Meta and other tech giants are justifying recent layoffs with the same narrative: AI can do the work now, so humans don’t need to.
Yes, it’s true we’ve seen AI rapidly approach the level of human expertise. However, through our decade of work crafting AI tools that offer tailored interview prep, we’ve also seen why designing systems driven solely by AI is the wrong bet.
Nick Charles Triscritti is the Co-Founder of InterviewAI and an AI product leader and builder.
Andy Kurtzig is the CEO of Pearl.
Yet, on paper, AI looks like it has already achieved human-level performance. Microsoft recently used OpenAI’s advanced o3 model to “solve” more than eight out of ten New England Journal of Medicine test case studies, compared with a 20% success rate for human doctors.
Meanwhile, AI research non-profit METR predicts that by 2030, AI could “execute complex software projects” that today require weeks or even months of human expert labor.
But our replacement is not as inevitable as the AI industry would have you believe.
The companies that today are choosing to implement AI solely for cost savings are about to discover an expensive truth: the most effective AI systems are those designed around human expertise, not in its absence.
Permanently and intentionally incorporating human validation throughout AI workflows is what will unlock each subsequent level of AI functionality.
Where AI Systems Fail
AI accuracy gains are still hard-won. Each 10% improvement in foundational models requires $1 trillion in expenditure. At the same time, Stanford researchers have discovered that large language models may never be able to reliably distinguish between what is true and what humans believe to be true.
Translation: without human judgment in the system, AI will confidently scale blind spots. But there’s a failure point that will persist even if AI is perfect: often, the systems containing AI don’t use it properly.
Here’s what it looks like when a company goes all-in on a faulty AI-driven workflow:
A product management leader assigns an agent to independently build a weekly performance dashboard. Everything looks right, so the PM ships it with only cursory oversight. A few months in, leadership shifts ad spend toward a new product based on the chart. Next quarter’s sales tank. The cause: the agent had quietly counted free trial users as paying customers, which a human expert would have spotted in seconds. Instead, a full quarter of ad budget got burned because the role of human oversight wasn’t properly defined.
Or take a company using AI for hiring decisions. Without human input to fine-tune its selection, the model calcifies around outdated assumptions about what humans believe a successful leader looks like. It passes on multiple candidates who fall short on paper but bring the perfect combination of disparate skills; people who a seasoned manager would instinctively pull from the pile. Here, more codified routing rules for each application could have kept these candidates in the running.
These aren’t imaginary scenarios. We’re seeing them play out in our own work with AI systems; from how they still only variably account for rapidly changing medical guidelines to how they incorrectly assume they can move past human oversight steps.
And the consequences keep mounting as companies race to integrate AI into their core processes and consumers increasingly forgo expert consultations for AI advice. During a code freeze, an AI agent for Replit deleted an entire live database without human review, and AI-driven facial recognition sent a Tennessee woman to jail for five months for crimes she did not commit in North Dakota.
These incidents are the predictable cost of failing to integrate humans properly and instead treating them as an inefficiency to be optimized away.
Building AI-Native Workflows
Company leaders are pushing teams to just “rub some AI” on existing workflows to improve them. However, bolting AI on only worsens underlying problems as processes scale. To build AI that delivers on its growth promise, we must redesign workflows to inherently play to the strengths of AI, humans, and software.
The truth is not every task requires an LLM; there are situations that call for the predictability of routing, validation, and constraints. If everything is probabilistic, workflows become more like suggestion engines than reliable systems.
The best AI systems combine rigid and flexible processes and clearly define task ownership across three separate layers:
1. AI Flexibility: embeds probabilistic models that interpret more abstract requests and turn them into recognized actions.
2. Deterministic Software: adds hard-coded routing rules, constraints, validation triggers, and exceptions that direct operations toward the right pathway. This can include a context-aware RAG layer which pulls from a pre-determined “knowledge base” while also controlling memory and compression.
3. Humans For Context: continually assess model assumptions, change workflows based on external conditions, clarify human behavior nuance or overall ambiguity, and make final decisions.
AI’s flexibility identifies inputs that don’t quite match up with what the system might be expecting while the deterministic layer helps put up hard blocks on tasks it is unable to do, instead of placating the user.
At this ideal level of functionality, AI can make a recommendation of where to go next but let the human choose which path to take.
Redefining Human-In-The-Loop
Yes, AI is often inaccurate, but that doesn’t negate its potential. Frequently, it is a faulty system around AI that leads to the most dangerous breakdowns.
To improve AI integration, the goal is to go beyond adding a “human-in-the-loop.” If a human is only there to check an AI’s final output, that workflow is flawed. Instead, we should be designing AI systems where each part best contributes to a reliable outcome.
Eliminating humans from AI-driven processes today is akin to swapping a compounding asset for a one-time win. Instead, centering and clearly defining the role of human expertise is what will make AI systems better over time.
The companies that dominate the next decade won’t be the ones with the smallest org charts but those that properly integrate humans with software to drive AI advancement.
We feature the best IT automation software.
This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.
The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit
