Two Weeks as an FDE at a Foreign-Trade Company: Look at the Stars, Stand on the Ground

Enterprise AI is easy to start with big words.

A negotiation copilot, a company knowledge base, customer insight, intelligent decision-making — each phrase looks like a page from a long-range plan. Look up, and AI can go a very long way. Walk into an actual foreign-trade company, and the first thing you meet is a stack of Excel and Word files.

Sales fill in the shipping coordination sheet, procurement adds lead times, carton counts and weights, and the documentation clerk moves that same information into the booking request, the customs paperwork and the clearance documents. The customer name gets copied once, the port gets copied once, gross and net weight get copied again. The company has two documentation clerks; one of them also handles admin and HR, and the other one's calendar is already full.

AI entered this company and first took over the copying. The tables are small, but the chain behind them is long. Sales, procurement, documentation and management are all strung along it. The plan that looks at the stars has to start from the one cell under your feet.

It does not sound grand, but it makes a good opening for an FDE project. The rules are clear and the results are easy to check. Whether the generated file is right is something the clerk can tell at a glance. The more tempting ideas can stay on paper for now. What a company is usually shortest on is human attention — save some of that first, and everything else becomes discussable later.

A documentation clerk's job does not live inside the spreadsheet

Only after following one order all the way down did we see how much is packed into the phrase “doing documents.”

Filling in forms is one segment of it. As the goods near completion, sales confirm the destination port and mode of transport, procurement reports the lead time, and documentation contacts the forwarder to book space, then handles trucking, warehouse delivery, customs declaration, the bill of lading and the final payment. Change one date and everything downstream moves. The cells in Excel are quiet; the people behind the cells never stop talking.

So the system only takes over the fields whose relationships are stable. The shipping coordination sheet, the invoice and the packing list go in as input; the program reads customer, goods, packaging and transport information and writes it into different templates. Coordination and exceptions stay with the clerk.

Where you make that cut matters. Frame automation by job title and the conversation runs straight to “will AI replace people.” Break it down by action and it becomes much plainer. Fixed information transfer goes to the machine; the parts that carry responsibility and on-the-spot judgment stay with the person. A role that looked like a tangled ball of yarn turns out to have a loose end once you follow the process.

The demo went smoothly; the rollout got lively

Once the core path was running, we were somewhat optimistic. Files went in, files came out, and the fields landed where they should. But software develops a personality the moment it leaves the developer's laptop.

One employee's run went four hours with no result. Old .doc files behave differently from .docx. An order with multiple product categories produced garbled text in the merged template. A file would be edited and saved, and the agent would occasionally still cling to the old content. The output even kept its red font — a certain candor, as if to say “I finished, but I also want you to know something went wrong here.”

Most of these problems have nothing to do with model intelligence. File format, machine environment, template, cache, toolchain — let any one of them act up and the user sees the same sentence: AI is broken again.

In a demo you verify whether one capability works. When employees bring real orders, what gets tested is the whole system: whether the input files follow the spec, whether legacy formats are supported, how you fall back after a failure, who checks the result. The model is only the actor on stage; if the curtain, the lighting or the teleprompter breaks, the audience still asks for a refund.

We later settled on a very practical rule: if there is no result after twenty minutes, do the document by hand and don't hold up the shipment. Keep the problem file and the environment details, and reproduce it afterwards.

The rule is not elegant, but it leaves the business a way out. Orders have sailing dates, and customers will not wait two extra days because we are exploring AI. Let the system fail, and don't let the business stop — that is roughly the first lesson production teaches idealism.

Some problems are better deleted than solved

The orders used to include a column of product images, and the agent often stumbled on it. The first instinct was natural: should we add multimodality and teach the AI to read pictures?

We kept asking, and found that the documentation flow does not use those images at all. The shipping-mark information that actually matters had already been turned into text by the sales rep. The final solution was simple: delete the image column.

When an engineer sees an anomaly, the hand reaches for a new feature. The field often has a different answer: remove the unnecessary input and the problem disappears. Technical capability is like kitchenware — buying another appliance is fun, but if the counter is already covered, clearing the table helps more.

The experience also changed how I think about AI product interfaces. The company already runs on WeCom, WPS and Excel; employees work inside those tools every day. The documentation system had no need to build another impressive chat window. An employee drops in the files they already know, the system fills them out, and a person checks a few key fields. That path is short enough.

AI Native is often imagined as everyone talking to an agent. After going into a company, I care more about something else: can employees learn less and change less of what they already do? If every user has to learn prompt writing first, the product still owes them a stretch of road.

One documentation chain brings the whole company into agents

Looking only at the final artifact, the system resembles form-filling automation. Follow the data one more time and the picture gets much larger.

