Library 5 articles
Structured explanations of models, algorithms, systems and engineering decisions.
An AI agent can prepare documents, check data, and perform limited actions in ERP. But access to the system is not the same as authority to make any decision. We break down how to separate recommendations, execution, and approval, which operations require human control, and why the "Approve" button by itself does not ensure oversight.
Reliable automation is possible primarily for verifiable administrative operations and supporting tasks, where AI searches, extracts, cross-checks or prepares a draft, but does not determine the outcome for a person. Candidate selection, employee evaluation, promotion, dismissal and the allocation of opportunities require separate risk classification, valid criteria, group-level control, logging, real human review and the right to appeal. Some scenarios, including emotion recognition in the workplace from biometric data in the EU, may be directly prohibited. The article shows how to choose automation scenarios and which changes to process, data, control and accountability are needed before deployment.
AI automates the production of many technical artifacts, but it does not take over the engineering function as a whole. The engineer still defines the goal and requirements, chooses the computational mechanism, specifies admissible states and constraints, evaluates the consequences of failure, assembles evidence of correctness, and is accountable for integration and operation. The delegation boundary does not run between “simple” and “complex” tasks, but between results that can be independently verified and safely rolled back, and decisions where an error is hard to observe, spreads widely, or is expensive.
For numerical classification, regression, scoring, and forecasting on financial tables, classical statistical models, Random Forest, and especially GBDT should usually be evaluated before LLMs. They are preferable with limited samples, a fixed feature schema, and strict requirements for arithmetic accuracy, calibration, latency, reproducibility, and auditability. LLMs provide more value at the language boundary of the task – working with reports, natural-language queries, and SQL or Python orchestration – but that is a comparison of systems, not just models.
A method for evaluating an AI agent is explained through a teaching agent that prepares payments: from authority boundaries and observable success conditions to test results and a decision not to approve deployment. The article also shows why usefulness, blocked attempts, actual violations, and the reliability of human approval cannot be reduced to one metric.