Voice Invoicing in 2026: What Works and What Is Science Fiction

No Polish invoicing SaaS ships voice input. The technology exists: Polish ASR (automatic speech recognition) reaches 5.4-6.3% word error rate in clean benchmarks. PKO BP's iKO assistant logged 360,000 conversations in its early months using a voice-to-draft-to-confirm pattern. KipBill supports Polish with confirmation cards.

The gap between "works in the lab" and "works for invoicing" is set by two numbers.

The Two Numbers That Set the Ceiling

Number 1: Mixed-alphanumeric recognition accuracy is approximately 85%. NIPs, bank account numbers, and invoice numbers are mixed alphanumeric sequences. ASR systems that handle Polish speech well still struggle with "seventy-zero-one-zero-zero-zero-one-two-three-four" as a NIP. Eighty-five percent sequence accuracy means one in seven NIPs will be wrong. A wrong NIP on a KSeF invoice requires a three-document correction procedure.

Number 2: Voice biometrics are defeated by 3-second deepfake clones. Combined with PSD2's two-factor authentication requirement, this means voice can request an action but never be the authorisation. Voice authorisation of payments, invoice submission, or bank-detail changes is not viable. It is not a regulatory question. It is a physics question.

What Works: Voice-to-Draft

The viable pattern is narrow and defensible:

1. Push-to-talk capture. The user speaks invoice facts right after a job: "Faktura dla Kowalski Budowa, usługa remontowa, pięć tysięcy złotych plus VAT."
2. ASR transcribes to text. The system converts speech to a structured draft.
3. CRM entity resolution. The system matches "Kowalski Budowa" to the contractor database. If the match is ambiguous, it asks for clarification.
4. Render as editable draft. The draft appears on screen with all fields filled. Low-confidence fields (NIP, amounts, account numbers) are highlighted.
5. Mandatory read-back. The system reads back the critical values: amount, NIP, account number. The user confirms or corrects.
6. KSeF submission is a separate explicit tap. Not voice. Not automatic. A button press.

This pattern works for field trades: plumbers, electricians, construction workers who finish a job and want to issue an invoice before driving to the next one. "Wystaw fakturę zanim odjedziesz z budowy" (issue the invoice before you leave the site). The value is time-to-cash, not novelty.

What Does Not Work

Voice-authorised submission. The user says "wyślij" (send) and the invoice goes to KSeF. This is not viable. ASR can mishear "wyślij" as something else, or background noise can trigger a false positive. KSeF submission is irreversible. The confirmation must be a deliberate physical action.

Voice-entered NIPs and account numbers. The 85% accuracy on mixed alphanumerics means too many errors. These fields must be resolved from the CRM (the contractor card) or entered manually. Voice can say "Kowalski" and the system pulls the NIP from the database. Voice cannot dictate "one-two-three-four-five-six-seven-eight-nine-zero" reliably.

Voice corrections. Corrections have tax consequences. A korekta do zera zeroes out an invoice in the tax records. The correction must be deliberate, reviewed, and confirmed. Voice is too error-prone for this workflow.

The Market Gap

No Polish invoicing SaaS ships voice input. The field-trades segment is underserved. A plumber who finishes a job at 6 PM and drives home has two choices: issue the invoice from their phone (typing on a small screen with dirty hands) or wait until they get home and issue it from a laptop (by which point they may have forgotten the details).

Voice-to-draft solves this. The plumber speaks the invoice facts into their phone, the system drafts the invoice, they review it on screen, and they tap submit. The time from job completion to invoice issuance drops from hours to minutes. Cash collection accelerates.

The feature is narrow. It is not a general-purpose voice assistant. It is a time-to-cash tool for a specific workflow in a specific market segment. That narrowness is what makes it defensible: a general voice assistant is copyable within a quarter. A voice-to-draft pipeline tuned for Polish field-trade invoicing, with CRM entity resolution and mandatory read-back, maps to willingness-to-pay proven abroad.

This material is information of a general nature and does not constitute legal or tax advice. For a specific situation, verify the current rules or consult a qualified adviser.