Bought One Software After Another, But Is the Accounting Still Done by People? Try Hiring an FDE

FDE stands for Forward Deployed Engineer. Our approach is to bring AI tools directly into the client’s company to work. Before going there, I thought the finance team at the clothing company was busy every day making vouchers and generating reports. After two days, I realized they spent most of their time reconciling accounts.

1. Accounts Split Across Two Systems

This company uses two systems:

• Multi-Customer Management for sales, order processing, cash register, and customer interactions;

• Clothing Management for supply chain, procurement, production, and factory reconciliation.

When the client explained the finance team’s work, they were very straightforward: “We actually have three types of money: factory money, fabric money, and daily expenses.”

The problem is that the two systems do not communicate. Multi-Customer shows that a batch of goods was sold, but only half was recorded in Clothing Management. To find the discrepancies, finance has to export the tables from both sides and compare them line by line. When we arrived, there were millions hanging in the system that had not been reconciled.

2. Daily Morning Payment Reconciliation

Every day, orders are processed and payments are collected through WeChat and Alipay in Multi-Customer. Finance needs to reconcile three things: channel bills, payment flows, and sales documents. There are no common transaction numbers between the two systems, so they can only match by amount; the same amount can appear multiple times in a day, and customer names may not match. Additionally, with returns, combined cash register entries, and cross-day accounting, it is common for manual reconciliation to take a whole morning.

Now, this task is handled by an AI assistant. Here are the results from one day:

Figure 1｜AI Assistant Completed Daily Payment Reconciliation (Example)

On that day, there were dozens of bills, but only 8 required manual intervention from finance, each with noted issues and suggested actions, such as suspected omissions, requests for return vouchers, or verification in the order backend. A few entries that seemed mismatched were actually due to differences in statistical criteria or accounting times, which the AI also explained. Previously, these types of judgments were the most labor-intensive.

In the generated summary table, green indicates matches, yellow indicates discrepancies in criteria, and red indicates items that need to be addressed. Now, finance opens the table in the morning and starts checking from the red entries:

Figure 2｜Excerpt from the Triparty Reconciliation Summary Table (Example)

3. Monthly Factory Reconciliation: A Few Days of Work in 45 Minutes

The finance team’s requirement is to reconcile all factory bills at the beginning of each month. Previously, they had to go through the system line by line, copying quantities, multiplying by unit prices, and reconciling each factory bill one by one. One factory could take an afternoon, and several factories could take several days.

Now, for factory bills, we let the AI assistant directly pull the product data for the month from the system and reconcile it with the Excel sheets in finance’s hands, completing all the bills in less than 45 minutes:

Figure 3｜AI Assistant Reconciling Factory Bills (Example)

The report separates consistent entries from those with discrepancies, highlighting those with significant differences. It even identified an item that was split between two factories, which needed to be combined to match. Ultimately, finance still needs to confirm with the factory, but they know exactly which items to ask about and what to inquire after receiving the report.

4. Two Systems Automatically Reconcile Daily

The millions of amounts pending reconciliation used to rely on finance’s experience and last-minute efforts at the end of the month. Now, the AI collects data from both sides daily, automatically comparing inventory, stock, and cost records, generating a discrepancy report, with abnormal orders entering a pending verification list the same day.

Figure 4｜Daily Automatic Reconciliation Records of Two Systems (Example)

This process has been running continuously for over two months. The amounts pending reconciliation no longer have to wait until the end of the month to be addressed; they are processed in small batches daily. Daily clearing and weekly reconciliation is a standard proposed by finance themselves.

In addition to these tasks, monthly costs, customer payments, and first-order commissions, which originally required manual spreadsheet compilation, are now also automatically generated by AI each month, with finance only needing to review.

Figure 5｜Excerpt from First Order 30-Day Sales Statistics (Example)

5. How These Rules Were Given to AI

During the first week on-site, I basically sat next to the finance team, watching them reconcile accounts. Whenever I saw a mismatch, I would ask why. If the money collected via bank card was not in the WeChat bill, was that a mistake? If a code scanned at midnight was recorded the next day, should it be accounted for? The finance team answered immediately, but none of this was ever documented; it was all in their heads.

What we did was document these answers one by one and organize them into skills. You can think of it as an operations manual written for AI, which follows it daily. If the rules change, we just update the manual without needing to redevelop.

Figure 6｜Daily Reconciliation Script Called by Skill: Read Data, Check Differences, Write Reports

AI connects to the Multi-Customer and Clothing Management systems through MCP, only observing and not modifying, and it does not handle account passwords. It collects data twice a day, at 8:30 AM and 6:30 PM, and if it loses connection, it stops and alerts someone.

Figure 7｜Scheduled Task List in AI Assistant (Example)

Later, the production team also inquired whether they could be informed daily about which orders were cut but not shipped. The same rules were asked one by one and added to the manual. Now, if someone wants data, they just type “Generate Production Daily Report” in the chat box, and the report comes out.

Figure 8｜Requesting a Production Daily Report in the Chat Box (Example)

6. “If You Can Help Me Solve 70-80%, I’ll Be Very Happy”

At the end of the requirements meeting, the finance manager said, “If you can help me solve 70-80%, I’ll be very happy.”

Finance does not expect a fully automated system. AI handles what it can automatically, and any issues are clearly documented for manual verification.

During the review after a month on-site, finance no longer had to work overtime until 8 or 9 PM at the end of the month; they only needed to spend about ten minutes each morning reviewing the discrepancy list. Previously, when the boss asked, “Which items made money this month?” they had to wait for the end-of-month report; now they can see it by mid-month. Any suspected missed payments or orders with half the goods missing are identified on the same day they occur, which translates to real money.

The entire process did not involve purchasing new software, changing systems, or adding personnel. What we did was connect the existing data in the systems and let it run automatically every day. From the first requirements meeting to daily automatic reconciliation, it took less than a month.

Final Thoughts

Many companies face similar situations: multiple systems each managing a part, with data being manually transferred in Excel, leading to a full reconciliation at the end of the month. Software has been purchased repeatedly, yet the problems persist. What is often lacking is someone sitting down to clarify the rules in the business one by one and then handing them over to machines for daily execution.

What FDE does is exactly this. If your company is also trapped by reconciliation, data transfer, and end-of-month overtime, feel free to discuss in the comments. The next on-site note may just be about your industry.