Redesign the workflow, or the agent does nothing

There is a familiar disappointment settling over a lot of agentic-AI programs: the pilot was impressive, the rollout was expensive, and the productivity numbers barely moved. This is not a sign that agents don't work. It is a sign that they were bolted onto workflows designed for humans — and a workflow designed for humans cannot be improved much by dropping a faster worker into one seat of it. The gains come from redesign, or they don't come at all.

The bolt-on trap

The instinct is understandable. You have a support process, a claims process, an onboarding process. An agent can do one step of it faster or around the clock, so you slot the agent into that step and expect the whole process to speed up. It rarely does. The step you automated wasn't the bottleneck, or the hand-offs on either side still run at human pace, or the agent now waits on the same approvals and queues the humans waited on. You have made one node faster in a system whose throughput is set by everything you didn't touch.

Worse, bolting an agent onto a human workflow inherits all of that workflow's assumptions: that a person reviews each item, that context is re-gathered at every hand-off, that trust is established by a human being in the loop. The agent is forced to imitate a human doing a job that was shaped around human limits. It is a fast horse pulling a carriage designed for a slow one.

The lesson of electrification

This has happened before, at least twice. When factories first electrified, they replaced the big central steam engine with one big electric motor — and saw almost no productivity gain. The leap came decades later, when factories were redesigned around what electricity made possible: many small motors, machines arranged by the logic of the work rather than the reach of a driveshaft, the assembly line. The technology had been available for years; the productivity waited on the redesign.

The computer age rhymed with it. Economists spent the 1980s puzzled that computers were "everywhere except in the productivity statistics." The gains eventually arrived — but not from putting a PC on each desk to do the old job faster. They arrived when firms re-engineered their processes around what software could do. The pattern is consistent: general-purpose technologies pay off only after the surrounding work is rebuilt to fit them, and there is a painful lag in between where the investment is real and the returns are not. Agentic AI is early in exactly that lag.

What redesign actually means

Redesigning a workflow around an agent is not the same as documenting the current one and handing it over. It means asking a different question: given an agent that is tireless, bounded, auditable, and able to act on the first turn, what should this process look like if we built it for that?

Concretely, redesign tends to move in a few directions at once:

• Collapse the hand-offs. Much of a human workflow is context being re-gathered as work passes between people. An agent that already has full context before the interaction begins can resolve on the first turn what used to take three hand-offs — but only if you remove the hand-offs, not preserve them.
• Move the human from the loop to the edge. Instead of a person reviewing every item, a person defines the bounds and handles the exceptions the agent gates to them. Oversight shifts from per-item to per-policy.
• Design for the agent's failure modes, not a human's. Humans get tired and inconsistent; agents fail differently — confidently, at the edges, in ways a good workflow catches with gates and audits rather than with a second reviewer.

Bounded authority as the enabling constraint

The counterintuitive part is that the constraint is what unlocks the gain. An agent you can trust to act unsupervised is one whose authority is narrow, explicit, and logged — read this order, refund up to this limit, and nothing more. That boundedness is what lets you take the human out of the per-item loop, which is where the throughput actually comes from. Teams that refuse to draw the bounds keep a human checking everything, and a human checking everything is the old workflow with extra steps. The bound is not a limitation on the agent; it is the thing that makes the redesign safe enough to ship.

Where the gains show up

When the workflow is rebuilt rather than retrofitted, the numbers stop being marginal. Coverage goes to 24/7 without a night shift. First-contact resolution rises because the agent completes the task instead of routing it. Handle time on routine work falls, and the humans you kept are spent on judgment rather than triage. None of that comes from the agent being clever. It comes from the process being redrawn so the agent's strengths are load-bearing and its weaknesses are gated.

The uncomfortable conclusion for anyone hoping to sprinkle agents onto the org chart: the productivity of agentic AI is not a property of the agent. It is a property of the workflow you are willing to redesign around it. Do the redesign and the gains are real. Skip it, and the agent — however capable — does nothing that shows up where it counts.