Suri Nuthalapati

The FDE shortage will not end because companies hire their way out of it. It will end when each deployment leaves reusable patterns behind for the next one. The teams that capture those patterns will ship the second workflow in days, not months.

The next production gate for enterprise agents will be identity, not capability. An agent that borrows a person's login will not pass review. Every production agent will carry its own scoped and revocable credential.

We do not need a better hinge for a foldable phone. We need to ask why the phone needs to fold at all. The smartest engineering decision is often to delete the thing instead of improving it.

Companies will stop giving agents more data and start giving them meaning. Meaning will live in a shared semantic layer that holds the definitions, metrics, and rules behind the numbers. The teams that build that layer first will trust their agents with real decisions.

Every production agent will soon have a named human manager. That manager will own its budget, its permissions, and its results. The agents that cannot prove their work will be retired like any role that stops earning its keep.

Companies will stop choosing agents by the model inside them. They will choose them by the harness around the model, the system that controls what an agent can see, do, and prove. The model will keep getting cheaper, and the harness will decide who reaches production.

The companies that scale agents will separate building them from watching them. They will put monitoring, permissions, and audit trails on infrastructure the agent itself cannot edit. An agent that can silence its own alarms will never be trusted in production.

Companies will stop dropping AI tools into old workflows. They will redesign one workflow around the agent and name one owner for the result. The teams that redesign the work will reach production, and the teams that only add a tool will stay in pilots.

Companies will not let agents run loose inside their core systems. They will give each agent a contained workspace with clear limits. Credentials and approvals will stay outside that workspace, beyond its reach. The winners will design the container before they scale the agent

Companies will soon onboard AI agents like they onboard employees. Each agent will carry a verified identity, limited permissions, and a clear owner. The teams that build that trust first will move agents into production while others stay stuck in pilots.

In a few years, "forward deployed engineer" stops being a special title and becomes the default way enterprise AI is built. Models will be rented. Outcomes will be owned; by engineers inside the business, shipping to production. FDE era is now, and it’s just getting started.

66% of companies now run agents in production. 45% lose more than a quarter of the expected gains to oversight and rework. Production was never the finish line. Without a named owner and real observability, every agent adds work it was meant to remove.

74% of finance teams now use AI. Only 14% report enterprise value. The difference is not the model. It's the foundation underneath it — data that can be traced, explained, and trusted. Scale the AI on untraceable data, and you scale doubt, not decisions.

AI coding agents produced 30% more code across 718 firms, and no more finished work. Review time rose 49%. Generation was never the bottleneck. Verification is. Scale output without scaling approval, and work doesn't ship faster. It queues.

The demo is no longer the sale. Buyers now ask a harder question: who is running this in production, on real work, and what number moved? That shift is healthy. It rewards deployment over promises — one workflow, live, measured.

An agent can pass every eval and still make the wrong call in production. Evals test it against yesterday's data. Production asks a harder question: is what it believes still true at the moment it acts? Build that live check in, or the demo was the peak.

A guardrail written in a prompt is a request, not a control. Production agents need limits enforced where the action happens — at the permission layer, with a named human approving anything that writes, pays, or deletes.

85% of knowledge workers use AI. Only 6% of executives can point to clear, company-wide ROI. That gap is not about access to models. Everyone has that now. It's about context — your data, your workflows, your people. That's deployment work, not software.

Most AI pilots don't die in the demo. They die in the handoff — to security, to data, to the team that has to run it in production. The fix isn't a better model. It's engineers embedded with the business, shipping one real workflow end to end.

Enterprises now average 183 AI tools. Only 15% operate as one system. That is not a technology problem. It is an ownership problem. AI does not compound when every team buys its own. It compounds when someone owns how it all connects.