A major trading and distribution company in Dubai was handling over 8,000 supplier invoices per month in email, scanned mail, supplier portals, and WhatsApp. Their accounts payable team wasn’t failing because it was undisciplined. They weren’t failing because they were too busy or lazing around; they were failing because the workflow, for the volume they were doing, was just too broken up, too manual, and too slow. Eight AP clerks went about their days opening invoice files, determining the supplier, matching invoices to purchase orders, verifying goods receipts, inputting data into SAP, routing approvals, and tracking down exceptions. Even simple invoices required 12-18 minutes. Problem invoices could take an hour. The average time it took to go through the cycle from getting an invoice to entering it into SAP was 8-12 days.
This delay led to truly high financial costs. The company was losing early payment discounts, incurring late payment penalties, processing duplicate invoices, and damaging supplier relationships throughout the UAE and Saudi Arabia.
At aTeam Soft Solutions, we developed an AI agent for invoice intake, intelligent extraction, three-way matching, SAP entry, approval routing, and supplier communications. The solution used OCR, document understanding, business rules, and human review if necessary. The result was a transition from manually processing AP to exception-based finance operations: 99. 2% matching accuracy, 72% invoices processed without human touch, 75%+ reduction in manual efforts, and estimated annual savings of approximately $650,000.
When we analyzed the client’s business process, we realized almost immediately that the problem with invoicing was not starting in SAP. It was created long before SAP ever got to see the invoice.
Invoices poured in from all directions. Roughly half landed via email as PDF attachments. Another quarter arrived by physical mail, which was then scanned by the mailroom into a shared folder. Approximately 15% were obtained through supplier portals that AP staff needed to access and manually download. And the last 10% of those documents came in via WhatsApp, often as phone photos, scans of the images, or documents shared by local firm suppliers.
That meant there was no real intake layer. There were simply channels.
Each day, AP clerks began by opening mailbox folders, accessing scanner directories, logging into vendors’ portals, and monitoring messages on WhatsApp that sometimes included invoices. They had to find out whether a document was even an invoice before they could run one through. Quotes, delivery notes, statements, and advertising material were also regularly dispatched through the same channels by suppliers. Someone would need to open that attachment, read it, and determine if it needed to be part of the AP flow.
Once an invoice was found, the real work began. A clerk would read the invoice, identify the supplier, find the applicable supplier code in SAP, search for one or more purchase order numbers, match line items, quantities, and prices, ascertain whether goods were received, confirm the currency, calculate or check VAT, and then he would post the invoice manually into SAP. If the invoice tied up nicely with the PO and the receipt had already been posted, the entry was pretty simple. But even then, it still took 12-18 minutes because it required going through several screens, verifying against references, and filing the documents.
The real challenge was the chaotic cases that were prevalent enough to fill up the daily workload.
Some suppliers issue one invoice for multiple purchase orders. These POs were often for different departments and had different approval routing. Some invoices contained revised supplier pricing, while the original PO continued to display the old price. Some invoices were received prior to the goods receipt note being posted by the warehouse, preventing AP from executing the three-way match, even though the goods might have been physically delivered. A few invoices were resent by the suppliers as they hadn’t received any confirmation, creating duplication risk. Global invoices were more complex, as they needed to be recorded using the right exchange rate at the date of the invoice and not the date of processing.
The AP team had no single way to handle this complexity. A clerk took care of the issue in front of them, and sometimes they exercised judgment informed by experience. If there were discrepancies, they emailed procurement, called the warehouse, asked finance for clarification, or held the invoice for later. The older the parked invoice was, the more likely it was that someone had to start from scratch to look at it because they’d lost the context.
This led to a long processing turnaround. The business was spending eight to 12 days just to get from receiving an invoice to entering it into SAP, on average. That meant the company’s 30-day supplier payment SLA was already under strain even before approval began. In practice, that only left them 18-22 days to get the payments approved, scheduled, and executed. Even with that, it was not enough runway for a company managing $150 million+ in annual procurement spend and engaging with over 400 suppliers across multiple countries.
There was another charge hidden and built into the delay: early-payment discounts. Most of the suppliers had terms like 2% if paid in 10 days. However, if the processing of the invoice itself took more than a week, these were days wasted for the finance team without any action. The client calculated they were failing to capture 2-3% of their available early pay discounts, equating to about $300,000-$500,000 a year.
