Most people using Salesforce, Figma, or GitHub Copilot have no idea they’re sitting on top of a sophisticated AI inference stack — and that gap in awareness is exactly where enterprise risk accumulates.
The list of apps that use AI has exploded past the point where any IT team can track it manually. This isn’t a consumer curiosity. It’s a procurement and governance problem. When AI runs inside tools your teams already approved, the data exposure and model behavior questions don’t disappear — they just go unasked.
Here’s a category-by-category breakdown of 50 products, what the AI actually does inside them, and what most users never think to ask.
Productivity and Collaboration Tools
Microsoft 365 Copilot routes your emails, documents, and Teams transcripts through Azure OpenAI Service. The model sees full document context — not just the sentence you’re editing. Google Workspace (Duet AI) does the same on Google’s infrastructure, with Gemini Pro handling drafting and summarization across Gmail and Docs. Users think they’re getting autocomplete. They’re actually submitting document content to a hosted LLM endpoint on every interaction.
Notion AI uses OpenAI’s API behind the scenes — your workspace data leaves Notion’s servers. Slack AI (available on paid plans) summarizes threads using a combination of Anthropic’s Claude and internal models. Zoom IQ transcribes and summarizes meetings using Zoom’s own ASR models plus third-party NLP. Otter.ai sends audio to its cloud for transcription via fine-tuned Whisper variants. Grammarly now runs generative suggestions through its own LLM, trained partly on user text. Calendly uses ML to suggest optimal meeting windows based on historical scheduling patterns.
What users don’t realize: every “suggested reply” or “summarize this thread” button is an API call that includes message content. Most enterprise agreements don’t prevent the vendor from using that data for model improvement unless you’ve negotiated it out explicitly.
Developer Tools and Code Platforms
GitHub Copilot is the most widely deployed AI coding assistant in enterprise. It runs on OpenAI Codex (a GPT-4 derivative) and sends your active file plus surrounding context to Microsoft’s inference infrastructure on every keystroke pause. Cursor operates similarly, with a choice of GPT-4, Claude, or its own model depending on tier. JetBrains AI Assistant routes through JetBrains’ own proxy service before hitting OpenAI or Grazie models.
Tabnine is notable because it offers a fully on-premise deployment option — one of the few apps that use AI in the coding category that lets you keep data inside your own infrastructure. Replit Ghostwriter runs on Google Cloud infrastructure. AWS CodeWhisperer routes through Amazon Bedrock. Sourcegraph Cody uses Claude or GPT-4 depending on configuration, and it indexes your entire codebase before querying — meaning your proprietary code structure exists on their servers as vector embeddings.
Snyk Code uses ML to identify vulnerability patterns, trained on millions of open-source repositories. Linear uses AI to auto-triage and categorize issues. Jira (via Atlassian Intelligence) now summarizes tickets and suggests work breakdowns using hosted LLM calls.
What users don’t realize: embeddings are not anonymized. A vector representation of your codebase is still a representation of your codebase. Vendors who store embeddings own a structured fingerprint of your proprietary logic.
Customer-Facing and CRM Applications
Salesforce Einstein uses a combination of proprietary models and the Einstein GPT layer (backed by OpenAI partnership) to power lead scoring, email generation, and forecasting. HubSpot AI runs content generation and deal prediction through OpenAI endpoints. Zendesk AI uses its own fine-tuned models alongside third-party LLMs to classify tickets and draft agent responses.
Intercom Fin is a fully GPT-4-powered support bot — every customer inquiry your users submit goes through OpenAI’s API. Drift uses ML-based intent detection plus LLM generation for conversation routing. Gong transcribes and analyzes every sales call using its own ASR and NLP pipeline, building behavioral models on your team’s speech patterns. Chorus (ZoomInfo) does the same with slightly different model architecture.
Klaviyo uses predictive ML to optimize send times and segment churn risk. Braze uses AI for next-best-action recommendations in customer journeys. Segment applies ML to identity resolution and predictive audiences. Amplitude uses AI to surface behavioral anomalies and forecast conversion.
What users don’t realize: your customer data — names, behavior patterns, conversation transcripts — is being used to generate AI outputs that get fed back to your sales and support teams. The AI loop is closed on your customers’ private interactions.
Enterprise Infrastructure and AI Control Platforms
This category is where the decisions with the longest consequences get made. Most apps that use AI in the productivity and CRM categories make a choice for you: their cloud, their model, their data handling terms. Enterprise AI infrastructure platforms exist to give that control back.
Peridot sits in this category as a platform built specifically for enterprises that need AI running inside their own infrastructure. Rather than connecting your data to a vendor’s hosted model, Peridot runs the AI layer within your environment — giving IT directors control over which models run, on what data, with what access policies. For regulated industries, this is the difference between an AI deployment you can audit and one you can only hope is compliant.
AWS Bedrock offers hosted access to multiple foundation models (Claude, Titan, Llama) through Amazon’s infrastructure, with strong isolation guarantees. Azure AI Studio provides a similar model marketplace with Microsoft’s compliance wrapper. Google Vertex AI hosts Gemini and open-source models with enterprise SLAs. Databricks Mosaic AI integrates model training and inference directly into your data lakehouse. Weights & Biases handles experiment tracking and model governance. Scale AI manages data labeling pipelines and model evaluation workflows.
Glean builds enterprise search using embeddings of your internal documents — notable because it can be deployed on-prem or in your own cloud tenant. Moveworks automates IT service desk workflows using a combination of intent classification and LLM generation, with options for private deployment. Writer offers a fully private LLM deployment for enterprises that need model customization without data leaving their environment.
What decision-makers don’t realize: the infrastructure category is where the real architectural choices live. Choosing a SaaS productivity tool with embedded AI is not the same decision as choosing your AI infrastructure layer. One is a vendor relationship. The other determines what’s possible — and what’s auditable — for every AI deployment that follows.
The list of apps that use AI will keep growing. What won’t change is the basic question: when that AI runs, whose infrastructure is it running on, and who controls what it sees. Every enterprise needs a clear answer to that before the list gets any longer.