back to earlier conversations without repeating the PO details, assuming the buyer was already familiar with the context.
The three procurement coordinators worked hard to create a single reliable master tracker. Each day, they opened the shared inbox, checked their WhatsApp and WeChat messages, skimmed through previous conversations, and searched for any clues about shipment movements. Then, they carefully mapped each update to the correct purchase order, ensuring it matched the right line item or batch.
It may sound straightforward, but the sheer volume was overwhelming. With over 200 active POs, many batches per PO, and more than 60 suppliers communicating at various speeds and formats, the tracker expanded to more than 2,000 rows. It was not just a convenience tool; it was crucial for live planning related to production scheduling, material availability, and timely decision-making.
If the estimated time of departure (ETD) for an essential raw material was outdated, the factory could mistakenly believe stock would be delivered next week when, in reality, it was delayed by ten days. This miscommunication could lead to production stoppages costing between $50,000 and $100,000 each day. Conversely, if the planning teams didn’t trust the tracker, they might hold onto extra safety stock, tying up valuable working capital.
The coordinators were really putting in a lot of effort, but the nature of the process made it nearly impossible to achieve perfect accuracy. They had to interpret ambiguous language, like transforming “End of April” into a specific date range. Phrases like “After Chinese New Year” were dependent on the calendar year, the supplier’s location, and how quickly the supplier typically resumed production after the holiday. Even a simple message stating “first batch next week” needed to be translated into a date that the planning teams could actually work with.
On top of that, the client was working on an Oracle APEX application for tracking purchase orders and shipments. However, the Estimated Time of Departure (ETD) information entering the system was often incomplete since the coordinators couldn’t keep up with the rapid updates. Manual data entry always lagged behind the incoming messages.
This effort, shared among three people, took about 4-5 hours each day. Yet, despite their hard work, the client determined that about 15% of ETD entries were outdated or incorrect at any moment. That’s the kind of unnoticed planning mistake that can disrupt operations long before anyone officially points it out.
This client had already explored the usual solutions.
They set up email filters and folder rules, which helped them sort incoming messages, but didn’t actually tackle the root issue. Organizing messages isn’t the same as pulling ETD data out of them. A supplier’s email could mention several POs, various batches, estimated timing, and potential delays, all without a consistent format. So, the filters just transferred that complexity to a different folder.
They also attempted to guide suppliers to use a Google Form for ETD updates. However, it didn’t work for a straightforward reason that many procurement teams quickly recognize: suppliers aren’t going to change their communication habits just because one buyer wants tidier data. As a result, compliance was below 20%. Suppliers who were already comfortable with email, WhatsApp, or WeChat continued to prefer those platforms.
Traditional RPA and ready-made automation tools just weren’t the right choice for this situation. While they excel at processing structured inputs, they struggle with multilingual, multi-channel, and human-written updates that can be ambiguous for businesses. For instance, a WhatsApp message with a photo attached is beyond the capabilities of a simple workflow tool to interpret. Similarly, a WeChat update that mixes Chinese and English cannot be converted into precise ETD records by a mail parser. If a supplier mentions, “delayed because of the component shortage you know about,” it requires understanding the context, rather than just matching keywords.
That’s why the workflow wanted an AI agent instead of a typical automation script. The main challenge was understanding the language alongside the business context. The system needed to identify which purchase order was being referred to, which batch or line item the message discussed, what the supplier intended regarding timing, how the new estimated delivery time varied from previous ones, and whether the change warranted escalation. This is truly an AI-based ETD tracking automation challenge, rather than just an inbox-management issue.
For a manufacturer running in the Middle East relying on international suppliers, timely updates are crucial for production planning. This is precisely the sort of workflow where agentic AI can make a significant impact—by reading, interpreting, suggesting, escalating, and ultimately acting, all under clear human supervision.
We intentionally chose not to start with “full automation.” In the world of supply chains, building trust is more important than just moving fast from the get-go. If the estimated time of departure (ETD) is incorrect, then production planning will also go awry. That’s why we organized the rollout into phases, allowing the client to verify the value at each step and gradually take on more responsibilities.
During our initial four weeks, we concentrated solely on email as it was the most significant communication channel and allowed us to create a controlled pilot environment. We set up an inbox-monitoring service linked to the procurement team’s shared mailbox through the Microsoft Graph API. Each incoming supplier email was processed, organized into the appropriate thread, and assessed by the AI agent.
The agent went beyond simply extracting dates; it aimed to address a more practical question: what exactly has changed here?
