Enterprise AI Applications and Copilots

Enterprise AI applications are where the money spent on chips, clouds and models is supposed to come back as productivity, and they are where the evidence is thinnest. They come in four shapes: copilots that sit beside a worker inside software the worker already uses (Microsoft 365 Copilot, Gemini for Google Workspace, Salesforce's Einstein and Slack AI); agents that take whole tasks off a worker's hands (Salesforce Agentforce, ServiceNow's AI agents, Sierra and Decagon in customer service); knowledge and search layers that let an assistant reach a company's own documents (Glean, Microsoft Graph, Writer); and build-your-own platforms on which companies assemble internal tools from model APIs (Azure AI Foundry, Amazon Bedrock, Google Vertex AI, Palantir AIP). Enterprise generative-AI spending grew from about $11.5 billion in 2024 to about $37 billion in 2025 by Menlo Ventures' estimate (December 2025), and Microsoft reported more than 30 million paid Microsoft 365 Copilot seats in July 2026. Yet the most quoted study of the period, MIT NANDA's GenAI Divide (August 2025), reported that about 95% of organisations it studied saw no measurable return from their generative-AI pilots. The two facts are compatible: spending is buying access, and return depends on whether time saved is turned into lower cost or more output. An enterprise AI business is therefore an answer to four questions: whose workflow does it sit in, who owns the data it needs, how is it priced relative to the value it creates, and can the buyer prove that value after the pilot? This section maps the product types, the incumbent-versus-AI-native contest, pricing (seat, usage, agent and outcome), adoption and ROI evidence, procurement and governance, the role of integrators, landmark cases and pitfalls. A worked example computes a stylized copilot rollout's return, the seat-versus-usage choice and per-conversation versus per-resolution agent pricing, and a checklist sets out what a buyer or investor should ask.