Enterprise AI spend hit $37 billion in 2025. That's 3.2x growth year on year. 88% of organisations are using AI in at least one function. Gartner says 40% of enterprise applications will embed AI agents by 2026. By 2030, 35% of SaaS point products will be replaced by AI agents or absorbed into larger agent ecosystems entirely.
Those numbers are worth sitting with. Not because they're big, but because of what they imply about user interfaces. Every one of those AI capabilities needs a way to interact with a human. And the way most software handles that interaction today is embarrassingly shallow.
I've been mapping the progression of how AI shows up in product interfaces. Not the marketing version. The architectural version. What's actually happening inside the products, and what it means for how people work with software. The pattern that emerges is a maturity model with four distinct levels. Each one represents a fundamentally different relationship between the user, the interface, and the intelligence behind it.
The first level is AI as a feature. Same interface, new button. The user explicitly chooses to invoke an AI capability within an otherwise unchanged product. This is where roughly 90% of SaaS lives today. Google Ads' AI Max auto-generates headlines and descriptions from landing page analysis. Meta's Advantage+ creates AI-generated creative variations and handles audience targeting. HubSpot's Breeze adds 80-plus AI features across content generation, prospecting, and customer agents. Canva's Magic Studio has been used over ten billion times. Adobe's GenStudio codifies brand rules through Firefly StyleIDs and produces on-brand content at scale.
The pattern is consistent. The product interface is unchanged. The user invokes, the user approves. AI accelerates a discrete task. Copy, images, targeting, one step at a time. The workflow itself is structurally identical to what it was before. This is table stakes by 2026. If your product doesn't have this, you're already behind.
The second level is where things get interesting. Ambient AI. The AI is no longer waiting to be invoked. It continuously monitors, detects patterns, and surfaces things to the user before they think to ask.
The gold standard here is observability. Datadog's Watchdog continuously monitors infrastructure, auto-detects anomalies and root causes without any configuration required. PagerDuty's AIOps runs always-on machine learning for proactive incident detection and achieves 87% noise reduction. Tableau Pulse detects hidden drivers, trends, and outliers, then pushes personalised digests via Slack and email.
This is the shift from reactive to proactive. The user doesn't ask. The AI brings things to them.
Within ambient AI, there's actually a progression of its own. At the basic end, you have detect and alert. Something unusual happened. Mixpanel, Amplitude, that kind of thing. Then diagnose and alert, where the system tells you what happened and why. Datadog Watchdog, Dynatrace. Then diagnose and recommend, where it tells you what to do about it. Salesforce Einstein, Gong, Tableau Pulse. And at the most mature end, diagnose and act. The system already fixed it. Meta's Advantage+ dynamically shifts budgets at the impression level in real time. Madgicx autonomously shifts spend toward best-performing ad sets and detects creative fatigue before it impacts performance. Fully autonomous.
The architectural shift here is significant. Level two requires an event-driven AI layer sitting across the entire data model. Continuous data access across all streams. Pattern recognition through ML baselines, not static rules. Push-based delivery through Slack, email, webhooks, in-app notifications. And pre-programmed actions so the system can act, not just alert. This isn't a feature you bolt on. It's a different architecture.
Level three is where the interface itself changes. Natural language becomes the primary interaction mode. And there are two distinct variants here that I think are worth separating.
Level 3a is the agent-navigated UI. A conversational agent orchestrates the existing product interface. The product still exists. The screens are still there. But the agent navigates them for you.
Mutinex's MAITE is a compelling example. It's a chat-based interface over marketing mix modelling data. Natural language in, charts and C-suite slides out. It reverses the workflow entirely. Question first, answer assembled. Google's Ads Advisor uses Gemini to deliver personalised answers, campaign analysis, and troubleshooting through conversation. Microsoft Copilot operates as a side panel chat across Office 365, manipulating the existing UI on your behalf. Salesforce Agentforce is a conversational copilot across CRM with multi-turn context grounded in Data Cloud. ServiceNow's AIx provides a conversational front door with multimodal input and AI Web Agents that navigate third-party applications.
