AWS just stood up a $1 billion internal group with one job: deploying AI inside Amazon, not building more models. OpenAI and Anthropic have done something similar.
AWS spending a billion dollars to wire AI into messy internal workflows, and what that says about where an AI project should actually start.
AWS just stood up a $1 billion internal group with one job: deploying AI inside Amazon, not building more models. OpenAI and Anthropic have done something similar.
Amazon’s stated reason is blunt: “enterprises don’t fail at AI because models are weak, they fail because nobody inside the building can wire the model into decades of messy workflow.” That is the sentence I would put on a poster in a distribution warehouse.
Legacy ERP systems, compliance workflows, safety documentation built up over 20 or 30 years, that is the messy workflow in this industry. The model was never the hard part.
Here is what I actually do when a client wants to start. I do not open with a strategy deck. I pick one workflow, usually something small and painful, like pulling SDS data into a quote or matching a purchase order against a compliance rule that used to take someone 20 minutes by hand. We wire the model into that one workflow, watch it run for a few weeks, and fix what breaks. Then we pick the next one.
If a company with Amazon’s resources still needs a billion dollars and a dedicated team just for the wiring, that tells a mid size distributor exactly where to spend the next 90 days. Pick the one workflow that is actually costing time and wire it up.