Until this summer, the morning numbers at LDC Import-Export Co., Ltd. were built by hand.
LDC is a Lao distribution and retail group. It started in Vientiane in 2011, distributing local and imported goods, and its retail arm now runs 39 convenience stores across seven provinces, including Jiffy shops on PTT fuel forecourts and its own Sky Mart chain. The ERP behind those stores keeps the business in 52 databases that the company could only read through manual spreadsheet exports. Every morning the business-analysis team spent hours assembling the Excel report packs from those exports. The cost of empty shelves across the chain had never been put in money, and a question like “which items sold at zero cost last week” had no one to ask.
I’m the CEO and co-founder of QUEST EDTECH, and LDC is one of its clients. I took the work on because LDC was losing money it couldn’t see, and its analysts spent the first hours of every morning assembling numbers instead of using them.
Since May, my company has rebuilt that layer with AI coding agents working under my direction. A chain now restores the ERP at 4:45am each morning, rebuilds a warehouse and emails the packs by 7:45am with nobody awake, and a check at one in the afternoon ties 160,000 cells back to the source system. The analysts got their one to two hours a morning back, several hundred hours a year.
The first replay of the year found about US$570,000 of sales lost to empty shelves, more than half of it while the item sat in LDC’s own distribution centre, plus a duplicate-sales event the ERP had hidden. The AI running cost LDC carries for all of it is about US$100 a month. Five department files that staff used to compile by hand now rebuild themselves every night.
Before this year, the business-analysis team assembled the morning packs by hand, with Power Query pulls, SUMIFS across every branch and copy-paste from the ERP. When I proposed live dashboards instead, the answer was that the analysts need yesterday’s number in the format they know, and that stays.
So the format was fixed before a line of code existed and the daily reports are the original manuscript. Agents rebuilt the packs as codes, by mid-July they had loaded the ERP estate into a warehouse of about 60 million rows, the nightly chain has run unattended since 24 August via automation.
QUEST EDTECH realised that three things had to be fixed before the output was usable, and none of them were codes. The first fix was definitions, because the “online sales” column in one pack had been carrying wholesale accounts and ran several times too high. The definition that fixed it came from the analysts’ own requirements document, because the ERP codes credit returns differently from cash ones and the second code had never been filtered. Both fixes came from asking the people who own the numbers, and that’s where I’d start next time.
The second fix was time, after a date computed on a server that runs on UTC produced a pack with the wrong day’s data. The rule now is to compute a Bangkok day on a box set to Bangkok time. The third fix was a rule for the agents themselves, using the semantic layer derived from the database as the fact, then write the code. Agents read the database’s own schema before they name a column.
Also Read: AI agents broke into 100 firms for pocket change. SEA’s online shops fit the target profile
When the AI was confidently wrong about LDC’s numbers
Once nobody was assembling the packs by hand, nobody was looking at each number on its way to LDC either. In August every target-achievement figure in one pack read exactly 100 percent, which should’ve been the first clue. LDC’s target workbook is a rolling file, and when a month closes, its target is overwritten with the actual result.
The agents couldn’t know that, so they loaded actuals as targets and the pack was checking itself. What caught it was a test that rebuilt the targets from the sales data and found the two were identical. That same month I gave the agents the rule I’d give any team verify the fact, then write the code.
What AI informs but can’t decide
Before I could put the cost of LDC’s empty shelves in money, someone had to decide what an empty shelf is. Four defensible definitions of out-of-stock on the same data gave 0.6, 10.1, 67.2 or 77.3 percent. The agent could compute all four in minutes, but it couldn’t tell me which one the business should manage. I settled it with the rule the buyers already work to, from the supply-chain team’s own workbook. The US$570,000 of lost sales is counted on that rule’s zero-stock test.
What reaches whom is a human call too, and the AI “chief strategy officer” I set up to write weekly briefs for LDC’s CEO showed why. Its first version flagged four branches as the week’s news, and all four were already red in the pack the analysts read every morning. I rebuilt it as a decision sheet with an owner, a date and the amount at stake on each line, and who gets it is my call.
Also Read: Tokens are the wrong yardstick for AI agents, says JFrog’s Sunny Rao
What stopped, and what it has been worth
The analysts stopped building the packs, and five department spreadsheets that staff uploaded by hand were retired once a system source replaced them. I stopped waking up for failures too. In September I told the agents that when something breaks between 4 am and 9 am I need them to fix it, not notify me, since silicon doesn’t need sleep, and a DevOps agent now runs four checks a day and re-runs known-safe steps itself.
The scale of what agents build is the part I’d have doubted a year ago: about 240 dashboard cards in one three-day stretch in September, at roughly 4.1 minutes of my time per card, and 766 saved questions and 198 dashboards on the BI site against the client’s requirement of 76 reports.
The honest way to count the money is in what the system has found, because realised savings are LDC’s to report. With AI, LDC has identified more than half was lost while the distribution centre had the item in stock. Annualised, that’s roughly US$770,000, which is why the next phase is a replenishment fix rather than another report.
Who does what and the outcome
On QUEST EDTECH’s side the agents work in lanes, with Claude Code building and operating. The bigger change is on LDC’s side, where the analysts got Claude Desktop skills that turn a question in Thai into SQL. As of October, the BA lead added her own card to the finance dashboard, written through that skill with AI in Thais and Laos, and the build was changed so her cards survive every rebuild. The people who used to assemble the numbers manually now ask questions to AI and it just builds reports, dashboard and identifies with simple languages instead of power query or even a line of SQL.
—
Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.
The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.
Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.
The post How a 39-store retail chain in Laos moved from Excel to AI agents appeared first on e27.