Generative AI in practice

Two and a half years into the generative wave, the gap between what is shipping to production and what is still demo-ware is wide enough to be a strategy question in its own right.

What is actually in production, reliably, across the clients we see: internal document search, meeting summarisation, first-draft generation for structured content (contracts, reports, standardised correspondence), coding assistance for developers, and tier-one customer support with human handoff. Boring, useful, measurable.

What is still mostly demo-ware: autonomous agents handling multi-step work without supervision, generative AI making customer-facing commitments without a human in the loop, and the full-stack "AI employee" narrative that vendors keep selling and that keeps not quite working outside controlled environments.

The gap is not permanent. Some of today's demo-ware will be next year's production. But the current cost of mistaking the second category for the first is high, and it falls on the operating team, not the vendor who sold the demo.

Our rule of thumb: for anything customer-facing or balance-sheet-affecting, assume the technology is one generation behind what the demo suggests. For internal productivity, assume it is roughly where the demo says it is. The two postures look similar on paper and lead to very different outcomes in practice.