In addition, the manual process also created relationship damage that never graduated neatly to an AP report. Suppliers called asking where the payments were. Procurement teams got dragged into invoice-status questions. AP staff had to waste their time responding to “Did you get this invoice?” or “Why is this still pending?” rather than on processing work. Finance functions were doing a lot of high-volume activity, but it was too much clerical catch-up work and not enough disciplined financial operations.
That’s where aTeam Soft Solutions came into the picture. The aim was not simply to speed up data entry. This was to redesign the entire AP process so that they only handled the invoices that really needed judgment.
The client had already attempted conventional approaches.
Once data was structured, SAP’s invoice processing ability was good, but it presupposed that the invoice had already been brought into the correct fields by someone or something. SAP couldn’t just glance at a scanned PDF, a WhatsApp photo, or a badly formatted supplier document and know what was on the page.
They also attempted an OCR-only solution with ABBYY FlexiCapture. It pulled out text, which is very different, like knowing what an invoice is. OCR can find numbers, dates, addresses, and tags. It doesn’t know which is the invoice total, which is the PO, which is the field for a bank account, or which line is a discount term. Even without a business context, the result was still a human having to interpret the page nearly line by line.
Template-based methods were also not well-suited. The invoices sent to this company were in English, Arabic, Chinese, and German, each with a different supplier layout, with varying image quality and formats. Creating and updating hundreds of templates for 400+ suppliers would have required a whole team to just keep the extraction system alive.
That was precisely the reason why the issue called for the invoice processing AI agent and not the scanning tool. The workflow called for a solution that could recognize document types, understand invoice layout, detect supplier intent, handle multi-PO references, detect discrepancies, and forward the invoice to the next business process. Or, to put it another way, this was more than just OCR. It was a genuine AI accounts payable automation problem with operational judgment at every stage built in
We engineered the solution as a phased rollout because AP is a financial control function. No self-respecting company should go straight from manual processing to full automation. The system had to gain the confidence of the AP teams, financial controllers, and external auditors.
Phase 1 has established a universal intake layer for all invoice streams. Email accounts were automatically watched. The mailroom scan folder is linked to the pipeline. Downloads from the supplier portal were automated using routines running in browsers. WhatsApp invoice submissions were directed to a unique business number, and they entered the same processing flow.
That was important, as the client didn’t really have “one AP process” before this. They had four different intake protocols. The AI agent unified them all into a singular, controlled intake stream.
Once in the pipeline, the system initially classified the document. Was it an invoice, a delivery note, a credit note, a quote, or just buzz? It sounds basic, but that eliminated a lot of wasted AP work. Clerks could stop manually opening everything just to know what they were dealing with.
Following categorization, the extraction pipeline was initiated. Running on Google Cloud Vision for OCR and Claude for document understanding, the AI agent reads the invoice in any language or quality of an image and extracts the supplier name, invoice number, invoice date, PO references, currency, VAT details, line items, quantities, unit prices, totals, payment terms, and bank details. For attachments to emails and for scanned documents, the original image was retained alongside the extracted data in the dashboard, so AP users could rapidly validate.
This is where the human-in-the-loop design made a difference. During Phase 1, the system did not post any messages automatically. AP clerks made a quick review of extracted fields and fixed errors in a few clicks. That transformed the AP team’s job from one in which they had to type manually to fast validation. A process that previously took 12-18 minutes became a brief review in action for most of the invoices.
The 2nd phase was to bring the AI agent out of reading invoices and into comparing them to business records.
We have integrated the system with SAP data for automated three-way matching: invoice versus PO versus goods receipt note. At the line-item level, the agent also validated that the billed quantity was less than or equal to the received quantity, that the price matched the PO within a predefined tolerance, that the description matched the expected item, that the PO was still active, and that the invoice fit within a purchasing context the company understood.
For clean matches, the invoice could be marked as “ready for SAP entry.” For the exceptions, the system did not simply say ”mismatch.” It explained the problem. If the invoice price was different from the PO price, the dashboard shows the quantity, the price gap, the total disruption impact, and if recent invoice histories indicated that the supplier was operating under the assumption of a new price while the PO was for an outdated one.
It came up a lot that the explanation layer mattered. AP teams don’t just need flags. They need enough context to know if they should fix it, approve it, hold it, or send it back. Proposed resolutions appearing to the AI agent based on past behavior shortened the resolution time for the recurrent categories of issues.
This is the phase in which the project began to feel like automated invoice matching AI, not document capture.
After extraction and three-way matching were robust enough, we enabled the AI agent to automatically generate SAP entries for matched and validated invoices.
