Takuma Domoto · Tokyo, Japan

I build LLM systems for work where being wrong is not an option — and I sell them to the executives who sign off.

Five years selling and deploying AI and robotics into Japanese manufacturers and logistics operators. Now shipping production Claude systems into regulated professional work.

4 systemsLLM automation in production — regulated tax work, running as Claude Skills
¥1.0Bdeal closed through four approval layers, up to a parent-group board
5 yearscustomer-facing enterprise AI & robotics, mostly board-level buyers
Case 1 · Probabilistic output, zero-tolerance work

Automating a tax accountant's filing work — with the accountant still signing

A certified tax accountant firm in Tokyo. The work: year-end closing and statutory filing — reading receipts and ledgers, generating journal entries, building trial-balance comparisons, consumption-tax worksheets, and statutory breakdown schedules. Errors here are not bugs; they are professional liability.

I delivered four production systems as Claude Skills running inside the customer's Claude Cowork environment, and ran the engagement alone: discovery, requirements, implementation, delivery, and adoption support.

The design problem was never accuracy. It was: where is probabilistic output allowed to live, inside work that a licensed professional signs their name to?
Source documentsreceipts, ledgers, trial balances
Deterministic coretransfer, arithmetic, formatting — never the model
LLM adjudicatesgenuine ambiguity only, against versioned master tables
Confidence flags3 levels on every output — staff read only what needs eyes
Reconciliation gatedebits = credits, or nothing ships
The accountant signsfinal responsibility stays human

The pipeline that all four systems share.

Three decisions that made it deployable

  1. The model adjudicates; it does not calculate.
    Transfer, arithmetic, and format application run deterministically. Claude resolves only what is genuinely ambiguous — e.g. mapping a receipt description to an account — against versioned master tables, so every judgment is reproducible and explainable.
  2. Uncertainty is visible, per item.
    Every output carries a three-level confidence flag (confirmed / review recommended / review required). Review effort tracks actual uncertainty instead of document count.
  3. Delivery is gated on the books balancing.
    Reconciliation — debits equal credits, net income matches the trial balance — is a hard output condition wired into the pipeline, not a report someone might read.

Operationally: the firm's own Excel templates are preserved (formula cells are detected and left untouched) so audit traceability survives; per-client differences live in CSV master tables; a new client is onboarded by deriving mappings from a single correct input/output pair.

Status: all four systems in production. In August 2026 a second regulated practice — a labor & social-security attorney firm — signed, showing the approach transfers.

Case 2 · Frontier tech vs. the approval chain

A ¥1.0B automated warehouse, carried through four layers of decision-makers

At RENATUS ROBOTICS (AI × robotics automated storage), I lead solution sales and business development. I closed a ¥1.0B (~US$6.5M) deployment — the company's second contract ever — over a twelve-month cycle.

Month 1
Operating teamwarehouse & logistics staff — does it survive the floor?
Customer's board of directorsdoes the ROI model hold?
CEOis this defensible as a company bet?
Month 12
Parent-group board — approved¥600B-revenue group, chairman present — group governance

What moved it

What selling frontier technology in Japan taught me: safety and interpretability are not a values statement. A visible failure mode — knowing exactly where a system can be wrong, and what happens then — is what gets a proposal through the approval chain.
Principles

How I think about enterprise AI

Adjudication over generation

In high-stakes workflows, the model's job is to resolve ambiguity inside a deterministic pipeline — not to produce the deliverable unchecked.

Failure must be visible

Confidence flags and reconciliation gates turn "trust the AI" into "check exactly here."

The approval chain is the real interface

An enterprise adopts what its people can defend internally. Design for the ringi, not only for the end user.

ringi (稟議) — the internal approval memo that carries a Japanese enterprise decision upward through each layer of management.