The product still exists behind the conversation. The agent is a navigator, not a replacement. That distinction matters for what comes next.
Level 3b is the north star. I'm calling it agentic UI. The traditional interface is substantially abstracted. The agent doesn't navigate existing screens. It composes the experience dynamically from a registry of capabilities.
The user states an intent. The agent composes the experience. The agent orchestrates tools and services. The outcome is delivered. This is a fundamental decoupling of the demand side, what needs to be done, from the supply side, the UI components, data sources, and actions available.
This isn't theoretical. The protocol stack to make it real is emerging right now. Anthropic's Model Context Protocol handles agent-to-tool connection, and it's already adopted by OpenAI, Google, and the Linux Foundation. Google's Agent-to-Agent protocol handles inter-agent coordination. CopilotKit's AG-UI connects frontend to agentic backends. Google's A2UI protocol, announced in December 2025, lets agents describe UI that the client renders. And Anthropic and OpenAI's MCP Apps enable interactive UI directly within agent conversations.
Together, these form a composable stack analogous to TCP/IP for the internet. Agents can access tools, coordinate with other agents, render interfaces, and interact with users in real time.
The examples are already appearing. OpenAI's Operator and Computer Use Agent are browser-based agents that see and interact with any website, now integrated into ChatGPT. Manus AI runs fully autonomous sessions on dedicated cloud VMs using twenty atomic tools to control a virtual computer. In ad tech specifically, PubMatic's AgenticOS provides an operating system for agent-to-agent execution in ad environments, achieving 87% less setup time. The IAB Tech Lab is extending OpenRTB, AdCOM, and VAST with MCP, Agent-to-Agent, and gRPC. Meta plans that by end of 2026, you'll input a business URL and AI will create the entire ad, decide targeting, and optimise budget. Full automation.
On the frontend, Vercel's AI SDK maps tool results to React components for dynamic UI rendered from conversation context. CopilotKit combines AG-UI, A2UI, and MCP to build in-app AI copilots that read app state, suggest actions, and modify the interface in real time.
The architectural vision is straightforward. User intent expressed in natural language. An AI agent that understands the intent and composes the experience. A protocol layer connecting everything. A capability registry containing data, actions, analytics, UI components, and other agents. And a rendered experience that's dynamic, composed, cross-platform, and personalised by nature.
No fixed UI to learn. A new MCP server equals a new capability. Cross-system by default. This is where the puck is going.
The competitive context sharpens the urgency. AI-enhanced products, the wrappers, bolt AI onto existing architecture. Same UI, new buttons. Static workflows with AI features. New capabilities require UI changes. Per-seat pricing. AI-native products design with AI at the core. Intent-driven, context-aware interfaces. Continuously learning and self-improving architectures. New capabilities added as tools and skills with no redesign required. Usage or outcome-based pricing.
Gartner notes that only about 130 of thousands of claimed agentic AI vendors actually offer legitimate agent technology. The gap between real and claimed is enormous.
Each level delivers value on its own. They're incremental, not all-or-nothing. Level one is table stakes by 2026. Level two is differentiation today. Level 3a is the leading edge. Level 3b is the strategic north star. The journey is incremental. Start where you are, build toward where the market is going.
The numbers tell you the trajectory. Gartner says 33% of enterprise software will include agentic AI by 2028. Deloitte reports 74% agentic AI adoption within two years, up from 23%. McKinsey estimates $2.9 trillion in economic value unlocked by 2030 through agent-redesigned workflows.
Bain's 2025 Technology Report put it plainly. Today's tech giants have proven unusually resistant to disruption, co-opting it through self-reinvention. Yet AI, with its ability to transform work processes and the unprecedented speed of its adoption, is this decade's disruption.
The competing platforms are not standing still. The window to lead is now.
The question is not whether. It's who leads it.
Damien Healy is the founder of Qanara, an Australian AI consultancy helping businesses accelerate from strategy to impact. He writes about AI-native workflows, frontier AI capabilities, and practical transformation.