The system generated the invoice, paired it with the source document, connected it to the associated PO, and submitted it for the company’s approval matrix. Different amounts, cost centers, departments, and invoice types triggered different chains of approvers, and agents handle them according to that routing logic consistently.
We also created a React Native mobile app for approvers. Without having to ask managers to sign into a desktop workflow and open up multiple tabs, they now had a single-screen summary of what they were approving, how it matched, the impact to the budget, and whether an early payment discount was available. If discount capture was on the line, the invoice was visibly routed faster and prioritized.
The system could also generate supplier communications with detailed reasons for rejections for invoices that were disputed or rejected. That took away the uncertainty, the cause of invoices being sent over and over again by suppliers.
The final stage took the solution out of invoice entry and into the management of AP operations.
We added invoice remittance advice automation so that the suppliers received advice of payments automatically with no AP manual work. We created a supplier self-service portal, which allowed vendors to check on invoice status directly, leading to a significant reduction in inbound “where is my payment?” queries.
We also layered in a cash flow and discount optimization. The AI agent scanned outstanding invoices, discount windows, due dates, and availability of cash to suggest payment schedules that would maximize discount capture without causing treasury strain. That was an important step, because speeding up invoice processing only generates real financial value if the company takes the additional time and uses it to make smarter payment decisions.
Duplicate invoice detection was another big boost. The system scanned for same-supplier, similar-amount, near-date combinations, and tagged potential duplicates before they were sent to payment. This saved the client from one of the most costly AP errors: paying the same invoice twice because a resubmitted document appeared to be a new one.
For aTeam Soft Solutions, this is what changed the project from a capture tool into a true AI agent for AP operations spanning Dubai, the UAE, and supplier networks extending into Saudi Arabia and beyond.
The core system was developed in Python with FastAPI, and asynchronous tasks were managed with Celery and Redis. This means invoices from multiple channels could be ingested, classified, extracted, matched, and routed without bottlenecking during high-volume periods.
Google Cloud Vision for OCR, as it demonstrated good support for Arabic, along with a range of scan qualities of scans. Claude also did the contextual parsing layer: parsing supplier identity, finding invoice references, telling relevant totals from non-numbers, and parsing document structure in ways rule-based OCR never could.
All invoice data, extraction results, matching results, exceptions, and audit events were stored in PostgreSQL. Source documents with attachments were stored in S3. Using the React-built dashboard, AP teams had a practical working interface for review, exception handling, and visibility into status, and the mobile app provided approvers with an easy way to expedite decisions on outstanding items.
SAP integration was achieved by using the SAP Business One Service Layer APIs and the DI API for your access. This was not just a basic write-back. The AI agent had to be able to create invoice records, attach supporting documentation, correctly split multi-PO invoices, allocate shared charges on a proportional basis where business rules mandated it, and adhere to the company’s approval path and control the boundaries.
The WhatsApp Business API was embedded in the intake layer, enabling invoice photos and local supplier messages to be handled within the same application rather than as side-channel documents.
We also executed strict confidence and fallback controls. The confidence of the system in a critical field was under threshold, especially for the handwritten invoices, so the document was sent for manual review. If the matching algorithms can’t cleanly resolve an invoice, it is flagged for a clear reason. In case of SAP posting error, the invoice remains in a retry or exception state with full traceability and does not just vanish in a silent error.
Finance automation is not like content or CRM automation, and at aTeam Soft Solutions, we get that. Every extraction, every match decision, every approval route, and every posting action should be auditable. That was the principle that the whole design was based on.
Multi-PO invoices were a really challenging practical problem. And ~25% of invoices cited more than a single purchase order, and those POs, at least more often than not, belonged to different departments. The AI agent had to split the invoice into multiple SAP postings while maintaining the reference to the original invoice document. Common costs, like shipping or insurance, had to be distributed proportionally among POs based on the client’s accounting rules. This was not a document processing problem alone. It was the business logic that was the problem.
Dealing with currency was another important aspect. International suppliers invoiced in USD, EUR, CNY, and GBP. The company’s policy was to convert at an invoice date, not a processing date, to AED. We tied exchange-rate information sources with date-specific querying and fallback to official rates so the system could interpret international invoices correctly and in a uniform manner.
Handwritten invoices from smaller local suppliers in the UAE introduced yet another level of complexity. Arabic handwriting is much less predictable for OCR processing than printed text. These accounted for only about 5% of the total invoices, but due to them, we had about 30% of the remaining cases of manual review. Instead of distilling them too much, we selected confidence thresholds. If a significant field had less than 85% confidence, the invoice was sent for total review. That maintained the risk at a low level and at the same time allowed for the rest of the processing to go on at a fast pace.
