Alienstacks Co., Ltd. — Bangkok · Tokyo · Reg. No. 0395569000469 (DBD)
EN / 日本語Contact
All services
SERVICES · AI · AUTOMATION

AI development and workflow automation that removes real work

We build AI agents, document pipelines, and LLM features into the systems businesses already run — quoting, accounting, approvals, reporting, customer replies. Not a chatbot on your website. Our AI systems are in production at more than 200 Japanese firms, and you can open three of our own AI products right now without registering.

AI agentsLLM integrationDocument extractionWorkflow automationRAG / retrievalHuman-in-the-loopAudit trailsNode.js / TypeScript
200+Japanese firms running our AI systems in production
6 hrs → 48sQuote turnaround for a Japanese manufacturer
3Live AI products you can open right now

What you actually get

Not a strategy deck and not a proof of concept that dies in a browser tab. A working system inside your business, measured against the manual process it replaced. In practice:

·A named process automated end to end — documents in, structured output into your existing system, with a human approving before anything is committed
·A review interface your staff actually use: what the AI proposed, how confident it was, and the source document it read, side by side
·An audit trail on every AI decision, so you can answer "why did the system do that?" six months later
·Integration with the systems you already run — major accounting platforms, ERP, CRM, email, shared drives — rather than another separate login
·A measured baseline: how long the process took before, how much AI handles unaided now, and the real per-document cost
·A model layer you can change. Provider choice is configuration, not a rewrite, so you can follow the market on price and capability
·Your own cloud, model provider, and repository accounts, owned by your company with us as a removable collaborator
·A written architecture note and handover pack in plain language

The measurement point is the one buyers underrate. Most AI proposals cannot tell you what "working" means. We insist on a number before the build — accuracy on your documents, percentage handled without a human touch, minutes saved per case — because that number is what tells you whether to fund phase two.

How we work: architecture before prompts

Most AI projects that fail in a small business fail for the same reason: someone wrote clever prompts and called it a system. Prompts are the easiest part and the least durable — models change, edge cases arrive, and a pipeline held together by prompt text has no way to tell you it has started being wrong.

So we design the plumbing first. Where does the document enter? What structured shape must come out, and how is it validated before anything touches your accounting system? What is the confidence threshold below which a human must look? What is written to the audit log? Which model does the cheap extraction and which does the expensive reasoning? Those decisions are the system; the prompt is a replaceable component inside it.

The commercial consequence is that you are not locked in. The model sits behind an interface, so when a provider halves its price or a better model appears, you change a setting rather than commission a new project. A shop that hard-wires one vendor's SDK into your business logic has, intentionally or not, sold you a rewrite in eighteen months.

Every consequential action goes through a human. In our own accounting product, AI staff read source documents and draft journal entries; the accountant reviews and approves, and every entry balances to the yen. That is not timidity, it is what makes AI usable in a business where being wrong has consequences — and it is why the system is in production rather than in a pilot. We operate that product ourselves, so we carry the cost of our own shortcuts.

What drives timeline and cost

No figures on a web page. But AI projects have unusually predictable cost drivers, and three of the four are about your data rather than our code.

How messy the inputs are

Consistent digital PDFs from three suppliers is a short project. Photographed handwritten forms in two languages from two hundred suppliers is not. Bring us twenty real examples on the first call and we can tell you which situation you are in.

How many steps the workflow has

Read a document and output fields is one step. Read it, price it against your rules, check stock, draft a reply, and file it in your accounting system is five — each with its own failure mode and review point. Cost tracks steps, not enthusiasm.

What it has to connect to

A modern system with a documented API is days. An older on-premise system with no API, or a platform whose approval process takes weeks, is the longest pole in the project and has nothing to do with AI. We surface it in week one.

The accuracy bar

Getting to broadly useful is fast. Getting to reliable enough that money moves on it takes validation layers, test sets, and review tooling. Deciding honestly which parts need which standard is where the budget is won.

Roughly: a single well-defined automation with a pilot on your own data is typically a 4–8 week build. A multi-step agent workflow spanning several systems, with role-based review and reporting, runs into months. Running costs are per-use model charges plus modest hosting, and we measure the real per-document figure during the pilot so you are not signing up to an unknown monthly bill.

