The Warehouse Scheduler That 'Thinks Like a Planner'

Calling it 'AI' is defensible only in the broad, historical sense. CogniFlow is a rule-based expert system: human experts encoded domain knowledge as if-then rules, and an inference engine applies them. That places it in the symbolic AI/GOFAI tradition that dominated early AI, not in machine learning. It is narrow: it handles replenishment scheduling in a fixed domain and has no general reasoning. It is not machine learning because no parameters are estimated from data and performance does not improve through exposure to examples; the 1,850 rules are manually maintained. The label matters because 'AI' is an umbrella term, while the grant requires ML specifically, so the product would fail that eligibility test. The 22% override rate and 60 conflicting orders in a +30% demand week are evidence of brittleness outside the encoded envelope, a classic weakness of rule-based systems. Procurement should ask for the rule-maintenance plan and stress-test results, not rely on the marketing label.

The blind test is an imitation game, not a definition or proof of intelligence. A 7/10 result on one 30-minute scheduling exercise is weak evidence: the sample is small, the judges are planners who may share the system's heuristics, and the task is narrow. The Turing test asks whether a machine's responses are indistinguishable from a human's; it measures behavioural resemblance, not understanding. Searle's Chinese room makes the point that rule-following can produce correct-looking outputs without comprehension: CogniFlow manipulates symbols according to human-written rules, but it does not know what a replenishment order means. The holiday failure is more informative than the blind test: when demand moved 30% outside the encoded conditions, the system produced conflicting orders. That shows narrow competence, not general intelligence. A better evaluation would use held-out stress scenarios, adversarial cases, and operational metrics such as override rate and stock-outs, not a single imitation exercise.

Separate the labelling question from the value question. (1) For the grant: do not certify it as ML; it is a rule-based expert system, so either seek a different grant or ask the vendor for an actual learned model with training data, validation method, and model cards. (2) For operations: treat CogniFlow as decision support, not autonomous control. Require a rule audit, a change-control process, and a stress test at +30% and +50% demand with conflicting-order detection. (3) Use the 22% override data: analyse where planners disagree, because those cases define the system's operating envelope and may reveal missing constraints. (4) Contract for transparency: rule documentation, versioning, incident reporting, and exit rights. (5) Set a pilot with human sign-off and measure stock-outs, spoilage, and planner time. The board should not let a 'Turing-style' marketing claim substitute for evidence of fitness for purpose; the relevant question is whether the system improves decisions under real constraints, not whether it sounds human.