How to Bring GenAI into Operational Work

Code became a productive environment for models not despite its complexity, but because compilers, linters, tests, and repeatable execution provide fast signals. Each tool reports a violated syntax rule, contract, or logical expectation. In sales, support, and management, feedback is sparse, an outcome may surface a month later, and the counterfactual cannot be replayed. Autonomy therefore depends less on model intelligence than on an environment where actions can be evaluated.

Writing a procedure does not make the environment ready. In one experiment, experienced support specialists documented their process, yet the document explained only about 20% of the operators' actual decisions. The rest lived in tacit knowledge, conversations, and professional judgment. Before automating the work, a team must isolate stable stages, define inputs and outputs, recover exceptions, and agree on what correct behavior means in each case.

In support, generating an answer is not enough: the system must resolve the customer's problem, escalate at the right moment, and pass quality control. In accounting, calculations and explicit rules belong in deterministic code, while a model can interpret heterogeneous documents, classify transactions, and locate the relevant contract. The business case must still include review effort, corrections, penalties, and the error rate of the human process being replaced.

In marketing, the bottleneck is often not writing copy but waiting days for product, editorial, legal, and design approvals. A model creates more value as an early reviewer that moves institutional rules closer to the author than as an endless content factory. In building design, GenAI can explore alternatives and reveal savings while conventional calculations verify strength and compliance; asking it to produce the complete structure would add risk without eliminating verification.

Bondar recommends taking responsibility for a vertical slice from process input to customer outcome instead of inserting an LLM into one existing step. This exposes the real constraint: an accelerated activity may only move the queue downstream, while mandatory manual review can erase the savings. Sometimes the largest benefit appears before any model is deployed, when process analysis removes a redundant approval, an unused labeling operation, or an obsolete rule.

A practical screen asks four questions: which metric should move, how much the end-to-end result changes under ideal automation, how poor quality is detected today, and whether control will cost more than the benefit. Human in the loop is not a guarantee; a tired specialist can become a ceremonial reviewer. Successful teams have an owner who treats every agent failure as their own and can change the procedure, context, guardrail, or escalation path.