Where AI fits in an SME

Quoting and order intake

Manufacturers and distributors receiving specifications by email and PDF, then cross-referencing spreadsheets to price them. This is exactly the six hours to 48 seconds case — the highest-return AI project we see in SMEs.

Accounting and back office

Receipts and invoices read into draft journal entries, expense classification, month-end preparation, and reconciliation — with the accountant approving. In production at 200+ Japanese firms via our own accounting suite.

Reporting and decisions

Turning the files a business already produces into a dashboard with cash runway, department profitability, and a scenario simulator — every number traceable to its source document, which is the only way an owner will trust it.

Retail chains use the same architecture differently: transaction data from the POS systems we build feeds demand forecasting and stock alerts. Property agencies use it for document handling and enquiry triage. The pattern is constant — connect the systems that already hold your data, then let AI do the reading and the drafting.

Proof you can open in a browser

We do not ask you to take AI claims on trust. Three of our own products are live, running on fictional data, no registration required — and each is a real system in commercial use, not a mockup.

And the client engagement most relevant to AI automation:

What we don't do

More nonsense is sold under the word "AI" than any other word in software, so here is our list plainly:

·We do not train foundation models from scratch, and we will tell you that almost no SME needs to.
·We do not build systems that take consequential action with no human approval. If you want money moved or contracts sent autonomously, we are the wrong firm.
·We do not sell AI strategy decks, workshops, or "AI readiness assessments" with no software at the end.
·We do not build a chatbot on your website and call it AI transformation. If a search box would do, we will say so.
·We do not sell developer-hours or staff augmentation for your existing AI team.
·We do not promise accuracy figures before testing on your own documents — anyone who does has not seen your documents.
·We will decline a project where the underlying process is undefined. AI cannot automate a process nobody can describe; fix the process first.

Related services

Questions buyers actually ask

How do we know AI will actually work for our process?

You do not, until it is tested on your real documents and your real edge cases — so we do that first. The engagement starts with a narrow, measurable pilot on one process using your own data, with a number attached: how much of it AI handles unaided, and how often a human has to correct it. If the pilot does not clear the bar we agreed, you stop there rather than funding a full build on a hunch.

What happens when the AI gets something wrong?

It will, so the system is designed around that rather than pretending otherwise. Every output carries a confidence signal and a link back to the source it came from, low-confidence items are routed to a person, and nothing consequential is committed without human approval. In our accounting product an AI drafts the journal entries and an accountant approves them — that is the pattern we build, not autonomous decisions nobody can audit.

Where does our data go, and is it used to train someone’s model?

We use commercial API tiers that contractually do not train on your data, and we design what leaves your systems in the first place — often only the extracted fields rather than the whole document. Where the data is genuinely too sensitive to send anywhere, we can run smaller open models on infrastructure you control, with a frank conversation about the accuracy you trade away. This is settled in writing before the build, not after.

Are we locked into one AI provider?

Not if it is built properly. The model sits behind an interface, so switching providers or mixing them is a configuration change rather than a rewrite — we routinely use a fast, cheap model for extraction and a stronger one for reasoning in the same pipeline. This matters commercially: model pricing and capability change every few months, and you should be able to follow that without paying for a new project.

How long does an AI project take, and what does it cost to run?

A single well-defined automation — one document type in, one structured output out, with human review — is typically a 4 to 8 week build including the pilot. A multi-step agent workflow touching several of your systems runs longer. Ongoing cost has two parts: per-use model charges, which we measure during the pilot so you know the real number per document, and hosting, which is usually modest by comparison.

Can you connect AI to the systems we already use?

That is normally the point. We integrate with major accounting platforms, ERP and CRM systems, email, and shared drives, so the AI works inside your existing process instead of becoming another place staff have to log in. Where a system has no usable API we will tell you honestly what that costs in workarounds before you commit.

Will this replace our staff?

In the work we have delivered, it removes the retyping rather than the person. A Japanese manufacturer we worked with cut quote turnaround from six hours to 48 seconds — the sales engineers did not disappear, they stopped cross-referencing five spreadsheets and started answering more customers. If your honest goal is headcount reduction, say so on the first call so we can be realistic about what is achievable.

Bring us the process you hate doing.

A 30-minute call with an engineer, in English or Japanese. Bring twenty real documents and the steps a person takes today, and we will tell you honestly whether AI helps and what it would take.

Request a consultation Open JP Accounting Suite demo