I set an anchor for myself: don’t replace an entire department — replace 10% of one business line, and call 500k RMB in annual revenue a success.

That number turns “helping businesses do the math” from an empty slogan into something verifiable. It’s not big, but it’s concrete — concrete enough for me to judge whether a solution is worth building.

The stack I can put together: n8n (self-hosted workflows), multimodal model calls (vision and OCR), an API relay station (its compliance status is unresolved, so it can’t go to production), and a machine that can run things. I used this stack to research one question: in traditional industries, those business lines where “10 to 20 people do data entry, slowly, with errors” — how much of that can I take?

A veteran in enterprise services pointed me to three directions — HR recruiting, asset management data entry, and financial verification — and gave me the underlying logic of interest alignment: irreplaceable, non-conflicting, information asymmetry. I researched those three plus six more, running 10 agents in parallel (8 research threads plus 2 synthesis threads), with search cross-validated across two engines, Tavily and Metaso.

First, a bucket of cold water

“One machine working for an hour matches what 3 to 5 employees do in a day” — I love saying this to business owners because it settles the math in one sentence. But this time I seriously stress-tested it, and it doesn’t quite hold.

The synthesis agent in charge of finding flaws flagged it: “rules of thumb, not independently verified.” In real scenarios, you’re not recognizing the clean PDFs from a demo — you’re recognizing asset labels photographed by employees on their phones: tilted, blurry, reflective. OCR error rates under ideal conditions are below 0.5%; real-world scenarios are far worse. The actual replacement ratio may drop to 1:1 or 1:2, not 1:3–5. The math behind that pitch would be shaken.

So this post isn’t hype. I’m writing down what can be built in half a day, how compliance can be resolved, and how to find channels — and I’m also flagging every place it could blow up.

Nine directions, one table

I put nine directions into a table and scored them on four dimensions (1 to 10): whether a prototype can be built in half a day, 500k RMB/year potential, compliance ease, and middleman reachability.

First place isn’t HR recruiting — it’s asset management entry. Three things work in my favor: a prototype runs in half a day (standard OCR plus multimodal, no complex system integration), compliance is lightest (business operating data isn’t personal privacy — it just has to stay within the country), and official AI competition is weak (Kingdee and Yonyou’s official AI focuses on invoicing and tax, and doesn’t touch asset entry).

HR recruiting ranks fourth — not because it can’t be done, but because resumes are full of personal information (names, phone numbers, ID numbers), and Personal Information Protection Law compliance is a medium-to-high-level risk: a candidate submitting a resume doesn’t equal consenting to have a third-party AI parse it. The veteran’s suggested order is: use asset entry to run a free closed loop first, then upgrade to HR and financial verification. I took that advice.

Financial verification ranks second with the highest per-customer value (a single large enterprise’s financial shared-service center can reach 400k–500k RMB/year), but it also has the highest compliance barriers: it touches tax filing and trade secrets, and reimbursement slips contain ID numbers and bank account numbers. Better to touch it after building experience with asset entry.

Where exactly is n8n’s thin layer of paper

This is something I only learn by doing it myself.

Conclusion: half a day gets you a demo-able prototype, but that’s not production-ready. Production-grade work takes another 3 to 5 days per customer. The research listed 8 real pitfalls:

1. Webhook production configuration — the #1 most frequent issue in the n8n community; requires an Nginx reverse proxy plus HTTPS, half a day to a full day.
2. Real OCR accuracy — phone photos are far below demo quality.
3. Unstable JSON from LLMs — you need schema validation plus retries plus fallbacks; debugging eats a third of the entire project.
4. Every customer’s ERP/WMS interface is different and needs separate adaptation, 0.5 to 1 day each.
5. Error handling and observability — at least one more day.
6. Data mapping expressions, credential management, pagination and rate limiting — these are the thin layers of paper that break at a single poke.
7. Reusing one workflow across different customers: each of the first three customers needs 15% to 30% of node logic customized (different ERP interfaces, different field definitions, different email templates); marginal cost only drops starting with the fourth.
8. Multimodal integration — the realistic path of calling vision/OCR models inside n8n (self-hosted LLM or relay station); credentials and timeouts are the traps.

So “standardization” isn’t something you do on day one — you have to survive the first three customers. “A workflow written in half a day” is true, but it’s a demo-able prototype, not a production system you can charge for. The gap between those two things is this industry’s biggest cognitive trap.

The relay station can’t run naked

My API relay station’s compliance status is unresolved, so it doesn’t go to production. The research gave a clear way out: lean on the mountain that is Alibaba Cloud Bailian.

Bailian (the Qwen series) was in the CAC’s first batch of 117 registered large models back in April 2024. Plug into Bailian as a secondary application (ISV status), use its already-registered models, and you only need to do two lightweight things yourself:

First, large-model registration — not filing. You’re just calling an already-registered API, so you go through the registration track: 1 to 2 months at the local cyberspace administration office, low review difficulty, and you just note the registration number of the Bailian base model in your materials. The heavy kind of large-model filing is for self-developed models — you don’t need to go through that.