By using Claude Sonnet, the system was able to read the email content, pinpoint referenced PO numbers, and identify specific batches or line items whenever possible. It also extracted the Estimated Time of Departure (ETD) language and converted vague timeframes into clear date ranges. For instance, “end of April” was transformed into April 25-30 instead of just a single day, and “around March 15” was converted into a defined planning range rather than an imprecise estimate. The results were organized into a structured Excel file that aligned perfectly with the client’s existing format, allowing the procurement team to directly compare the AI output with their manual tracking system.
This phase served two purposes. The first was technical: to demonstrate that we could accurately interpret the often-messy communications from suppliers. The second was operational: to foster enough trust in the system so the coordinators would start relying on the outputs. We aimed for an extraction accuracy of over 90% on the pilot POs, not because that was our ultimate goal, but because it was sufficient to begin shifting the team’s workflow.
Once we successfully established stable email extraction, we expanded our system to include WhatsApp and WeChat. This marked a significant transition for the project; it evolved into a true AI agent for supplier management instead of just a message parser.
We utilized the WhatsApp Business API along with a WeChat bridge layer to funnel supplier messages into a single pipeline. This development allowed the agent to now enjoy a cohesive intake stream across all major channels, eliminating the previous fragmented communication that required human oversight. Each message was saved with important metadata: supplier name, timestamp, communication channel, message body, language, attachment status, translated version when necessary, and any inferred purchase order references.
Next, we created a review dashboard in React that centralized all ETD-related updates. This was a game-changer! Procurement coordinators no longer had to toggle between Outlook, WhatsApp, WeChat, and Excel to figure out what was happening. They could simply open one interface and review each update, seeing the original message, the translated or adjusted content, the AI-extracted ETD values, confidence scores, and any discrepancies with the prior ETD for that purchase order line.
This is where the importance of human-in-the-loop design really shines! Every update that was extracted was carefully reviewed and approved. Any necessary corrections were noted and incorporated back into the extraction logic. Plus, our dashboard showed whether the new Estimated Time of Departure (ETD) was earlier, unchanged, or delayed compared to the last known status, transforming review from just passive reading into quick operational decision-making.
We also added automated follow-up logic. If an ETD update wasn’t received within the expected communication window for a supplier or PO batch, the AI agent would send a friendly follow-up message using the supplier’s preferred communication method. This approach isn’t about automation for its own sake; it’s a straightforward method to ease one of the most time-consuming tasks for our coordinators: manually tracking down status updates.
Our next step involved integrating with Oracle APEX, as extracting ETDs is only beneficial if the business can reliably trust and utilize them in their processes.
To achieve this, we created an API layer that connects the ETD database with the client’s Oracle APEX application. After an update is reviewed and approved, it automatically transfers into the PO tracking module, which significantly enhances data accuracy within the client’s internal system—removing the previous bottleneck caused by manual re-entry.
Additionally, we implemented two-way synchronization. If a procurement manager needed to make manual adjustments to an ETD in Oracle APEX following a supplier call or internal issue, that update would then be sent back to the agent database. This measure ensured that the two systems stayed aligned, which is crucial since many projects can hit roadblocks when AI-generated data and operational data start to deviate.
At this point, we also introduced historical ETD tracking, which was a significant enhancement. We recorded every ETD update for each PO line over time, allowing the client to monitor ETD changes, not just the current ETD. This way, they could assess supplier performance by looking at how frequently the stated departure dates changed, how much those dates were delayed, and how responsive suppliers were when requested for confirmation. Ultimately, this provided a much more solid foundation for supplier reviews compared to the old Excel tracker!
By months 5 to 8, the system had gathered enough data to become more proactive.
We rolled out predictive ETD (Estimated Time of Departure) logic informed by supplier behavior patterns. If a supplier regularly shipped five to seven days later than their stated ETD, the system could suggest a planning buffer. It didn’t replace the supplier’s declared ETD but instead provided planners with a more realistic picture of operations.
We also implemented escalation logic. If the forecasted risk of delay suggested that a shipment might miss an important production date, the AI agent would automatically escalate the issue. It would first alert the procurement coordinator, and if the situation wasn’t resolved, it would notify the procurement manager. Additionally, it included a suggested course of action: follow up with the supplier, expedite an alternate batch, evaluate substitute stock, or inform planning.
At this stage, the system evolved beyond just reading messages. It functioned as an AI agent to track supplier communications, interfacing with procurement operations, internal systems, and planning priorities. In the context of a manufacturing environment in Saudi Arabia with large international purchasing flows, that’s where the real value lies: not in having cleaner dashboards, but in reducing the number of surprises that make it to the factory floor.
We developed a solution that features a multi-service architecture specifically created for asynchronous communication, language interpretation, review workflows, and enterprise system integration.
