Which steps AI runs, and where people check: decided at design time.
The table below is the standard split. Steps and roles are agreed with you and adjusted to your work.
Who does what, step by step
| Step | AI | Your staff | Licensed or authorized |
|---|---|---|---|
| Document collection (receive and sort) | Does it | Checks what AI can’t classify | — |
| Data entry | Reads and proposes | Confirms against the original | — |
| Deficiency and deadline check | Flags them, with proposed fixes | Decides on exceptions | — |
| Requests and reminders | Drafts | Approves and sends | — |
| Document pack and history for handoff | Assembles | Final check | — |
| Residence-status or hiring decisions | Never | — | You, or a licensed professional |
| Applications and filings | Never | — | The person authorized to file |
How one case moves.
An illustration of a workflow WATARI designs, based on Specified Skilled Worker onboarding and using synthetic data. Choose a stage to see what AI does, what your staff do, and what gets recorded.
Stage 1 / 6
Illustration · synthetic dataReceive and sort documents
- What AI does
- Takes in email attachments and photos, identifies each document type and links it to the case. Password-protected PDFs and wrong file types are caught here.
- What your staff do
- Checks only the items whose classification is uncertain.
- What is recorded
- Time received, sender, file hash
Documents received
- Passport copy.jpgFrom the worker · converted to PDF
- CV.pdfFrom the sending organization
- Employment terms.pdfFrom the employer
- Skills test certificate.pdfFrom the worker
- Employment certificate.pdfPassword-protected, cannot be opened
Stage 2 / 6
Illustration · synthetic dataRead each document into fields
- What AI does
- Turns the content into fields such as name, date of birth and company name, and shows where each value came from. AI output is a proposal; nothing is final yet.
- What your staff do
- Checks the proposed values against the originals and confirms them.
- What is recorded
- Proposed value, who confirmed it, before and after any edit
Structured fields (proposed)
- Name (Latin)NGUYEN VAN A · passport p.1
- Date of birth1999-04-12 · passport p.1
- Company nameサンプル食品 株式会社 · employment terms p.1
- Start date2027-04-01 · employment terms p.1
Stage 3 / 6
Illustration · synthetic dataFlag missing items, mismatches and deadlines
- What AI does
- Checks the case against the checklist agreed with you and lists missing documents, conflicts between documents and upcoming deadlines. Formatting errors such as full-width vs half-width characters or length come with a proposed fix.
- What your staff do
- Decides how to treat exceptions the rules do not cover.
- What is recorded
- Which checklist, and which version, was applied
Check results
- MissingHealth check certificate not received
- MismatchName: passport “NGUYEN VAN A” / CV “Nguyen Van An”
- FormatHalf-width space in company name → proposed fix
- ConfirmedSkills test certificate
Stage 4 / 6
Illustration · synthetic dataRequests and reminders, sent after approval
- What AI does
- Drafts each request in Japanese or English for the recipient and prepares a reminder at the agreed time if there is no reply.
- What your staff do
- Reviews the message and signs it off before it is sent.
- What is recorded
- Content, recipient, time sent, and whether a reply arrived
Request draft (English)
- ToThe worker
- MessagePlease send your health check certificate by 20 November. A clear photo of each page is fine.
- StatusAwaiting staff sign-off
Stage 5 / 6
Illustration · synthetic dataYour staff review and sign off; specialist questions go to a professional
- What AI does
- Lists every point that needs a check, next to the place in the document that supports it.
- What your staff do
- Staff accept, edit or return each proposal. Anything needing specialist authority, such as residence-status questions, goes to a licensed professional.
- What is recorded
- Who accepted or changed what, and why
Review log
- Company name fixAccepted (Tanaka · 12 Nov 10:04)
- Name spellingDecided to ask the worker (Tanaka)
- Start dateMatches the original (Tanaka)
Stage 6 / 6
Illustration · synthetic dataHand over the document pack and its history
- What AI does
- Assembles confirmed fields, the document list, ordered and merged attachments and the review history into one pack, and checks file-count and size limits.
- What your staff do
- Does the final check. Applications and filings are made by the person with the authority to make them.
- What is recorded
- Files in the pack with their hashes, and the review history
Review pack
- 01_worker_documents.pdf3 files merged
- 02_employment_documents.pdf2 files merged
- confirmed_fields.csv18 fields
- Review history14 actions
Document types and content are examples (synthetic data). Required documents vary by procedure and case.
Open the synthetic-data demo (hiring to arrival, Japanese interface)What we settle around the work
- Exceptions
- Uncertain readings, cases the rules do not cover, deadlines at risk and unexpected documents are never pushed through automatically; they go to a person. Who receives them, when and through which channel is decided at design time.
- Monitoring
- Progress per case, stalled work, upcoming deadlines and how often AI proposals were corrected, visible in one list.
- Reviewing outcomes
- Against the workload measured before we started, we regularly review minutes per case, rework and issues caught before submission, and adjust the rules and the split of work.
- Information handling
- Before any live case, we agree confidentiality and personal-data terms. First trials use samples without personal data or past cases you have de-identified. Where live data is stored, who can access it, how long it is kept and how it is deleted are agreed per engagement. If an AI model is to read documents, we get your approval for that use first.
Start by measuring today’s workload with us.
Tell us the work, the volume and the organizations involved. We will work out together where to start.
Request materials