Second, algorithm filing. Here I have to honestly make a correction: the research wrote this too absolutely at first. The Interim Measures for the Management of Generative AI Services govern “services provided to the public within China.” What I do is B2B targeted enterprise delivery (used internally by an agency bookkeeping firm) — strictly speaking, it most likely doesn’t fall under “facing the public,” and neither registration nor algorithm filing may be triggered. The right move is to first consult the local cyberspace administration office in writing to confirm — don’t spend 10k–50k RMB on an agency right away; you may not need to spend that money at all.

How to run during the transition: Bailian direct connection only; demote the relay station to an internal model routing layer that provides no external services; do free cases first without charging (no charges means the ICP commercial license hard threshold isn’t triggered); keep all data in a domestic closed loop. Never use overseas models, never run the relay station naked facing the outside.

How not to get locked into Bailian alone: DeepSeek is MIT-licensed — commercial use, redistribution, and modification allowed, no geographic restrictions. Once the business is running with revenue, self-deploy DeepSeek as your own compute foundation, then go through the full large-model filing, running Bailian and self-deployment on dual tracks. So DeepSeek is the fallback insurance, not the production choice right now.

There’s another threshold I have to add: the ICP commercial license. The research only said “apply when you start charging,” missing the hard requirements — registered capital of at least 1 million RMB, domestic ownership, social insurance for 3 employees in most provinces, and a pre-existing ICP filing. For an indie developer this isn’t something you do casually; it can take months. Either evaluate it early, or adjust the business model (don’t build your own platform — embed into the customer’s existing systems, which may not trigger the “commercial internet information service” threshold at all).

Model selection: Qwen-VL-OCR (image recognition, roughly ¥0.005 per thousand tokens) plus DeepSeek-V3 (text analysis and verification, ¥1 per 1M input) — both on the registered list. Incremental cost per resume or invoice is under ¥0.01; processing 10,000 documents a year costs ¥50–100 in API fees — the LLM is so cheap it’s almost not a cost. The main costs are n8n self-hosting ops plus my own development and maintenance time.

Find middlemen, not end consumers

My first target segment is agency bookkeeping firms and tax advisors. The reasoning is practical: 70% of the country’s micro, small, and medium enterprises are reached through agency bookkeeping firms, and the agency’s “invoice entry and reconciliation” is a pure cost center — its boundary with the revenue center that makes them money, “tax planning,” is crystal clear, which happens to satisfy the “non-conflicting” principle.

Two legs of cooperation:

White-label — I build the invoice OCR workflow, the agency slaps its own brand on it and sells an “AI smart bookkeeping value-add package” to clients; agency fees rise from 200 to 300–500 RMB/month, they keep the premium spread, and I collect a technical service fee.

Embedded — the workflow plugs directly into the agency’s daily operations, and I charge a monthly technical service fee of 500–2,000 RMB per company.

How the three principles of interest alignment turn into action:

Irreplaceable — I exclusively hold the n8n workflow orchestration plus LLM tuning stack; the agency can’t build it itself. Non-conflicting — I only touch the entry-and-verification cost center, never their tax-planning revenue center. Information asymmetry — the boss doesn’t know the machine costs only 1/5 to 1/10 of human labor; the middleman knows the customer’s budget ceiling; I know the real cost of AI infrastructure. Both sides agree not to disclose this spread to the boss, and monetize the information gap.

The opening line: “I don’t want your money — I’ll run it free for a month. One machine working an hour matches what three to five of your accountants do entering invoices all day, and you don’t lose a cent of what you charge clients. If it doesn’t work, you pay nothing.” Use free to drop the trust threshold to the floor, then let results speak before talking money.

But there’s a competitor risk I have to flag: ERP distributors may prefer pushing Kingdee and Yonyou’s own official AI modules over my third-party n8n solution. I have to find cross-system gap scenarios the official offerings don’t cover to survive. The veteran didn’t expand on this — I’ll have to figure it out in the field.

The pitch to the boss, and honest math

To the boss, only a few sentences: how much an employee costs per year, how much a machine costs per year.

The research gave a full script:

One of your accountants, salary plus social insurance and housing fund, costs you roughly 70k to 100k RMB a year. One machine — server fees plus API fees — tops out at 3,000 to 6,000 RMB a year, and an hour of it processes what three to five of your accountants enter manually in a full day. I won’t take your money; let me run it free for a month and show you the results. If I can cut a tenth of your bookkeeping entry work, you keep every cent you charge clients. If it works, pay me a modest monthly technical service fee — far cheaper than hiring.