But the largest non-engineering challenge was in governance above. The AP clerks weren’t the most skeptical users. The financial controller and external auditors were. They wanted to know that an AI-powered agent inputting invoices in SAP could be audited more closely than a single human clerk. This meant that every action had to be logged with the rationale, confidence, source evidence, and policy state of approval.
This learning became the core of how aTeam Soft Solutions now handles agentic AI finance Dubai projects. Auditor confidence isn’t a final checkbox. It is a design feature needed from the start.
The first big enhancement was speed. The average processing time for an invoice was reduced from 8-12 days to approximately 1.5 days. That one thing changed the entire economics of the AP function because now the business had time to capture discounts, move approvals properly, and not be pressured into paying at the last minute.
The manual effort per invoice also fell significantly. What used to require 12-18 minutes of manual processing now takes less than two minutes to validate on the majority of audited invoices. More significantly, is that 72 percent of invoices were turned into no-touch transactions and passed through without any human interference. The AP team was no longer required to devote the bulk of its energy to routine entry and matching.
Three-way matching accuracy improved to 99.2%. That gave the finance team confidence that matched invoices were truly good to post, and that exceptions were being surfaced with enough details to address them quickly.
The staffing impact was dramatic. The AP team went from eight clerks to three people whose main roles were dealing with exceptions, coordinating with suppliers, and conducting control reviews. Instead of routine data processing, the five team members were redeployed into vendor management and financial analysis. It matched with the client’s larger objective: employ automation to lift finance personnel to high-value-added tasks, rather than just cut the headcount blindly.
Early payment discount capture increased from approximately 15% to 78%, resulting in approximately $380,000 of recovered annual value. That was a particularly important result because it showed the system was doing more than just saving labor. It was having a positive effect on the cash economy.
The duplicate invoice detection layer flagged 45 duplicate invoices within the first six months, mitigating over $127,000 of risk of double payment risk. Late payment penalties dropped 85%. Supplier inquiry volume decreased by 60% as vendors could check the status via the self-service portal rather than repeatedly calling or emailing AP.
Collectively, the labor savings, discount recovery, penalty reduction, and duplicate prevention resulted in an estimated annual saving of around $650,000.
For a company processing thousands of invoices every month through the Middle East supply network, those were not just judicious results. Day to day, they altered the finance operation.
This solution utilized Python with FastAPI for backend orchestration, Celery and Redis tracking for processing asynchronously, Google Cloud Vision for OCR, Claude for understanding the invoice and the matching intelligence, SAP Business One Service Layer API and DI API for ERP integration, WhatsApp Business API for intake, React.js for the AP dashboard, React Native for the approver mobile app, PostgreSQL for structured data and audit storage, and AWS EC2, RDS, S3, and Lambda for infrastructure, storage, and event-driven processing.
The largest takeaway was that financial AI is successful or not based on auditability, not just accuracy.
We anticipated that the AP team would be the hardest to convince. They came on board quickly, in fact, because it took out the most repetitive parts of their day-to-day work. The more profound resistance came from controllers and auditors, who had to be assured that an invoice processed by AI could be traced as easily as one processed by hand manually.
That changed our approach in terms of implementation. We will now design the audit trail as a first-class product feature within every finance workflow. Any extraction, matching result, discrepancy flag, approval routing, monetary conversion, or SAP posting decision must be explainable, contain a timestamp, and link back to the source document. During this project engagement, the external auditors ended up telling the client that the AI-created audit trail was more consistent than the paper trail it replaced.
We also realized that the real value in AP automation comes after capture. Faster intake is, of course, a good thing, but the actual gains were from three-way matching, speeding approval, detecting duplicates, and optimizing discounts. This is why aTeam Soft Solutions now considers AP automation software development in India projects as a complete workflow redesign process rather than OCR implementation with a dashboard on top.
If your finance team in Dubai, UAE, Saudi Arabia, or the larger Middle East is still manually opening invoices, keying data into ERP, chasing approvals, and responding to vendor status inquiries, the problem often isn’t just labor cost. At the heart of it are cycle time, controlling the quality, missed discounts, and unmanaged friction between procurement, warehouse, and finance.
At aTeam Soft Solutions, we design and develop practical AP systems that leverage an AI agent to read invoices, validate them, match them, route them, and escalate only what really needs human judgment. That’s really when finance teams begin to transition from document-centric handling to decision-centric operations.
