The pitch for AI agents has always been seductive in a specific way: stop stitching together ten apps, just tell the agent what you want, and it handles the rest. Anthropic’s Claude, OpenAI’s operator-mode models, and a stack of startups built on top of them have all leaned hard into this framing. Agentic AI was supposed to eat the SaaS category whole.
What’s actually happening is more bureaucratic.
Agents Don’t Replace Tools - They Interface With Them
Look at any production deployment of an AI agent today - the ones actually running inside companies, not demos - and you’ll find the same structural reality: the agent is calling APIs, reading from databases, writing to ticketing systems, and logging outputs into dashboards that humans then review. The underlying tools are still there. The agent is just a new orchestration layer sitting on top of them.
This isn’t a failure of capability exactly. It’s a failure of the premise. Enterprise software doesn’t disappear because something smarter can now operate it. Salesforce didn’t die when people built better reporting tools. The data gravity is too strong, the integrations too entrenched, the compliance requirements too specific.
So instead of collapsing the stack, agents have been absorbed into it. And that absorption comes with costs that don’t get talked about clearly enough in product announcements.
The Token Burn Is Real
Running an agent through a multi-step workflow - pulling context, deciding on actions, calling tools, evaluating results, looping back - burns through tokens fast. An agentic loop that takes thirty seconds of wall-clock time might represent a context window the size of a short novel. At current API pricing tiers, the economics only work if the task being automated is genuinely high-value and high-frequency.
Most of the things people want to automate are neither.
The Monitoring Problem Nobody Mentions
- and this is the part that tends to surface six weeks after deployment rather than six minutes after the demo - agents make mistakes in ways that are hard to audit. A traditional script fails loudly. An agent fails softly, often completing a plausible-looking version of the task that is subtly wrong. Teams end up building monitoring layers to watch the agents, which requires engineering time, which offsets the efficiency gain.
The honest framing is that agentic AI works well in a narrower band of use cases than the marketing implies: highly repetitive, well-defined tasks with clear success criteria and low cost of error. Outside that band, you’re not replacing workflows. You’re adding one.