Many in the software industry assume that AI’s natural endpoint is autonomy: agents that replace workers, write code, or run workflows without human intervention. Robert Englander, republishing a piece from his blog with permission, argues that this focus may underappreciate a different, highly practical opportunity — natural language interfaces layered over deterministic systems.
What works today, and where it falls short
Generative systems and large language models already deliver value: they can speed up coding, help draft documents, summarize information, and accelerate repetitive tasks. However, LLMs are statistical machines: they sometimes hallucinate, improvise, or provide approximations. That behavior is acceptable in contexts that tolerate uncertainty — brainstorming, drafting, translation — but is problematic where correctness and determinism are required.
Financial calculations, scheduling, medical applications, and accounting demand precise, reliable results. In those domains, probabilistic suggestions can introduce unacceptable risk. Reliability remains the bedrock of useful software.
Natural language as a bridge between people and systems
Englander’s central point is that LLMs may find their most important role not by replacing deterministic systems, but by lowering the friction between users and those systems. Historically, users adjusted to the discipline of machines: learning command syntax, navigating menus and workflows, and filling forms exactly as the application required. Natural language interfaces invert that relationship: the system moves closer to how people naturally express intent.
Examples of conversational requests Englander mentions include:
- “Show me how delaying Social Security by two years impacts long-term spending.”
- “Transfer $500 from checking into savings next Friday.”
- “Why did my tax liability increase this year?”
- “Find the contracts signed after January that contain auto-renewal language.”
None of these queries removes the need for deterministic back-end systems; they depend on them. The natural-language layer acts as an interpreter between human intent and authoritative execution.
Two different capabilities that complement each other
Current language models are particularly strong at interpretation: extracting meaning from ambiguous human communication, maintaining conversational context, translating between representations, and helping users express intent more clearly. These are interaction problems.
Weaknesses typically appear in areas requiring guarantees, consistency, accountability, and deterministic correctness — the system-of-record problems. Industry discussions sometimes blur the line between these domains.
Far from diminishing the importance of deterministic software, conversational interfaces increase it. When users speak in natural language, the validation and execution layers beneath the surface become even more critical: systems must safely interpret intent, validate operations, preserve constraints, and maintain correctness when requests are conversational and ambiguous.
Why this reframes how we think about software’s future
Every major era of computing introduced an interface transition: mainframes with specialist operators, personal computers with graphical interfaces, the web with hyperlinks and forms, and mobile with touch and gestures. Natural language could be the next major abstraction layer — not because machines became human-like, but because we can now translate human communication into machine discipline at scale.
This perspective changes how we value autonomy. The current AI zeitgeist often treats autonomy as the inevitable destination: partial automation today implies full replacement of people tomorrow. Englander suggests the most durable value might instead lie in reducing interface complexity. For many applications, the friction isn’t execution — it’s expressing what the user actually wants the software to do.
Enterprise systems, financial platforms, creative tools and even simple applications often require significant onboarding before users become comfortable. Natural language interfaces can let software meet users where they already are: ordinary human communication.
Implications for design and engineering
If interfaces become conversational, underlying architectures must be more rigorous about validation and execution semantics. Ambiguity does not disappear; it moves into interpretation and validation layers. Engineers will need to build systems that clarify intent, normalize inputs, enforce policies, validate operations, and provide authoritative execution beneath the conversational front end. Probabilistic interpretation and deterministic execution are not competing ideas — they are complementary.
The likely outcome is greater accessibility from the conversational layer and preserved trust from deterministic back-ends. The transformative impact of LLMs may be greatest when they help people interact more naturally with existing systems rather than when they supplant those systems entirely.
Closing thought
For decades, people adapted to computers. Now software may finally begin adapting to people: accepting requests in natural language while rigorous, authoritative systems ensure correct, accountable outcomes. That combination — conversational interfaces over deterministic systems — deserves more attention than the single-minded pursuit of full autonomy.
