Here's something I've seen repeatedly while building one of the world's largest enterprise automation platforms at Salesforce: companies are deploying AI agents without any real visibility into what those agents are actually doing — or what they're costing.
That gap is going to become a serious problem.
As agents get woven into enterprise workflows, the companies that win won't just be the ones with the best AI models. They'll be the ones that build a proper control layer around those models — one that gives operators visibility, accountability, and the ability to measure what's actually working.
First-generation agent tools are largely isolated and free-standing. The management experience of the future has to be something far more integrated. Here's what I think it needs to look like.
"The question is never whether humans should be in the loop — it's how to design the loop so that humans stay effective at scale.
The Hybrid Workflow Problem
Consider a support center handling incoming customer calls. Some calls are handled end-to-end by a frontline AI agent. Others route to a human rep who may use an AI agent as a real-time assistant — feeding it context mid-call to get a quick diagnostic, then acting on the recommendation.
Three distinct pathways. Three different cost profiles. Three different customer experience outcomes.
Three pathways, one control layer — the hybrid workflow model.
The problem is that in most deployments today, there's no unified view across all three. Token consumption from the AI agent is tracked somewhere. Human handle time is logged somewhere else. Resolution outcomes might live in a third system entirely. None of it is linked in a way that lets you answer the question that actually matters: which pathway is delivering the best outcomes at the best cost?
To answer that question, you need unified logging across the entire workflow — and a way to tag agentic activity with metadata that connects it back to the broader process. Something as simple as a PathID, passed from the orchestrating workflow into the agent as context and returned in the agent's log output, makes it possible to analyze cost and resolution rate by pathway rather than in isolation.
Agents Need a Deterministic Wrapper
There's a tempting alternative architecture: put the agent in charge of everything. Let it manage routing, logging, state — the whole workflow.
Burdening an LLM with the full complexity of enterprise process management doesn't play to its strengths — and in enterprise workflows, reliability is non-negotiable.
Large language models are exceptional at reasoning over language and generating responses. They're not designed to be reliable state machines. When you ask an agent to manage deterministic business logic, you're trading reliability for flexibility.
The better architecture wraps agents in deterministic automation. The automation layer handles routing, generates the universal issue ID, manages state transitions, and collects the structured logging data. The agent does what it does best: reasoning and language. The automation does what it does best: predictable control flow.
OpenAI's Agent Builder reflects this thinking. I expect most mature enterprise platforms will converge on this pattern.
Visualization: Bring Back the Kanban
Once deterministic automation is your control layer and logging is unified, something valuable becomes possible: a single Kanban view that spans human and AI work simultaneously.
In-Session
Assistant AI Agent
A unified Kanban spanning human and AI agent work simultaneously.
From that single view, a supervisor can drill into live sessions to observe conversations in real time, roll up completed work by cost — in tokens for AI sessions, in salary/time for human sessions — and review the prompts that human agents are sending to AI assistants, using that data to improve training materials and refine agent instructions.
None of this requires new AI capabilities. It requires classic orchestration discipline applied to a new kind of workforce — one that happens to include both humans and agents.
What This Means for Product Teams
The vendors who figure this out first will define the next generation of enterprise AI management tooling. Right now, most platforms are optimizing for agent capability. The next frontier is agent observability — giving the humans responsible for these workflows the visibility and control they need to actually trust and scale them.
The technical building blocks already exist: deterministic workflow engines, structured logging, and Kanban-style process visualization have been around for decades. The opportunity is in connecting them to the new agentic layer. The pieces aren't new. The urgency is. And in my experience, the teams who move first on infrastructure moments like this don't just win — they become the infrastructure.