Dictation at Work: Enterprise and Government Policies

Organizations can allow dictation at work, but not from just any tool. They approve tools that have a contract limiting what the vendor may do with speech, a security certification that IT can check, and retention and training settings that staff cannot override. They restrict or ban consumer apps that may keep, review or train on what employees say. Governments apply the same logic through formal authorization programs.

On a managed Windows PC or Mac, the simplest control is the dictation that comes with the operating system, because IT can switch off cloud recognition for every device at once. For a separate cloud tool, the decision rests on the contract and the vendor's default settings. This guide covers what organizations have actually done, what IT can and cannot see, and what a dictation policy should contain. For the basics, see our overview of dictation at work and our guide to voice privacy.

Organizations restrict dictation tools because speech carries the same secrets as documents

Speech is data. A dictated client name, product plan or patient note carries the same secrets as a typed one. What changes is the route. When audio goes to a cloud service, the question becomes what that service keeps, for how long, and who can read it. Samsung's 2023 memo put the concern plainly: the company was worried that data sent to AI platforms is stored on external servers, where it is hard to retrieve and delete ¹.

The concern was not abstract. Bloomberg reported that some Samsung staff had uploaded sensitive code to ChatGPT ². Dictation creates the same exposure through a different input. Anything a person says at work can end up in a transcript, and a transcript is a document.

Speech also leaves traces that a typed note does not. Audio can be recorded, transcribed, kept as a sample for quality review, or used to tune a model, depending on the vendor. A policy that only covers typed documents misses all of these.

High-profile leaks in 2023 led companies to ban consumer AI tools

The clearest corporate response came from Samsung. CNBC reported that Samsung confirmed it was temporarily restricting generative AI use on company computers ³. Samsung's own staff survey, run in April 2023, found that 65 percent of respondents believed generative AI services pose a security risk ⁴.

Bans like this are not limited to chatbots that type. Notetaking tools that record meetings have drawn the same scrutiny. A federal class action filed in August 2025 alleges that Otter.ai's Otter Notebook service, by default, does not ask meeting attendees for permission to record ⁵. The company's privacy policy lists training its AI on de-identified audio recordings and on transcriptions ⁶. The lawsuit's claims are allegations, and we found no ruling in the case when we checked. The pattern is what matters for a policy: a tool that joins meetings and keeps transcripts needs the same review as any other data processor. We explain the reasoning in more detail in our answer on why companies ban AI notetakers.

The problem is often the account, not the software. Cyberhaven, a data security vendor, reports the share of each AI tool's use that happens through personal accounts. Its figures are shown below. The vendor's page does not give a sample size or collection period in the text we read, so treat the numbers as an indication of scale.

The figures are the share of usage, not a count of employees. For ChatGPT, the share was 32.3 percent ⁷. For Gemini it was 24.9 percent ⁸, for Claude 58.2 percent ⁹, and for Perplexity 60.9 percent ¹⁰. Cyberhaven's point is that personal accounts bypass SSO enforcement, centralized logging, enterprise retention policies, and controls related to data usage or model training ¹¹. A policy that approves a tool for work but does not require a work account leaves those controls behind.

Governments reached the same conclusion from the other side. The US House of Representatives barred staff from using Microsoft Copilot on government-issued devices ¹². The report says the memo cited the risk of data leaking to unauthorized cloud services ¹³.

Governments approve cloud services through frameworks such as FedRAMP

In the United States, the cloud products that federal agencies use go through FedRAMP, the Federal Risk and Authorization Management Program. The statute defines a FedRAMP authorization as a certification about a cloud computing product or service ¹⁴. Federal rules for AI in agencies come from OMB memoranda. The memo M-25-21 says agencies should develop a policy for acceptable use of generative AI within 270 days ¹⁵. The memo uses the word "should" for that deadline, so it is guidance to agencies rather than a rule for private companies. A private organization can borrow the structure: a named owner, a written acceptable-use policy and a fixed deadline for putting it in place.

Some state governments have acted against specific apps. Oklahoma's governor banned the DeepSeek app on all state-owned devices, citing security risks ¹⁶. Action of this kind names one app. It does not decide whether dictation is allowed in general, so for a state agency the policy should cover the tool category and the device, not only the app on the banned list. Government employees have their own question, which we answer separately in whether government employees can use AI dictation.

Procurement checks security certifications and data processing terms

A certificate is the start of a procurement review, not the end of it. The contract decides what the vendor may do with speech. Under the EU's GDPR, a processor that handles personal data for a customer may not bring in another company without the customer's approval. Article 28(2) requires "prior specific or general written authorisation of the controller" ¹⁷. For dictation, that means asking which speech engine, cloud host or analytics provider sits behind the app, and whether each one has been approved.

For health data in the United States, the HIPAA rules add a second requirement. HHS says that a covered entity and a cloud service provider holding its health information must sign a HIPAA-compliant business associate agreement ¹⁸. We cover the clinical side in HIPAA-compliant dictation, and the term is defined in our glossary entry on business associate agreements.

