Cognitive computing, meaning systems that read text, interpret images and speech, classify, rank and recommend, is increasingly practical for enterprise use. The practical question for most organisations is not whether it works in a demonstration but where it fits in a workflow they already run.
Good candidates are narrow, frequent and checkable: sorting incoming documents, extracting fields from a standard form, routing a service request, flagging an unusual reading for review. They have a clear definition of correct, and an existing process that can be compared against.
Poor candidates are open-ended, rare or high-stakes without a check: decisions that affect a person's rights, anything where a confident wrong answer is expensive and unnoticed.
A human review step is not a sign that the technology failed. It is how an organisation keeps accountability and learns where the system is weak. A sensible pilot has four features:
- A narrow task with a defined input and output.
- Review before action, at least until measured error rates justify less.
- A baseline: how long it takes today and how often it goes wrong.
- A decision rule agreed in advance for expanding, adjusting or stopping.
The NIST AI Risk Management Framework gives a useful structure, organised around governing, mapping, measuring and managing risk. Even a small pilot benefits from asking: what data trains or informs the system, who may be affected by errors, how will drift be detected, and how can a person contest an outcome? Records of inputs, outputs and reviews make later audit possible.
A model that classifies documents is only valuable if the classification lands in the system where work happens: the ERP queue, the case record, the approval flow. Much of the effort in enterprise projects is this integration, along with permissions, logging and performance, rather than the model itself.
Cognitive systems make mistakes, and the mistakes can be confident. They also change as underlying models and data change. Plan for monitoring and periodic re-evaluation, and avoid describing results beyond what a measured pilot has shown.
QDT's research and innovation work explores cognitive capabilities where they fit enterprise systems, starting from a bounded pilot and an agreed measure of success. We prefer to show a modest result with evidence over a broad claim without one.