To err is human — Latin proverb
To err is machine too — probably some data scientist
AI is probabilistic by design. It is imperfect, and it will never be 100% right. Can this imperfect, “usually-right” AI ever be trusted in business workflows where mistakes have real, costly consequences? This question is especially critical at Justworks, where we offer HR software, payroll, benefit access, tax compliance, and international hiring for our customers. Putting it another way, we work in domains where mistakes could be life-changing.
As software engineers, we respond to this risk by leaning on deterministic, rules-based systems. We don’t take chances: we meticulously consider business processes and craft them into code. Good code, by its very nature, is consistent: same inputs, same outputs, every time. We make this consistent code accurate through rigorous testing.
However, reality is messier: exceptions arise, business logic changes, and data arrives in unexpected formats. Deterministic code, if flawed, is also consistently wrong, repeating the error senselessly until human intervention. Moreover, it can only automate processes with structured, unambiguous data as input. The real world, meanwhile, is full of unstructured, ambiguous inputs: scans of documents, handwritten notes, and PDFs generated by countless legacy systems are an unavoidable part of the landscape, particularly with benefits and payroll. This mismatch creates an automation gap.
To bridge this gap, where interpretation and pattern recognition are needed, we traditionally place humans — our indispensable, intelligent, “organic interfaces” — between chaotic reality and structured code. This is paradoxical: however indispensable and intelligent humans may be, we are fallible, mistake-prone, and understandably, do not enjoy doing repetitive, mind-numbing tasks day in and day out. Meanwhile, today’s multimodal LLMs seem so capable and accessible that it begs the question: why not use them to ease this burden? The question is often met with concerns about the criticality of the process. As asserted before, critical workflows, where mistakes have real, costly consequences, demands 100% accuracy: therefore, the argument goes, AI must necessarily achieve 100% accuracy. If AI erred with your payroll, would “well, it’s usually right” suffice? Critics, and our instincts, would say “No!” The stochastic nature of AI seems fundamentally at odds with the deterministic needs of high-stakes business processes.
Critical business processes are already fault-tolerant
We agree with the caution that AI is not a drop-in replacement for rigorously validated code, nor can it be blindly trusted with critical tasks. However, we challenge the underlying assumption that machines must achieve 100% accuracy solely on their own. Consider the human role: people, not technology, administer benefits, process payroll, and ensure tax compliance. Technology assists. And, for the fallible yet indispensable humans, our processes already incorporate checks, balances, review steps, reconciliation procedures, and support mechanisms to catch and correct errors.
This reframes AI’s potential. If systems already accommodate human errors, is it not worth considering that these systems could also accommodate machine errors, since AI drastically cuts manual effort and related human errors? The answer is yes: stochastic models can thrive in deterministic environments precisely because our human-centric designs already provide safety nets.
Let us be clear: this isn’t about replacing people. It’s about smart augmentation. Instead of demanding unattainable perfection from AI, we should ask a more pragmatic question: “Can AI boost efficiency and reduce specific error types within our existing, human-verified, fault-tolerant workflow?” Recognizing that processes are already designed to handle imperfection unlocks enormous potential for probabilistic AI to automate the drudgery, enhance human capabilities, and free teams for high-level judgment and complex problem-solving — the tasks where their expertise truly shines.
So, yes, probabilistic models have a vital role to play, even when the demand is for deterministic outcomes. It turns out, the secret to unlocking AI’s power in critical business functions isn’t just about better models; it’s about better partnership with the humans those models are designed to serve. And ironically, it’s the inherent fallibility and adaptability designed into our human-centric processes that makes this powerful collaboration possible.
Interested in working with us to unleash the potential of probabilistic models on processes that help small businesses work fearlessly? Check out our Careers page!
