Agentic AI Needs Operating Model Design, Not More Agents
2026-05-27 · 4 min read · Janaina Maia
The agent conversation is growing up. The interesting question is no longer whether an AI agent can complete a task. It is whether the organisation around that agent is designed well enough for the work to be safe, useful, and accountable.
MIT Technology Review Insights wrote about the organisational shift required for agentic AI, including the gap between companies that want agents and companies whose workflows, data, metrics, and decision rights are ready for them. The piece is sponsored, so I would treat the vendor framing with caution. But the core point is right: dropping agents into a human operating model is not transformation. It is often just another layer of complexity.
That matters because most enterprise AI failures will not look like a dramatic model failure. They will look like unclear ownership, duplicated work, invisible assumptions, weak review, and teams quietly disagreeing about who was meant to make the final call.
Agents do not remove operating model design.
An AI agent can search, draft, classify, summarise, route, monitor, and sometimes act across systems. That sounds powerful because it is. But as soon as an agent moves through a workflow, product teams have to answer very old organisational questions in a new form.
Who gives the agent its goal? Which data is it allowed to use? What counts as a low-risk action? Where must it pause? Who reviews the output? Who is accountable if the work is wrong but plausible? These are not backend details. They are the user experience.
If those questions are not designed deliberately, the agent becomes a polite chaos machine. It may make the work faster, but it can also make responsibility harder to see.
Decision rights are a design surface.
Enterprise software has always encoded power: who can approve, edit, escalate, publish, delete, override, or sign off. Agentic software makes that more visible because the system is no longer only storing decisions. It is participating in the path toward them.
That means design leaders need to care about decision rights as much as screens. A good agentic workflow should make it obvious when the AI is suggesting, when it is preparing, when it is acting under instruction, and when a named human has taken responsibility.
The interface should not let accountability dissolve into the word “automation.”
The metrics have to change too.
A common mistake is measuring agentic AI by activity: how many tasks completed, how many prompts sent, how many tickets touched, how many minutes saved. Those numbers can be useful, but they are not enough. An agent can create more activity while making the overall system more fragile.
I would rather see metrics tied to outcomes and quality: fewer escalations caused by missing context, less rework, faster cycle time with the same or better decision quality, clearer audit trails, fewer policy exceptions, and better customer or employee experience.
If the agent changes the workflow, the measurement system needs to change with it.
Design implication: map the human-agent team.
Before adding an agent to an enterprise workflow, I would map the team as if the agent were a new participant with limited judgement and very high stamina.
- Role: what work should the agent do, and what work should it never own?
- Inputs: which sources are trusted, stale, sensitive, or incomplete?
- Checkpoints: where does the work require human review before it continues?
- Escalation: what happens when confidence is low, instructions conflict, or policy is unclear?
- Accountability: who signs off when the agent-assisted work becomes real?
This is where product design and organisational design meet. The interface has to help people understand the work, not just admire the automation.
My take.
I am optimistic about agents, but not about agent theatre. Adding more agents to a broken workflow will not make the organisation more intelligent. It will make the broken parts move faster.
The teams that get value from agentic AI will be the ones that redesign the work around clear roles, visible evidence, review points, and accountable humans. The agent is not the operating model. It is a new actor inside one.
Design the operating model first. Then decide how much autonomy the agent has earned.