# When to process locally and when to use the cloud Start with a permitted data route, then quality, deadline and the cost of an accepted result. A decision diagram and example show why 30% less audio need not mean 30% savings. L00P.AI · 1.0 · 2026-10-13 https://www.l00p.ai/en/resources/series/kiedy-przetwarzac-lokalnie-kiedy-w-chmurze/ ## 01 / Start with the data route A leader choosing transcription needs three answers: may this recording be sent out, will the result be good enough, and will it arrive on time? Compare costs among options that meet those requirements. A cheap service does not resolve source confidentiality, and a local application does not guarantee that a name will be recognised correctly. Our starting point is three workshop notes from June and September 2026: an internal source-protection rule, an account of cloud risk, and a cost-routing proposal. We adapt the decision method. We do not claim that the proposed system has been deployed or that we have run a new quality test. ## 02 / Four separate questions for the provider Check use of content for training, retention periods, processing location, and access by people and subprocessors. Record the service, plan, setting and verification date. A provider's name alone does not describe your configuration. ElevenLabs documentation checked on 13 October 2026 describes opting out of training on newly submitted data through Data use settings. It states a separate default for Enterprise customers, qualified by what is necessary to provide their services. Do not infer that earlier data was deleted or its influence on a model reversed. [Source: data use](https://elevenlabs.io/docs/help-center/legal/is-my-data-used-to-improve-eleven-labs-ai-models). Zero Retention Mode is a separate mechanism for eligible Enterprise customers and specified products. The documentation covers STT audio input and text output, but specifies API traffic only; the web interface and playground are excluded. This is the provider's description, not a test of our account configuration. [Source: retention](https://elevenlabs.io/docs/eleven-api/resources/zero-retention-mode). Choosing a region does not automatically enable zero retention either. The documentation describes separate environments and conditions for restricting processing to a region; optional integrations may change the data route. [Source: data location](https://elevenlabs.io/docs/overview/administration/data-residency). ## 03 / A local interface is not enough Check where the model runs, where the application sends requests, and where recordings, transcripts and backups are stored. An application running on your computer may call an external API. Local processing also needs access controls, updates and file-deletion rules. In this method, confidential material stays local until an acceptable external route has been documented. An automatic cloud fallback would violate that rule. A recording being publicly available is not, by itself, a complete assessment of rights for further use. ## 04 / Decision diagram ```text Recording | +-- External route not authorised --> process locally or hold | +-- External route authorised | +-- Compare quality, deadline and cost on the same sample | +-- Choose local / cloud / an explicitly documented split ``` This original text diagram models a decision; it does not measure a particular product. A split might mean local preparation followed by sending only permitted excerpts. Removing names from a transcript does not guarantee that the voice or conversation is anonymous. ## 05 / Count savings after correction This is an illustrative example, not a price list: 10 hours of recordings at 1 currency unit per hour, with no minimum charge or rounding. Processing everything costs 10 units. If you correctly exclude 3 hours of unwanted sections, the charge for the remaining 7 hours is 7 units. Gross savings are 3 units, or 30%. If preparation, boundary checks and corrections cost 2 units, the total is 9 and net savings are just 1 unit, or 10%. Additional fees and retries may change the result. A historical note suggesting 30% savings does not establish that result for other recordings. Preserve the original. When cutting a copy, record a time map so an excerpt can be found in the source. Send uncertain sections and speech over music for listening review. An aggressive filter can reduce the bill by deleting speech you needed. ## 06 / Same sample, full cost Compare on a permitted sample containing difficult names, numbers, silence, overlapping voices and conditions typical of your work. Record model versions, parameters, waiting time and human correction time. Check the reference transcript against the recording; agreement between two transcripts is insufficient. For local processing, include energy, hardware, maintenance and competing workloads. For cloud processing, include the billing unit, plan, extras, retries, transfer and correction. Compare the cost of an accepted result. We report no new WER, speed measurement or winner: that test was not performed here. ## 07 / A decision you can reconstruct Record the material category, permitted data route, configuration, deadline, acceptance threshold and failure procedure. If local processing misses the deadline and cloud processing is prohibited, the outcome is delay or a hold, not a silent upload. Workshop sources provide the method's context; current service conditions come from the linked documentation. Codex prepared the text, translation, diagram and example. No independent second-model review or account audit was performed. Changes to the documentation, a calculation error or evidence that the actual data route differs from this description trigger a correction. Which constraint comes first in your workshop: data, deadline or quality?