Skip to content
fintechno AI
In production

Faktiko

The client's timesheet becomes the finished invoice — no retyping, no spreadsheets. Every figure is checked before the invoice leaves the building.

A real run

Ten stages from source document to an invoice ready for approval. Recorded from a real run, anonymised.

Demo — recorded output

Check chain

  1. 01Email / uploadTimesheet for week 12 (214 KB) taken from the mailbox.
  2. 02Document identificationSummary list · site [WRK] · [41000] Muster-Personaldienste
  3. 03Reader A · deterministic34 rows, total 1,042.90 — columns located via the table header, not hardcoded.126 ms
  4. 04Reader B · vision model34 rows, total 1,042.90 — read without any knowledge of Reader A's result.8437 ms
  5. 05Reconciliation A ↔ B34 against 34 rows · complete agreement in every cell.1 ms
  6. 06Checks0 errors · 2 warnings · 10 notes across five layers.16 ms
  7. 07Generate invoiceR-2026-0117 · 1,042.90 hrs × EUR 2.50 = EUR 2,607.254 ms
  8. 08Post-render checkGenerated PDF read back: number, hours, rate, amounts, period and German number format all match.23 ms
  9. 09ReviewerArithmetically, chronologically and formally consistent; mandatory content complete.51880 ms
  10. 10Release & sendAwaiting release. Nothing leaves without a click.

Result

Invoice number
R-2026-0117
Period
KW 12/2026 · 16.03.2026 – 22.03.2026
People
34
Hours
1.042,90
Rate
2,50 EUR
Net amount
2.607,25 EUR

Findings

  • Arithmetic

    Row sum 1,042.93 against printed total 1,042.90 — a 0.03 hr difference from rounding of the individual values, within tolerance. The printed total is what gets billed.

  • Calendar

    Week 12/2026, Mon 16.03.2026 – Sun 22.03.2026. The period matches the ISO week; this week has not been billed before.

  • Plausibility

    Hours distribution: 1× 0 shifts, 6× ~1 shift (7.70–9.44 hrs), 4× ~3 shifts (23.10–24.05 hrs), 23× ~5 shifts (37.80–43.52 hrs).

  • Plausibility

    One person with 0.00 hrs on the list — included in the total, but with no hours worked.

  • Master data

    The country in the stored IBAN and the country in the BIC do not match. Correspondent bank or a typo — clarify before sending.

  1. 01

    Read twice

    A conventional program and an AI model read every document independently. Only when both agree line by line is an invoice created — otherwise, a query.

  2. 02

    Recalculated

    The printed total is held against the individual lines. A misplaced decimal point shows up before it costs money.

  3. 03

    Read back

    The finished invoice is read in again and compared field by field with the source. That catches a fault in the template too, not just in the invoice.

  4. 04

    A person approves

    Everything up to the invoice runs automatically. The last click stays with a person — ten seconds against an invoice sent out wrong.

  5. 05

    Live DATEV integration

    Approved invoices go straight to DATEV. The tax adviser gets a booking proposal instead of a file to retype.

  6. 06

    Data stays in Germany

    Timesheets contain names and personnel numbers. The AI therefore runs in a German data centre, with no training on the documents.