StormView

AWS just put $1 billion behind a point many delivery teams have already learned the hard way: enterprise AI does not usually fail because the model is weak. It fails because nobody has connected it to the business properly.

On 30 June 2026, AWS announced a new Forward Deployed Engineering organisation, led publicly by Francessca Vasquez, VP of Frontier AI Engineering and Services. The pitch is direct. AWS will embed thousands of experts with customers to build and deploy agentic AI systems inside live workflows, not sit outside them as chatbots with a search box.

The named early customers include the Allen Institute, Cox Automotive, the NBA, the NFL, Ricoh and Southwest Airlines. CIO Dive reported that many engagements will start with pods of five or six engineers working in compressed 45-day sprints. It also reported a key technical pattern: a semantic layer deployed into the customer’s AWS account, connecting enterprise data sources into a knowledge graph that agents can reason from.

That detail matters. The agent is not the hard part by itself. The hard part is identity, permissions, messy records, audit trails, approvals, fallback paths and the odd spreadsheet that still runs half the operation. If those pieces are wrong, a clever model just makes bad decisions faster.

This is why the AWS move is bigger than another cloud product launch. It says implementation capacity has become the scarce resource. Strategy decks are cheap. Demos are cheap. The expensive work is turning a claims process, sales handover, field service workflow or finance review into something an AI system can assist, escalate and sometimes complete without creating new risk.

Mid-market companies should pay attention, but they should not copy the hyperscaler playbook blindly. Most firms will not get AWS’s top embedded engineers. Most do not need a Palantir-sized programme either. They need the smaller version done well: pick one workflow, map the data and decisions, set the security model, define the human override, measure cycle time and error rate, then decide what should be automated.

Sometimes the answer is an agent. Sometimes it is a retrieval assistant. Sometimes it is a plain integration between two systems. In regulated or sensitive settings, a local LLM may be the better fit because the data boundary matters more than having the newest hosted model. The label matters less than whether the system improves cost, quality, speed or scale.

Gartner’s numbers underline the stakes. On 1 July 2026 it said up to $234 billion of enterprise application software spend could be exposed to agentic arbitrage by 2030, where agents complete work across multiple systems instead of forcing users through separate software screens. But Gartner also warned in 2025 that more than 40% of agentic AI projects may be cancelled by the end of 2027 because of rising costs, unclear value or weak controls.

Both can be true. Agents can replace awkward software work. Bad agent projects can also burn money quickly.

The next phase of AI consulting will belong to teams that leave clients with running systems, documented patterns and staff who know how to operate them. The useful question for July 2026 is not “where can we add AI?” It is “which workflow deserves autonomy, and what has to be true before we let it act?”