Alpha · first test version in early November
Questions across all your paperwork.
The AI sees documents, emails, appointments and tasks together and finds what you would miss by hand. It proposes, you click. How well each model does, we measured.
What you can ask
Questions from the life of Jonas Falk — a cabinetmaker in Cork, landlord, father of two; made up, with an archive of 299 documents. Every question cuts across his paperwork, every one has a finding you would miss by hand, and it ends in a proposal that one click accepts. Below each: how often each model solved it fully in three runs.
Family
“What have I paid privately for digital subscriptions and services over the last twelve months?”
What Undan finds
- The music subscription runs twice — individual on the card, and the family plan through the payment service.
- A standing order to a friend is a shared subscription, with no invoice at all.
- A free trial quietly became an annual plan; it renews in February.
- The AI assistant and the design software belong to the business — and what runs through payment service and card counts once.
Bookkeep on the business receipts, tasks for cancelling.
Fully solved, in three runs
Claude- Opus
- Sonnet
- Haiku
Local- Gemma 4
- Qwen3.8
Family
“Which doctor's and dentist's bills since 2025 haven't I submitted to my health insurer yet? Mark them.”
What Undan finds
- A GP visit from November is missing from the insurer's claim statement — the twelve-month deadline runs out in six weeks.
- Three new receipts were only filed, never marked for submitting.
- The insurer no longer pays for the brace; the maximum benefit has been paid.
Submit on the four receipts, a task with the deadline.
Fully solved, in three runs
Claude- Opus
- Sonnet
- Haiku
Local- Gemma 4
- Qwen3.8
Holiday home
“Which quote for the roof is better?”
What Undan finds
- The cheaper one leaves out scaffolding and VAT — complete, it costs £1,166 more.
- It only starts in March; the other one before winter.
The comparison in the project note for the roof.
Fully solved, in three runs
Claude- Opus
- Sonnet
- Haiku
Local- Gemma 4
- Qwen3.8
Rentals
“Tomasz is moving out on 30 September. What do I need to do?”
What Undan finds
- The deposit of €1,380 has to go back within ten days — that is what the lease says.
- The scratch in the floor is already in the check-in report of 2021; it must not be deducted.
- The insurer wants to know about a vacancy over 30 days, and the new tenant must be registered within a month.
The missing steps as tasks in the project note.
Fully solved, in three runs
Claude- Opus
- Sonnet
- Haiku
Local- Gemma 4
- Qwen3.8
Which model? Measured, not claimed
The same questions to five models, three times each, through the same chat as in the application — every answer checked against what is in the archive. Two levels: look up, one question to one source (“How much is Tomasz's deposit?”), and cross-check, the four questions above, across many documents.
Either the best model — or straight to one on your own computer. Any model can look things up, a local one too, without anything leaving the house. Only Opus reliably cross-checks across your paperwork. In between there is little to gain: Haiku solves it less often than Gemma 4 or Qwen3.8 on your own machine.
| Model | Look up | Cross-check | Time per question | Cost per question |
|---|---|---|---|---|
| Claude OpusClaude subscription | 15 of 15 | 11 of 12 | 2 min | ≈ $1.50 |
| Claude SonnetClaude subscription | 15 of 15 | 6 of 12 | 45 s | ≈ $0.45 |
| Claude HaikuClaude subscription | 13 of 15 | 1 of 12 | 1.5 min | ≈ $0.30 |
| Gemma 4 31Blocal, two graphics cards | 15 of 15 | 3 of 12 | 1.3 min | — |
| Qwen3.8 27Blocal, two graphics cards | 14 of 15 | 5 of 12 | 8 min | — |
Measured on 4 October 2026 on the sample archive (made up): three runs per question and model; “fully solved” means every finding is named — computed, not judged by hand. Cost: the API price Claude reports — with a Claude subscription you don't pay it per question. Local: Gemma 4 in Google's 4-bit build, Qwen3.8 in 8 bit, each on two RTX 3090. Time: one question including the follow-up.
How it works
Mail in one program, files in a shared drive, appointments in the calendar, tasks in the next app: each one knows only its own piece, and no AI can see that the invoice belongs to the email and the payment to the invoice. Undan keeps everything in one archive — documents with their facts, accounts, emails, contacts, appointments and tasks that refer to each other. Only on that can an AI answer questions that cut across everything.
How the AI works
Proposes, never acts
The AI has no tool to change anything by itself. It proposes — filing, facts, tasks, appointments, email drafts — and only your click carries it out. Every click can be undone with
Ctrl+Z.Code does the maths
Sums, dates, deadlines and checksums such as the IBAN's are computed by Undan itself. The model only comes in where judgment is needed: what kind of document is this, where does it belong, what does the email want?
You decide what it sees
The AI can access your whole archive — calendars, mail accounts and contact fields only as far as you allow, the web per session. This is enforced inside Undan, not in the prompt: what you haven't allowed doesn't exist for the AI. A finer permission model, changeable per session, is planned.
Your model
Claude through your own subscription — or a local model on your computer, then nothing leaves the house. A model per task: one for the inbox, one for email, one for the chat. More providers are coming.
And without AI? Undan is a complete tool without AI as well: it reads amounts, dates and IBANs from the text itself, the inbox form is the same, and search, mail, calendar and tasks need no model. The AI is an offer, not a requirement.
Join the waitlist
We'll write to you when there's something to try — and otherwise not.