When I've been brought into AI transformation projects, they often take shape in
By Ericseufert · November 10, 2026 · Curated by George's Blog
When I've been brought into AI transformation projects, they often take shape in three stages:
1) "Throw an LLM at it." Integrate an LLM or some sort of generative tool at every workflow. This is usually driven by fear; management thinks the company is falling behind on AI adoption, so every team is given a mandate to adopt some new AI-assisted tool, or to build one, and this can take the form of simply giving many members of the team large-limit Claude / Codex plans, to procuring enterprise contracts with those companies, to building proprietary tools that work through VertexAI, Bedrock, or AI Foundry.
2) "Efficiency benchmarks." The first stage becomes very expensive, so teams identify workflows that don't benefit from these models and revert to their prior operating models. Ironically, this is often interpreted as cost savings.
3) "Token diet." Teams start considering whether they need frontier models for workflows and tasks that are made more efficient by integrating LLMs or other generative models. What if we can use 3.8 Flash instead of 3.1 Pro? What if we can use 3.5 Flash instead of 3.8 Flash? This requires custom evals and is time-consuming, but it often delivers substantial cost savings with little to no trade-off in output quality.
If I had to guess, 20% of the companies in the consumer app space have even started this sequence. Maybe ~10% of those have progressed past stage one.