Defaults matter as much as contracts. Google says Cloud Speech-to-Text does not log customer audio or transcripts by default ¹⁹. A purchasing team that checks only the contract can miss a default that decides what is kept. The checklist below turns these points into questions for a vendor.

Ask for the audit report or certificate itself, and read its scope. A logo on a website does not tell you which product, region or service was tested.

IT can manage built-in dictation settings centrally on Mac and Windows

For most staff, the built-in dictation on Windows and Mac is the first option to evaluate, because IT can control it without buying a second product. The controls differ by platform.

On Windows, speech recognition comes in device-based and online forms. Device-based recognition processes the voice locally and sends no voice data to Microsoft ²⁰. Voice typing in Windows 11 uses online speech recognition ²¹. The switch between them is a policy, and Microsoft's reference describes it as AllowInputPersonalization: a policy that specifies whether users on the device can enable online speech recognition. When the policy is disabled, speech services are disabled and users cannot enable them in Settings ²².

To change the setting on one PC, follow Microsoft's steps for Windows 11 in its support page on speech and privacy:

1. Open Start, then select Settings.
2. Select Privacy & security, then select Speech.
3. Turn Online speech recognition to Off.

Microsoft says that with online recognition off, only device-based features such as Narrator remain available. Test any dictation app you approve, because the page does not list which third-party apps use device-based recognition.

For a fleet, set the policy rather than asking staff to change the switch. The Group Policy equivalent is named Allow users to enable online speech recognition services, and the policy reference lists the Windows versions it covers.

On Apple devices, the rules depend on who owns the device. Apple's guide for IT says the MDM framework can never access personal information, including email, messages and browser history ²³. The guide says that on personal devices managed under User Enrollment, corporate and personal data are kept apart through a Managed Apple ID and a personal Apple ID. For company-owned devices, Apple says supervision gives IT controls that are unavailable for other deployment models ²⁴.

That answers the question many staff ask about their employer. On a personal phone enrolled through User Enrollment, IT should not be able to read what you dictate into personal apps. On a company-owned, supervised device, IT has more control, and the settings it applies are the ones you should expect. The details are in our answer to whether your employer can see what you dictate. Apple's guide is several years old, so confirm current names and key settings in its deployment documentation before you write a configuration.

On-device dictation fits strict environments, including air-gapped ones

On-device recognition is the lowest-risk option for an organization that cannot send speech to a cloud service at all. Apple says that on a Mac, voice inputs and transcripts for general text Dictation are processed on the device and are not sent to Siri servers ²⁵. Its SpeechAnalyzer interface keeps transcription on the device, but the language models must be downloaded first ²⁶. The download step is a planning point: in a closed network, the models must come from an approved source before dictation works.

Microsoft's embedded speech documentation, which covers a developer SDK rather than Windows dictation, says the feature is designed for on-device speech to text where cloud connectivity is intermittent or unavailable ²⁷. That condition describes a network with no cloud route, which is what an air-gapped system is. The page does not name air-gapped networks, so test the configuration in the environment where it will run. The term is defined in the glossary entry on on-device speech recognition, and our guide to voice privacy explains where each option sends audio.

The table compares the main options a policy might approve. The cells draw only on the sources cited in this guide. Where a source is silent, the cell says so, and the gap becomes a question for the vendor.

Yap, which we make, is an on-device dictation app for Mac and Windows, so it can be one of the options above. Like any tool, it should go through the same review of settings, network behavior and contract before it is approved. Details for Windows are on our Windows page.

A good dictation policy names approved tools, banned content and retention

A policy that staff can follow is short and specific. It should answer four questions: which tools are approved, what content may not be dictated into them, which settings must be on or off, and how long speech and transcripts may be kept. The question on what a dictation policy should include walks through a sample. The outline below is the structure to adapt.

1. Approved tools. List each approved app or service, the account type required (work or school, not personal), and the review date.
2. Banned tools and accounts. Name the consumer AI apps and notetakers that may not handle work speech, and say that personal accounts are not allowed for work content.
3. Restricted content. Keep health data, legal material and unpublished financial results out of any tool without the required contract. Health data needs a business associate agreement where a cloud provider holds it ¹⁸. For legal material, see dictation and attorney-client privilege.
4. Required settings. Specify the baseline for every approved tool: online recognition off where it is not needed, training opt-outs set, and history or logging switched off where the vendor allows it.
5. Retention and deletion. State how long audio and transcripts may be kept, who may delete them and how deletion is confirmed. Our guide to voice data retention explains what the major platforms keep.
6. Review and deadlines. Set a date for the first review. Federal guidance asks agencies to write their generative AI policy within 270 days ¹⁵, which is a reasonable model for a private deadline.

The approval path can be written as a short flow. A tool that processes audio on the device still needs settings checks, and a cloud tool needs its contract and certifications reviewed before anyone uses it at work.

Policies also need a rule for AI training. Otter's policy shows why: a vendor can list training on de-identified audio among its purposes ⁶, so the policy should require an account and a setting that prevents training on company speech. Our guide on whether dictation apps train AI on your voice covers how to check this for a specific app, and human review of voice recordings covers who may listen to audio. Finally, write the policy in plain language. A rule that staff cannot follow without calling IT will not hold.