Navigating the AI Transition - What Happens to "Old Tech"?

A Moment of Pause in the Age of AI

Spending time looking at solutions that are pushing the boundaries of what’s possible with AI—where it can automate everything from summarizing alerts to provisioning user access to dynamically building entire apps—makes me pause and ask: what exactly happens to "old tech"? Not just the apps we use, but all the scaffolding that surrounds them—the glue scripts, dashboards, batch jobs, and approval chains. Do these fade away quietly into LLM-powered interfaces? Or do they stick around under the surface, wrapped by smarter systems until they’re no longer necessary?

More importantly, how does this "replacement stack" propagate? Is this a clean sweep or slow seep? Does AI replace software by function? By interface? Or do old habits and tools lingering until newer ones just feel easier, more natural, more adaptive?

The Application Layer as We Know It

At its core, an application is software that performs a specific function for an entity—whether it’s a person or another system. In enterprise software, these applications typically live in a layered model: user interfaces on top of business logic, on top of databases. CRUD (create-read-update-delete) operations power everything from sales dashboards to HR portals. Users click through menus and forms, triggering backend processes that shuffle data around.

Now, contrast that with where we’re headed with AI. If an application is “a specific function for an entity” and an agent is “a skill,” you start to see how systems of agents could go much further. As long as business logic and need are clear and explainable—in natural language—we’ve made enough progress to credibly say that the hypothesis of "AI will replace applications" is no longer ludicrous. Instead of a human navigating through dropdowns, you have a system of agents interpreting natural language goals, fetching data across systems, and assembling workflows on the fly. With enough progress in memory, context awareness, and orchestration, these AI agents might not just complement applications—they may become the next evolution of them.

What These AI Agents Need to Work

The system of agents – I don’t know what it’ll be called in 3 years – an AI powered application? Agentic app? AI employee? But here’s what I think that’ll need.

Self-prompting: Agents that know how to break down ambiguous tasks into subtasks, in the right order, and make intelligent decisions about what comes next. Think of a junior analyst who figures out that “generate insights” means pulling data, correlating it with recent incidents, ranking risk, and presenting options.

Context awareness: This isn’t just about loading Slack threads into an LLM. It’s about understanding your org’s specific language, workflows, and risk tolerances. Whether “incident” refers to a security breach or a supply chain glitch depends on the company. Agents need real-time telemetry, historical data, and embedded business logic to make intelligent decisions in context.

Memory: Most public LLMs today have session-based memory. But what we really need are persistent, layered memory models: working memory for short-term reasoning, episodic memory to remember past interactions, semantic memory to store org-specific knowledge, and procedural memory to retain workflows. Done well, this lets agents act more like teammates than tools.

Orchestration: In a system-of-agents world, different agents handle different skills—enrichment, triage, report generation. The magic lies in orchestration. Some agents might operate in a sequential flow, others in a parallel mesh where they negotiate outcomes and pick the best response. It’s less about a single genius agent and more about a team working in sync.

The Post-UI Era

This does change how we think about apps. Traditional apps require engineers to predefine every flow and error state. Agentic systems reason. They can handle edge cases. You don’t click through a report builder—you just say, "What underperformed last week and why?" The agent figures out how to get you the answer.

That's what people mean by "post-UI." Intent replaces interface. The tool isn’t a rigid system anymore—it adapts.

What’s Getting Replaced (just an opinion)

When we talk about AI agents potentially replacing applications, it’s worth unpacking what that might actually mean. It probably isn’t as simple as swapping out one software tool for another. If anything, the idea of “replacement” seems more about shifting the stack and the way we interact with systems—less about the application itself and more about what sits underneath it: glue code, approval workflows, dashboards, scripts, and the layers of coordination that tie everything together.

A lot of enterprise IT today still revolves around stitching. Custom scripts to shuttle data between systems, ETL pipelines, spreadsheets passed around for approvals, dashboards that require manual checks, notification rules that ping the right people at the right time. It makes me wonder: if AI agents can access the right data sources and tools, could they absorb a lot of that glue? Maybe not today, maybe not fully—but the shape of it is starting to come into view.

You could imagine a world where, instead of a data engineer maintaining a scheduled job to transform and load data for reporting, an agent simply pulls fresh data when asked. Instead of an analyst checking a dashboard every morning, an agent watches the metrics, flags anomalies, and proposes corrective action. The shift isn’t just in who executes the task—it’s in how the interface changes, from dashboard and dropdowns to something more conversational or goal-oriented.

In that framing, it’s not the business process that disappears—it’s the infrastructure beneath it that gets reimagined. You're not eliminating access reviews, you're replacing the spreadsheet and brittle workflow engine behind them with a more fluid, agent-driven layer. The goal stays the same; the means just evolve.

None of this is absolute or inevitable. But the idea that we’re moving from rigid interfaces and hardcoded glue toward more dynamic, intent-driven systems feels increasingly plausible. Maybe the future of enterprise software isn’t “apps” in the traditional sense—but something more modular, more agentic, and much more “tools adapting to needs” than “users adapting to tools”.

What’s Staying (Also an opinion)

Amid all the excitement around AI agents potentially reshaping the application layer, I think it’s worth grounding the conversation in a simple reality: some layers of enterprise infrastructure aren’t going anywhere. At least, not anytime soon.

Large organizations have made deep investments in the foundations—identity systems, data platforms, orchestration layers—that everything else runs on. And any AI-powered system worth using will need to build on top of those foundations, not replace them.

Think about identity providers like Okta or Active Directory. User lifecycle systems, data stores like Postgres and Snowflake, orchestration platforms like Kubernetes or ECS, observability stacks like Datadog or Splunk and core security infrastructure. These aren’t going to vanish because we have a more conversational interface. An agent might change how we interact with these tools—but the core functionality they provide is still essential.

