Generative AI has produced more enterprise pilots and more abandoned pilots than almost any technology in recent memory. The organisations getting durable value from it share a common pattern: they treat it as a data and process problem first, and a model choice second.
01 Start with a narrow, high-friction task
The best first projects are narrow, repetitive, and currently painful: summarising claims documents, drafting first-pass customer responses, extracting structured data from contracts. Broad 'AI strategy' initiatives rarely ship; narrow ones do, and build the case for the next one.
02 Your data is the real bottleneck
A generative AI system is only as good as the context it can retrieve. Most pilots that stall do so not because the model is weak, but because the underlying documents are unstructured, duplicated, or scattered across five systems with no single source of truth.
03 Evaluate against a business metric, not a benchmark
Model benchmarks tell you little about whether a tool solves your problem. Define success in business terms up front minutes saved per document, error rate versus a human baseline, or resolution time and measure the pilot against that, not a leaderboard score.
04 Plan for human review from the outset
Every production generative AI workflow we've deployed keeps a human in the loop for high-stakes decisions. This isn't a temporary safety net to be removed later it's a permanent part of a well-designed system, and it's what makes stakeholders trust the output enough to actually use it.
05 Scaling: from pilot to platform
Once a pilot proves value, the scaling questions become about infrastructure: cost per request at volume, data governance for what the model can see, and monitoring for drift in output quality. Treat this stage with the same rigour as any other production system because it is one.
Salcon Team
Feb 2025
