Why Your Software's UI Is About to Become Irrelevant

Microsoft Word has over 4,500 commands accessible via the UI. That number comes from the Wall Street Journal in 1996, and it’s only grown since. Think about that for a moment. A tool designed to help you write documents has thousands of commands, each added for a reason, each making sense at the time. But the cumulative weight of four decades of software evolution has created something that demands expertise just to find the option you need.

This isn’t unique to Word. It’s the story of enterprise software.

Salesforce has grown from a simple contact database into a labyrinth of clouds, objects, workflows, and automation rules. SAP S/4HANA implementations average 18 months, according to ASUG research, and routinely run 30% longer than planned. Large enterprises now manage an average of 275 SaaS applications. Each one has its own settings screen, its own menu structure, its own logic that made perfect sense to the team who built it.

We’ve been solving the wrong problem. We kept asking “how do we help users navigate complexity?” when we should have been asking “why are we making them navigate at all?”

The Inversion

For 40 years, we built software with a simple assumption, humans translate intent into interface actions. You know what you want. The software provides buttons, forms, and menus. Your job is to find the right sequence of clicks to achieve your goal.

This assumption shaped everything. We built increasingly sophisticated navigation systems, contextual help, training programmes, and certification courses, all to help humans get better at translating their intent into software actions.

Agents invert this entirely.

Instead of humans learning to speak software, software learns to understand humans. Instead of navigating to Settings → Integrations → API → Create New Key → Set Permissions → Generate, you say “I need an API key that can read customer data but not modify anything.” The agent knows what that means. The agent knows where the settings are. The agent handles the clicks.

This isn’t just a better interface. It’s a different paradigm. The burden of complexity shifts from the user to the system. Every action a human could take through a UI becomes a tool the agent can call. The UI doesn’t disappear, it becomes the agent’s toolkit rather than your obstacle course.

What This Actually Looks Like

We’ve spent the last six months building exactly this.

We are about to release inmydata Studio, an AI powered dashboard builder. Today, creating a simple sales dashboard is an exercise in UI endurance. You open a dashboard designer. Pick a data source from a list of forty. Find the field selector. Drag Revenue into values. Drag Month into categories. Open the chart menu. Choose a line chart. Open the formatting panel. Pick a colour scheme. Open the filter panel. Set a date range. Then realise you also need Region, delete the chart, and start again. You repeat this six more times. Then you manually arrange everything into a layout and save the dashboard. And when you notice you forgot to add a comparison period, you reopen every single chart and do it all again.

Now you just say “Build me a sales dashboard showing revenue trends by region, compare to last year, highlight any regions underperforming their targets.” That’s it. The agent understands your data model, knows what fields are available, creates an appropriate layout, selects suitable chart types, and handles the visual design.

But here’s where it gets interesting. Traditional dashboards are dead endpoints, you look at them, and if you want to explore further, you’re back to building another report. In inmydata Studio, when you click on a chart element, you enter a conversation.

Click on “North region” showing declining revenue. The agent responds “I can see you’re exploring the North region’s revenue decline. I’ve created a breakdown by product category. It looks like Enterprise Software dropped 34% in Q3. Would you like me to show customer-level detail, compare North’s performance to other regions, or break this down by sales rep?”

The drilldown isn’t a predefined hierarchy someone configured in advance. It’s a conversation. The agent generates charts on the fly based on what you’re trying to understand.

Behind the scenes, this works through a two-phase architecture. In the design phase, the AI agent analyses your request, understands your data schema, and creates a structured plan, what charts, what fields, what layout. In the render phase, deterministic Jinja2 templates execute the plan with live data. No LLM calls, instant refresh.

This separation matters. AI handles the creative, ambiguous work of understanding intent and designing solutions. Deterministic code handles execution, ensuring reliability and speed. Every dashboard refresh is instant and consistent because the LLM did its work once, during design.

The Giants Are Already Moving

The largest software companies in the world are racing to rebuild their products around this paradigm.

Salesforce has bet the company on it. At Dreamforce 2025, they announced Agentforce 360, calling it “the world’s first platform designed to connect humans and AI agents in one trusted system.” Marc Benioff told Bloomberg: “AI is doing 30 to 50% of the work at Salesforce now.” Reddit, using Agentforce, deflected 46% of support cases and cut resolution times from 8.9 minutes to 1.4 minutes.

