What it takes to move AI agents into production
Getting an AI agent into production depends more on redesigning work than on the model itself. Caylent, an AWS-only cloud and AI services firm, argues that value comes from reworking workflows, human roles and accountability. Almost 60% of enterprise leaders are ready for autonomous agents, but only behind guardrails, and the hardest remaining work is access, identity and proving the return. Caylent had to rebuild its own delivery model and incentives to make that shift itself.
Takeaways
- Autonomy needs guardrails before it ships. A Caylent survey found almost 60% of enterprise leaders are ready to let agents run autonomously in production, but only with defined guardrails. Henderson says a demo that cannot be turned on and show ROI over time is not enough.
- AI redesigns workflows, not just tasks. Work that ran for decades on handoffs and syncs now happens concurrently, with humans deciding and correcting agents in the loop. Henderson's analogy is factories that dropped electric engines into unchanged floors and waited thirty-plus years for productivity.
- Caylent rebuilt its delivery model and pay. Its delivery had been human-based and billed on time and materials, which AI changed by changing how its developers work. Incentives now reward customer satisfaction and speed, and Henderson says success means every customer can say what the value has been.
- Autonomy is a trade-off to keep revisiting. Many customers run read-only workloads already, but true autonomy such as self-healing infrastructure is limited to the most advanced, including Caylent's own managed service practice. Agents can also run up bills while doing nothing useful.
- Access is harder than model accuracy. Henderson says accuracy is only a small part of production readiness, and access is among the hardest problems. Single sign-on is difficult for humans, let alone agents, and security keeps needing inspection and visibility.
- Smarsh freed millions with an agentic SDLC. Caylent customer Smarsh rebuilt its whole software delivery process with an agentic SDLC on Amazon Bedrock and Claude Code, freeing millions of dollars of engineering capacity for reinvestment.
Caylent's path into production
- Find a problem where automating a couple of steps adds huge value, and start simple.
- Assume nothing, and build and test before anything reaches production.
- Start in HR or go-to-market motions, where an agent's mistakes are less critical.
- Keep humans accountable, since they direct the agents and grant their access.
- Reevaluate the autonomy versus risk trade-off as accuracy and reliability change.