Documentation sits at the intersection of the foreign-trade process. Sales provide customer, port and transport information; procurement adds lead time, cartons and weight; the clerk checks the files and talks to the forwarder; management watches whether the shipment goes out smoothly. A single documentation chain already spans sales, procurement, documentation and management — most of the company's key roles.

So the agent serves as two kinds of interface here. Technically, it catches files and data from different departments and produces results under one set of rules. Organizationally, it gets different roles handing work to AI in the same way. Who provides which input, who checks which fields, who takes over when the system fails — these agreements are becoming an early human–agent collaboration protocol.

That protocol becomes the foundation for the company's move toward an AI Native Company, or ANC — AI genuinely entering daily process, data and organizational collaboration. Whatever comes next, a negotiation copilot, customer analysis, or A2A letting multiple agents exchange tasks and state, all of it lands on the same base: the people in the company are already used to handing work to an agent. The documentation system happens to lay down the technical interface and the human habit together.

An employee gives files to the agent, waits, then checks the output. If something is off they say so; if it truly will not run they go back to the manual process. These few actions look simple, but they contain a whole set of capabilities: task handoff, result verification, exception feedback and human takeover. However complex the later copilots, customer analysis and knowledge bases get, they cannot route around these basics.

Trust accumulates through the same process. It works like a small ledger: every document completed successfully adds a little to the balance. When the agent gets something wrong, as long as the cause can be found and the business has a fallback, the trust saved up earlier is not spent all at once.

Once employees are used to this kind of handoff, more complex agents no longer have to explain themselves from zero. A negotiation copilot can reuse the existing data entry points, permissions and feedback channels, and multi-agent collaboration already has a shared user base. What the company really gets, besides a documentation tool, is a group of people who can use agents and are willing to hand them real work.

These few spreadsheets underfoot are paving the road for the more distant, more complex features. Before enterprise AI goes into deep water, people have to learn to get in the water first.

The group chat opened, and the real requirements showed up

After go-live we created a WeCom group with sales, procurement, documentation and management in it. The moment the group existed, the problems came through the gate.

One machine cannot run it — an environment problem. Garbled template — a system defect. A redundant image column — an input-spec change. An occasional unusual order — maybe it is cheaper to keep it manual. Coming out of a user's mouth, all of them are one sentence: “this doesn't work.” The FDE's job is to take that sentence apart and find which layer the problem lives on.

People are more complex than software.

The boss wants to see what the investment bought. The business lead wants clear, verifiable outcomes. The internal AI champion worries the project will lose continuity. Frontline staff care whether today's documents get done and who is responsible when something goes wrong. Everyone is reasonable, and put together they pull against each other.

When the system works, the saved half hour is quiet. When it fails once, “AI has a problem” travels across the office immediately. AI can even become a shield, hiding operational, environmental and process problems behind it. So the feedback group is not only for collecting bugs. Who describes the problem, when to fall back to manual, which files to keep for reproduction, who decides the fix priority — those rules are part of the system too.

Whether an employee is willing to use it has little to do with how progressive their attitude is. They pay a learning cost and carry the risk of failure; if the benefit is invisible, going back to the manual process is perfectly normal. Landing technology is a bit like hosting dinner: good food matters, but you cannot make the guests wash the vegetables, cook, and scrub the pans afterwards.

How I understand FDE now

Two weeks ago I thought an FDE was an engineer who understood the business and wrote code quickly. That is not wrong, only a little too clean.

The field is not clean. Old files, new templates, employee habits, permissions, sailing dates, the boss's expectations, and one machine that inexplicably will not run — all of it is crowded onto the same table. The FDE sits at that table and separates the pile piece by piece.

The boss says “efficiency,” and you have to ask where the time is saved. An employee says “it's not usable,” and you have to separate environment, input, rules and model. An engineer sees a bug, and still has to judge whether it is worth fixing. The system is live, so failure needs a door left open for it.

The work is part translation, part housekeeping. You do write code, and you do read spreadsheets, but most of the time you are running errands between several languages. Run them long enough and vague wishes turn into process, tangled problems sort themselves into layers, and AI can finally hold a stable position inside the company.

Plenty is still unfinished on this project. The documentation system is still collecting issues, the procurement database is waiting to be organized, and the business knowledge base is further out. What can be confirmed right now is small: a batch of forms is being filled by the system, employees have handed it real orders, and there is a way back when it fails.

Small, but it carries more weight than a hundred successes on a demo stage.

That is roughly what AI entering a company looks like. Watch the stars, certainly — and also fill in the cells underfoot, one at a time and correctly. AI first finds a place inside a single spreadsheet, then slowly learns to work alongside a group of people.