In fact, it feels like clean data becomes even more important in an agent-driven world. If agents are meant to reason, act, and automate across systems, they’re going to need consistent, well-maintained inputs. So yes, agents might streamline how we interact with the stack. They might take over parts of the glue, the UI, the coordination. But the stack itself and the data that powers it? It’s still the foundation. And it still matters.

To be clear, this doesn’t mean we won’t see innovation—or that AI won’t influence how we build these layers going forward. Of course it will. A company that can harness AI to ship faster or operate more efficiently will have an advantage over one that can’t. And sure, we might see organizations deciding to build certain components in-house rather than buying point solutions—especially if AI makes that path more viable. But that’s true of any major shift.

All I’m really saying is: AI alone won’t be a force multiplier in every part of the stack. The deeper you go into core infrastructure, the more defensibility and architecture will matter. And in those layers, simply having “AI” won’t be enough.

The Dissonance Between AI Possibility and Enterprise Reality

While the technical breakthroughs in AI are thrilling—and in some cases, genuinely transformational—it’s important not to get too swept away. The truth is, most enterprises don’t live on the cutting edge. They operate in a world defined by legacy systems, manual workflows, and deeply entrenched operational processes.

And that’s where, some enthusiastic technologists miss the mark. It’s tempting to imagine a world where every workflow is agentified, every decision path inferred, and every process rebuilt from scratch with LLMs and APIs. But let’s ground that against some sobering realities:

Most organizations, even in the most advanced economy in the world, are still coming to grips with the last paradigm shift. As of 2024, only ~30% of North American companies have the entirety of their workloads run in public clouds (AWS, Azure, GCP). That number has grown steadily but is slowing, according to Flexera and Uptime Institute surveys. Most organizations (70%+) still operate in hybrid environments, where legacy on-prem systems coexist with newer SaaS tools. At most large banks, the most critical applications are still hosted on-prem and adoption of modern cloud-native platforms remains uneven.

And it’s not just with core infrastructure, but more generalizable to function specific solutions –many companies still manage processes manual scripts, internal tooling, and spreadsheets. Let’s take access reviews, for example. In many mid/large enterprises, these are still handled in spreadsheets. GRC checklists are passed around by email. Onboarding workflows often involve HR exporting lists from SAP, emailing IT, and hoping that accounts get created in time. In fact, a 2023 survey found that over 60% of identity and access reviews are still done manually—often via email or Excel. A McKinsey report from 2022 similarly noted that 70% of GRC workflows globally remain spreadsheet-based, despite the availability of commercial GRC tools.

These aren’t exceptions. They’re the norm.

And it’s not because enterprise buyers are slow or uninformed. It’s because they have valid reasons to be cautious:

• Legacy complexity: Most enterprises run dozens—often hundreds—of deeply integrated systems. Replacing even one can have ripple effects across the stack.
• Regulatory pressure: Financial, healthcare, and critical infrastructure orgs face real constraints. They need traceability, auditability, and explainability. A black-box AI, no matter how magical, won’t pass that bar.
• Operational risk: A small error in an AI-driven access decision isn’t just an inconvenience—it could mean a breach, a fine, or worse.
• Cultural inertia: Systems don’t change without people. Retraining, reassigning, and reworking cross-functional responsibilities isn’t a side project—it’s an organizational transformation.

So no, these buyers aren’t laggards. They’re rational actors making risk-adjusted decisions. They’re focused on ROI. On what works. On what fits. The gap between what’s possible and what’s deployed isn’t a failing. It’s a reflection of real-world complexity.

Why Buyers Are Right to Be Skeptical

Enterprise buyers have been through waves of over-promising: cloud, no-code, blockchain, RPA, AIOps. Many of those categories are valuable—but each came with its share of vaporware and failed rollouts.

They’ve learned to ask three questions before they adopt anything new:

1. Does it clearly save time, improve accuracy, or reduce risk?
2. Can I adopt this incrementally without re-platforming my stack?
3. Will I be able to explain it—to my engineers, my board, my auditors?

If the answer to any of those is no, they’ll wait.

And that’s fair.

Because they know they’ll be accountable when the “AI strategy” doesn’t deliver.

The Equilibrium Ahead

This dissonance between what AI can do and what enterprises are ready for? It’s not a flaw. It’s a natural part of the curve.

This isn’t a revolution. It’s a transition. Will it take three years? Five? Hard to say. But IT IS ALREADY happening.

And the products that win won’t assume the future has already arrived. They’ll help customers navigate the messy middle. They’ll build the bridge—and walk across it with them.

We’re seeing it already:

• A Fortune 500 company using agents to reduce alert triage from weeks to hours.
• A payroll firm using agents to correlate vulnerabilities with asset risk and ownership.
• Risk and compliance agents that help users surface insights faster across distributed data.

These aren’t hypothetical. They’re live. They’re working. But they’re happening slowly, carefully, and with clear ROI.

How Work Gets Rewritten

The companies that succeed in this transition will:

• Solve real, painful problems today.
• Integrate with what’s already in place.
• Build trust through transparency, not just speed.
• Stay relevant tomorrow—by evolving as their customers evolve.

So if you’re building in this space: optimism is good. Grounding is better.

Meet customers where they are. Help them transition. Deliver outcomes—quietly and effectively.

Because eventually, the new workflows won’t just be smarter. They’ll be easier. More flexible. More trustworthy. The old ones—those spreadsheets and brittle scripts—won’t be replaced because they failed. They’ll be replaced because they’re no longer worth maintaining.

That’s not a flashy AI story. But it’s the one that will actually get deployed.

And when it works, it won’t just replace software. It will quietly, patiently, and fundamentally change how work gets done.