Oracle unveiled 600+ role-based AI agents across finance, HR, supply chain, and customer service at AI World 2025. Their new “Ask Oracle” interface replaces traditional navigation with natural language. NetSuite Next now lets users “retrieve data, execute actions, and trigger workflows without navigating menus.”

SAP announced Joule is now omnipresent across the entire SAP portfolio. They’ve released RPT-1, a foundation model trained specifically on structured ERP data, designed not to generate fluent text, but to generate correct numbers. And through SAP AI Core, they’re integrating multimodal capabilities, a building supply retailer now goes from blueprint image to SAP quote in minutes, with AI interpreting the architectural drawing and generating a bill of materials automatically.

These aren’t incremental improvements. They’re architectural shifts. The menu structures, form layouts, and workflow designers that took decades to build are being bypassed entirely.

How This Plays Out

Not every category faces the same disruption timeline. The key variable is interface complexity, the more convoluted the current UI, the faster agents can disrupt it.

Business Intelligence and Analytics faces immediate pressure. The traditional model requires users to understand data modelling, know which fields to select, configure joins, and design visualisations. Natural language collapses all of this. “Show me which products are trending down in margin” is vastly more accessible than navigating a BI tool’s interface.

ERP Configuration is similarly ripe. SAP implementations take 18+ months partly because the system has so many options, each requiring expert knowledge to configure correctly. When an agent can interpret business requirements and configure the system directly, the implementation model transforms.

Core Systems of Record face slower disruption. Not because agents can’t improve them, but because they’re embedded in workflows, integrated with dozens of other systems, and carry years of institutional data. Replacing Salesforce isn’t as simple as preferring a better interface.

Here’s my prediction, existing systems of record won’t be replaced. They’ll be hollowed out.

The data stays in Salesforce, SAP, or Oracle. The business logic stays there too. But the interface layer, the menus, forms, and workflows that humans interact with, gets bypassed by agents sitting on top.

This is already happening. Startups are building what Bessemer Venture Partners calls “AI Trojan horse features”, tools that tap into the flow of data without requiring customers to rip out their system of record.

The pattern looks like this. First, an AI layer integrates with your existing CRM via API. Users start asking it questions instead of navigating the interface. Then the agent gets new capabilities, create records, send emails, schedule tasks. Gradually, it can do most of what the underlying software does. Users stop opening the main application. They interact through the agent. The traditional UI becomes a legacy admin interface.

The protocols enabling this are maturing fast. Anthropic’s Model Context Protocol standardises how agents connect to external tools and databases. Google’s Agent-to-Agent Protocol enables cross-platform agent collaboration. These are becoming the HTTP of the agentic era, the foundational standards that make everything composable.

What This Means If You’re Building Software

If you’re building software today, the implications are significant.

Design for agents, not just users. Every capability should be exposed as a tool the agent can call. If a human can do it through your UI, an agent should be able to do it through your API.

Think about intent, not actions. Traditional UX asks “what actions can users take?” Agentic UX asks “what outcomes do users want?” The agent figures out the actions.

Separate design from execution. AI is good at understanding requirements and creating plans. Deterministic code is good at reliable execution. Don’t use an LLM where a template will do.

Embrace the hybrid. The UI doesn’t disappear. It becomes the feedback layer, showing users what the agent did, surfacing confirmation for risky actions, providing visual results. The agent navigates, the UI renders.

The Timeline

How fast does this happen?

The infrastructure is ready. MCP and A2A are deployed. The major vendors have shipped their first agentic releases. The models are capable.

The bottleneck is integration. Connecting agents to legacy systems, building reliable tool APIs, handling edge cases, maintaining security boundaries. Enterprise adoption is measured in implementation cycles, not press releases.

My estimate, within a year, agent-first interfaces will be the default expectation for new software, Salesforce, Oracle, and SAP have already made that bet. Within three years, every major enterprise system will have an agentic layer. Within five years, training someone to navigate traditional software menus will feel like teaching them to use a command line, useful for experts, but not the primary interface.

The 4500 commands Word application won’t disappear overnight. But it will become the exception rather than the rule. The settings screen, the menu hierarchy, the form with 40 fields, these will feel as dated as DOS prompts.

Software spent 40 years accumulating complexity. Agents collapse it in one conversation.