The backend was crafted using Python with FastAPI, and we utilized Celery and Redis for asynchronous processing. This setup enabled inbound messages from platforms like email, WhatsApp, and WeChat to be queued, parsed, translated when necessary, assessed for purchase order relevance, and routed into the review pipeline—all without interrupting other processes. We also employed the Microsoft Graph API to keep an eye on shared inboxes. The WhatsApp Business API and a WeChat connector integrated messaging data into the main system seamlessly. Additionally, attachments like Excel sheets, images, and PDFs were stored in S3 and connected to their corresponding message records.
For multilingual text comprehension, thread-aware ETD extraction, mixed-language parsing, and vague-date interpretation, we relied on Claude Sonnet. However, we didn’t just take its outputs at face value. We implemented validation layers to verify details such as whether the specified purchase order was valid, if the line or batch quantity was sensible, if the supplier matched the purchase order owner, and if the extracted date range aligned with known shipment sequences.
PostgreSQL was used as the main database for messages, neatly organized ETD records, approval statuses, supplier mappings, and ETD history. The review dashboard, built with React, provided coordinators with a quick way to check out the original supplier message, the translated content, the extracted ETD, confidence levels, previous ETD comparisons, and suggested next steps. This dashboard was thoughtfully designed to assist coordinators in making their crucial decisions: whether to approve, correct, escalate, or simply disregard as non-ETD content.
We handled Oracle APEX integration using REST APIs. Approved ETD updates were pushed right into the client’s PO tracking module, while any manual corrections made in Oracle APEX were synced back to the AI system to keep everything in harmony.
We also set up fallback logic. If the system couldn’t confidently identify the PO reference, it wouldn’t take a guess. Instead, it would send the message into a review queue with reasons like “multiple candidate POs” or “implicit supplier reference only.” If a date phrase couldn’t be normalized smoothly, it would show the original phrase along with a suggested date range for approval. Whenever the message context relied on a previous conversation, the agent would pull in the thread history to enhance understanding before escalation.
At aTeam Soft Solutions, this is our approach to AI procurement automation projects in the Middle East: viewing it not just as a model beside a spreadsheet, but as a well-managed operational layer that reads messages, interprets context, validates data, and integrates seamlessly with the business systems that the people really use.
The most challenging communication platform we encountered was WeChat. This wasn’t due to a high volume of messages but rather because of its highly complex language use.
For instance, a single update could state: “PO4521 第一批 500K pcs ETD 3月15号左右, 第二批 end of April.” It blends both Chinese and English in one message, includes shorthand for quantities, informal language, and a somewhat unusual structure. The AI agent needed to decipher that the supplier was referring to two different batches, not just one, and that “3月15号左右” indicated an approximate date of March 15, rather than a firm deadline. To enhance our extraction accuracy, we employed multilingual prompting techniques and examples drawn from actual supplier communication styles.
It turns out that working with approximate language is more complex than we initially thought. For instance, saying “After Chinese New Year” isn’t an exact date. Instead, it reflects business assumptions associated with holiday schedules, regional restart patterns, when the workforce returns, and the intricacies of the products involved. To tackle this, we developed a contextual date resolver that took into account holiday calendars, supplier locations, and recognized lead times before suggesting a planning range.
Another challenging situation arose with messages that mentioned multiple purchase orders (POs) without a clear organization. For example, a supplier might say, “PO-4521 and PO-4523 will have their first batches next week, while PO-4530 is delayed until May due to the component shortage you’re aware of.” To accurately process this, our system needed to break down the message, associate each estimated time of departure (ETD) fragment with the correct PO, and occasionally refer back to earlier conversations to clarify any implied references.
Then came the adoption phase! At first, some coordinators were a bit concerned that by approving the AI’s output, they might be held accountable for mistakes they didn’t make themselves. Others felt hesitant due to past experiences with automation that was too inflexible. To counter this, we made sure the review interface was easy to understand. Users could see the original message, the translation, the extracted ETD, the confidence score, and the reasoning cues. As a result, trust started to build because the system was clear enough for quick reviews.
That’s why we always emphasize that agentic AI supply chain projects in Saudi Arabia thrive only when the system can navigate both messy data and foster human trust.
The results we saw were impressive, but what really stood out was the boost in operational stability.
We saw ETD data accuracy jump from about 85% to an incredible 97.5%! This was significant because ETD wasn’t just a reporting field for our client; it played a crucial role in planning that influenced production schedules, supplier follow-ups, and inventory choices. When the quality of ETD improved, it naturally enhanced downstream planning as well.
