Articles

Notes on building AI for work where being approximately right isn't good enough. Payroll, HR, and regulated documents, mostly — where the interesting problems are about knowing when a result can be trusted, not about making models bigger.

Truth Has a Date — HR systems have used effective-dated data for decades. AI knowledge architectures still answer today's question when you asked about 2023.

AI in HR: Where LLMs Help, Where Rules Matter and Where Humans Still Belong — Not every problem is an AI problem. A map of which technology fits which kind of work, and where accountability has to stay with people.

Why HR Needs Both Rule-Based Systems and LLMs — I asked the same model the same payroll question five times and got five different answers. None was wrong. That's the problem.

Which business documents does Document AI actually work for? — Some documents record facts you can check against something real. Others make a case. That distinction decides whether the output can be held to a standard of correctness at all.

The problem with AI economics — Pilot ROI looks impressive at fifty users. Enterprise AI behaves less like fixed-cost SaaS and more like a variable utility with external dependencies.

Agentic AI Governance: The Missing Layer — Governance frameworks scrutinise what the agent decides. Far less attention goes to what the agent reads before it reasons, and how those inputs became facts.

A question for HR leaders deploying AI — A vendor says 95% accuracy. Measured how, against whose ground truth, and where do the errors fall? An honest question about how HR validates the tools it buys.

The AI feature that sounds good—until you ask one question — "Our product gets better every time you use it" sounds like continuous improvement. Better for whom, and using whose data?

Five questions every payroll professional should ask before trusting AI with compliance work — Payroll checks everything twice by instinct. The same discipline should apply to the AI tools now arriving in the process.

Deployment-Agnostic AI: Why Training and Deployment Matter for Trust in Regulated Industries — Compliance teams are really asking two questions: will you train on our data, and will our data leave our infrastructure? Most approaches answer only the first.

Architecture vs. Retrofits: Building Document AI for Regulated Industries — Most vendors manage the privacy tension with data-processing agreements and anonymisation pipelines. Those are retrofits on architectures never designed for regulated use.

The Myth of the Single Source of Truth — Fragmentation is treated as a temporary state to be migrated away. In practice it's the operating reality, and it's where documents become the final reference point.

Why We Build Bounded AI — Generic AI optimises for breadth. In payroll and finance, breadth means you can no longer tell when a result is trustworthy.