Hidden AI the Research Findings

AI is already part of many services businesses use for email, meetings, marketing, sales and customer support. Across the 13 services across three busieness functions:

it appears both in obvious tools, such as writing assistants, and in everyday functions, such as sorting messages, recommending content and classifying customer enquiries. A business may therefore encounter AI even when it has not deliberately bought an AI product.

In the studied list everyday business services AI now does more than help somebody write a message or summarise a document. It can influence what people see, sort incoming enquiries, assess potential customers or applicants, retrieve business information and, when authorised, take actions. Some of this operates in the background or is enabled by default. A business can therefore be affected without deliberately choosing to use an AI assistant.

The benefits include less repetitive work, quicker access to information, faster customer responses and help identifying opportunities. However, those benefits depend on what information the AI receives and how much authority it has. An assistant that proposes a reply creates a different exposure from an agent that sends it, changes a customer record or carries out a transaction.

The app’s name does not define the boundary of the AI’s access. Connections can allow it to work across email, calendars, document stores, customer databases, websites and other business services. Information may also reach AI through uploads, shared screens, meeting transcripts, public posts or another participant submitting a conversation. Across the reports, the main forms of exposure are:

• Reading: AI can retrieve information from the app and authorised connected sources. Existing permissions matter, but information shared too widely can become much easier to find and combine.
• Sharing: Internal information can influence customer replies, appear in summaries or be passed to another provider. Copies can remain outside the original app after a feature is disabled.
• Acting: Configured agents and workflows can send messages, change records, manage appointments or perform other operations. Some require approval; others can run automatically or under previously granted approval.
• Learning: Providers distinguish shared-model training, business-specific learning, personalisation, feature improvement and retained memory. Refusing one use does not necessarily stop the others.
• Influencing: Background AI can affect visibility, advertising audiences, enquiry routing and recommendations without producing an obvious AI answer.

A recurring limitation is that 'not used for training' does not mean 'not read, processed, transmitted or stored'. Similarly, switching off a visible assistant does not establish that background AI, connected tools or automated workflows have stopped. Removing a connection also does not necessarily delete information already imported, distributed or incorporated into a model.

For a business, the practical question is therefore broader than whether an app 'has AI'. It is who controls it, why it needs access, what it can read or change, where information goes, and whether the benefit justifies that exposure. The strongest position is to grant the information and authority needed for a defined task, check the actual settings and behaviour, and retain a workable way for people to review, intervene and continue the work.

Share this page