We also noticed a big reduction in manual tracking time, dropping from 4-5 hours daily across three coordinators to just about 45 minutes a day for one person handling exceptions and special cases. The other two coordinators could then focus on more valuable tasks like negotiating with suppliers, discussing lead-time reductions, and pursuing cost optimization strategies. The client was clear that they didn’t want AI to take away the human element of procurement; instead, they aimed for it to eliminate repetitive tasks like reading, copying, and updating.
The speed of updates improved significantly more than the effort required. Before the new system, a supplier’s message could linger in an inbox or chat for hours before a coordinator got around to updating the tracker. But after we implemented the new system, the average time between the arrival of a supplier message and its update in the system dropped to about 12 minutes, which includes both AI extraction and human approval. This change allowed planning teams to work with more current information throughout the day instead of relying on the previous day’s understanding of the current status.
The key takeaway for our business outcome was this: production stoppages caused by material shortages dropped from three or four each quarter to none within six months of launching Phase 3. In a manufacturing environment, that’s more than just a process KPI; it’s a significant operational and financial achievement.
Additionally, the client estimated that improved predictability allowed them to free up around $2 million in working capital. With enhanced ETD confidence, they didn’t need to maintain the same level of safety stock to hedge against uncertainties. This positive cash impact was just as valuable to leadership as the savings on labor costs.
Supplier responsiveness saw a boost as well. Thanks to automated follow-up messages sent via the suppliers’ preferred channels, the response rate jumped to 85% within 24 hours, compared to about 50% with manual follow-ups. This really shifted the procurement team’s dynamics from a reactive approach to managing exceptions more effectively.
Oracle APEX data completeness has jumped from around 60% to an impressive 98%. This improvement is thanks to the continuous flow of ETD updates, eliminating the need for manual re-entry. As a result, management now enjoys a much more reliable planning perspective and has a solid record of ETD changes for each supplier and PO line.
Currently, the client is operating with a well-established AI agent workflow that integrates multi-channel extraction, supervised approval, system synchronization, predictive ETD buffering, and automated escalation. For manufacturers in Saudi Arabia overseeing global supplier networks, this shift makes all the difference between chaotic communication and structured, clear visibility.
The solution utilized Python with FastAPI to handle the backend services, while Celery and Redis took care of asynchronous message processing. For email intake, we leveraged the Microsoft Graph API, complemented by the WhatsApp Business API and a WeChat bridge for messaging ingestion. We used Claude Sonnet for multilingual ETD extraction and date normalization. PostgreSQL was chosen for structured storage, and React.js powered the review dashboard. For internal integration, Oracle APEX REST APIs played a crucial role, and we relied on AWS EC2, RDS, S3, and Lambda for our infrastructure, storage, and event-driven processing. Altogether, these components formed a practical agentic AI system aimed at enhancing supplier communication management within the manufacturing landscape of the Middle East.
The most important takeaway was that we should prioritize channels based on urgency rather than convenience.
Initially, we opted for email since it was the technically cleaner choice and had the highest volume. While that was a smart technical move, it wasn’t the best operational decision. This became evident when the client faced a production delay due to a time-sensitive WhatsApp message that lingered in the manual workflow for six hours before being seen. This experience reshaped our approach to future projects.
Now, as we scope a plan for an AI agent for communication in multi-channel procurement, we make sure to ask two distinct questions: which channel receives the most volume of messages, and which channel has the most urgent information? These two factors often don’t align.
We’ve also discovered that understanding approximate language requires a bit of business context, not just NLP on its own. Phrases like “after Chinese New Year” or “mid-April” serve as planning signals influenced by supplier habits, geography, holidays, and the complexity of production. That’s why a robust supplier management AI agent needs to blend language comprehension with operational guidelines and historical patterns.
Lastly, we found that the design of reviews is just as crucial as the accuracy of extractions. Users tend to embrace these systems more quickly when they can see what the model processed, what conclusions it reached, and where there’s still some uncertainty. At aTeam Soft Solutions, this insight has become one of our key design principles for any AI agent involved in procurement, supply planning, or regulated operational workflows.
At aTeam Soft Solutions, we specialize in creating user-friendly systems for teams overwhelmed by communication-heavy operational tasks across areas like procurement, manufacturing, distribution, and enterprise planning. If your business in Saudi Arabia, the UAE, or the broader Middle East is still keeping track of supplier updates through emails, chat threads, and spreadsheets, there’s usually a more efficient way to organize that workflow.
We don’t start by asking your suppliers to change their communication style. Instead, we utilize the channels they’re already familiar with, address the exceptions your team handles, and build upon the internal systems your planners rely on. From there, we develop an AI agent that starts supervised, demonstrates its accuracy, and gradually takes over repetitive tasks while ensuring human judgment remains in place where it still counts.