But I have to honestly correct the 500k math. Three pricing calibers in the research conflict: the mid-size-customer caliber of 120k–160k RMB per client per year, the agency monthly-fee caliber of 6k–24k per client per year, and the revenue-share example caliber of 9.6k–14.4k per client per year. The claim “sign 5 agencies plus 3 RPOs equals 400k–600k” only holds under the most optimistic caliber; by the realistic volume-driven agency revenue-share caliber, 8 clients bring only 70k to 100k-odd RMB a year — a 4 to 5x gap.

The reality: agency revenue-sharing is high-volume, low-unit-price. To touch 500k, either go direct-enterprise sales (120k–160k per client times 4–5 clients — weaker channel but higher per-customer value) or push the agency count to 15–20. A more honest expectation is 300k–400k in year one, reaching 500k in year two after accumulating experience.

Annual cost per position (recruitment-platform figures, not official statistics — treat as market reference values): data-entry specialists in tier-2/3 cities earn 4–6k RMB/month; full employer cost including the five social insurances and one housing fund runs about 60k–80k RMB/year; recruiting specialists in tier-1 cities are higher, 9,000–11,000 RMB/month full cost, 110k–130k RMB/year. Machine annual cost: LLM APIs are extremely cheap (¥50–100 for 10,000 documents a year); the main cost is the self-hosted n8n server plus API fees — a few thousand RMB a year at most. The ratio holds up; it’s just not as exaggerated as “1 to 5.”

The first free case: diagnose, calculate, deliver, distill

The chosen direction: validate with asset entry first, then scale with financial verification.

The first free case is asset entry (asset ledgers and inventory entry for property/warehousing companies). A four-stage closed loop:

Diagnose — find out how this company currently does asset ledger entry: how many people, how slow, where the errors are, what WMS or asset system they use (Kingdee / Yonyou / Jiandaoyun).

Calculate — compute the labor value of replacing 10% of the entry workload and show the boss an ROI table. Honestly note that the real OCR replacement ratio may be 1:1 to 2, not the demo’s 1:3 to 5.

Deliver — a half-day n8n workflow: Webhook trigger (photo / email / WeCom) → call Qwen-VL-OCR to recognize asset nameplates, labels, and inventory sheets → Code node parses into structured JSON (asset number, model, serial number, quantity, location, person responsible) with schema validation, retries, and fallbacks → AI Agent validates numeric ranges and dedupes → IF branching (qualified entries written to the WMS/asset system API, unqualified ones sent to a manual review queue with a WeCom notification) → inventory variance reports. I now know in my head which parts can be built in half a day and which need 3–5 days of productionization.

Distill — parameterize the workflow using n8n Variables plus environment variables plus external JSON, export it as a reusable template, and cut the marginal cost for the second customer.

The channel for asset entry goes through ERP/WMS distributors (less mature than agencies), so finding the first free customer may be harder than with agencies. That’s this direction’s real tradeoff: easy to build and compliant, but finding customers takes more effort.

Places it could blow up

Flagging them in advance:

1. Real-world OCR accuracy is far below demos and may drag the replacement ratio down to 1:1–2, shaking the math behind the pitch.
2. Delivery can’t be standardized; each of the first three customers needs 15%–30% customization; an indie developer struggles to scale beyond 5 customers.
3. Middlemen may push official AI modules; you have to find gap scenarios the official offerings don’t cover.
4. Enterprise asset data contains trade secrets; customers may demand on-premises deployment, increasing delivery complexity.
5. B2B customer acquisition cycles are long; free case to paid can take 6–12 months; the 500k may only arrive in year two, and middlemen take a 30–40% cut.

Delivery is the main thing

I’ve stopped thinking about “AI-native replacing traditional industries.”

The veteran was right: delivery is the main thing. One machine working an hour matches what three to five employees do in a day; how much an employee costs per year, how much a machine costs per year — just those few sentences, said only to the boss.

Don’t talk tech stack with the boss, don’t talk model parameters, don’t talk automation principles. He doesn’t care what n8n is. He cares that an accountant costs him 100k a year, a machine costs him 5,000 a year, and he can try it free for a month first. Do the math for him; the rest is his decision.

About the method and transparency of this piece

This research used 10 agents in parallel (8 research threads plus 2 synthesis threads), with search cross-validated across two engines, Tavily and Metaso (the Grok engine was unavailable the whole time). Salary figures come from recruitment platforms (58.com, Liepin, Guipin) as unofficial statistical calibers; full FTE costs and “one machine hour equals 3–5 people’s day” are rules of thumb, not independently quantified and verified. All of this is flagged in my internal memo — don’t take it as hard numbers into contract negotiations.

The full internal research report (workflow details for each industry, the 8 hard parts of n8n delivery, the registered-model list with commercial terms, middleman-type comparisons, 152 source URLs) I’m not publishing — that’s my own battle map. This post is the publishable side: direction, compliance, channels, the math, the pitch, and the honesty of checking claim by claim.