01
Read the document
OCR and classification across 20+ document formats. Each document is tagged with its type before extraction begins.
Trust & Fiduciary · Liechtenstein
A Liechtenstein trust company
hands routine bookings to AI
An AI accounting pipeline posts the routine entries automatically - and routes every uncertain document to a specialist.
30×
faster per document
72%
manual workload reduced
100+
booking rules encoded
16
weeks to production
The story
For years, one person held it all. Two decades of cross-border booking rules - which jurisdiction applied, which exception to make, what to do with the document that didn't fit - lived in a single specialist's head. A new hire needed about a year before anyone could trust them with the work.
Every week the team ran the same loop on 80,000 documents: read, judge, book, repeat. Too important to rush, too repetitive to keep doing by hand. The brief from the client was blunt - automate it, but never post a wrong entry.
So we built a system that takes the routine and leaves the judgment. It books what it is sure about in seconds, and the moment a document is unclear it stops and asks a person. The expert stayed the final word. The keyboard work didn't.
The challenge
Routine work only an expert could do.
A specialist team was booking 80,000 client documents a year by hand - tax-relevant entries pulled from dozens of financial institutions, every one weighed against multi-jurisdiction rules. You'd normally hand the routine to automation and keep people on the hard cases. Here, the routine work itself demanded specialist judgment, and onboarding a new specialist takes about a year.
- 120 hours a week of manual data entry, at 2.5 minutes per document.
- A 1% error rate - small in percentage terms, large in trust accounting, where one mis-booking becomes a compliance event.
- 20+ years of expertise in one person's head - jurisdiction rules, internal-vs-external logic, exception handling, none of it documented.
The brief was specific: automate the workflow, but never post incorrectly. The system could route uncertain documents to a human - it was not allowed to guess.
The solution
Five layers, cross-checked.
We built a multi-model pipeline that replaced the manual keying step. Five components work together to handle every document; anything they're not all certain about goes to a human reviewer.
02
Apply the rules
A deterministic engine encodes 100+ trust-accounting rules across jurisdictions, and enriches each document with the firm's reference data.
03
Decide, explainably
Explainable decision trees handle the cases that need probabilistic judgment - chosen because the path through them is human-readable for audit.
04
Cover the long tail
A regionally-deployed LLM absorbs the format variations the rules don't anticipate. No public LLM API is ever touched.
05
Cross-validate
All four layers check each other. Agreement posts the entry; disagreement routes the document to a human reviewer.
Governance & compliance
Conservative by design.
Data stays in jurisdiction
The system deploys in the client's chosen cloud region or on-premise. For this engagement, all processing and storage stays in the client's home jurisdiction.
No public LLM API
The LLM runs in a regional managed deployment. No prompts, completions, or training data touch a public LLM API - OpenAI, Anthropic, or otherwise.
Every decision logged
Input document, output booking, model versions, confidence score, the auto-vs-review routing, and the human who confirmed it. Every step is reconstructable.
The expert is the final layer
Around 30% of documents route to a human accountant. Manual override is available end-to-end. We built around the senior accountant's judgment, not over it.
Results
Six months after go-live.
Manual data entry, hours per week
BeforeAfter deployment
A 72% drop in routine keying. Per document, the same work fell from 2.5 minutes to 5 seconds - a 30× speed-up, with no posting errors observed in production.
- Team capacity went up without hiring. The same team now absorbs routine volume that previously forced a long-onboarding hiring conversation. Specialist time goes to the documents that need judgment.
- The audit trail holds up for regulators. Every decision is reconstructable.
- 20+ years of tribal knowledge is now a system. 100+ booking rules - documented, transferable, and updateable.
- Better-than-human accuracy on the auto path. No posting errors have been observed in production.
Enterprise-grade controls and 30× speed are compatible:
this is governed AI, in production.
Questions
Can a regulated firm use AI without sending data to OpenAI or Anthropic?
Yes. We use a regional managed LLM deployment. No public LLM API is touched at any point in the pipeline - no prompts, completions, logs, or backups leave the client's environment.
How do you control hallucination risk in a regulated workflow?
We don't rely on the LLM as the final judge. Deterministic business rules, explainable ML, and an LLM layer cross-check each other. When the models disagree, the document goes to a human reviewer. The system never books with low confidence.
Can AI work with legacy systems that have no API?
Yes - file-based output. The system generates an automated export feed compatible with the client's legacy booking platform. A manual edit portal handles the fields documents don't carry.
How long does an AI accounting deployment take?
16 weeks for this engagement. 12 weeks is the repeatable target for similar scope - 3–4 transaction categories, a single target booking system, historical data available. A Phase 0 readiness assessment runs ahead of the 12-week timer.
What audit trail does AI-driven booking need to pass a regulator review?
Every decision is logged with input, output, model versions, confidence score, routing decision, and any human confirmation. Every step is reconstructable.