What it looks like when agents do a ten-person team's work

FinalPeace, my newest company, makes end-of-life planning software. Its development and marketing work, meaning the ads, the emails, the landing pages and the product fixes, is done by AI agents doing the work of a ten-person team. The agents send finished work to an approvals queue. I approve what they ship in about half an hour a day.

I’m writing about it this week because of a number in the Wharton and GBK Collective study released today. Of the 1,416 US enterprise leaders and employees surveyed, 67% say their organization uses AI agents in some form. Only 17% let an agent complete a task without a person approving each step.

GBK’s Brian Smith described most companies as sitting in a “supervised middle, paying for autonomy and then reviewing every action.” He says that is where returns are weakest.

I agree with the diagnosis. My answer to it is a different kind of review.

By the numbers

17%

in the Wharton-GBK study let an AI agent finish a task without a person approving each step

Source: natlawreview.com

Review the finished work, not every step

There are two ways to keep a person in charge of an agent. One is to approve each step as it happens: the agent proposes an action, waits, gets a yes, proposes the next one. The other is to let the agent do the whole job and review the result before it goes anywhere.

The first feels safer. In practice it turns the person into a full-time babysitter, and it means the agent can only move as fast as someone is watching. That is the supervised middle the Wharton study describes.

The second is how you would manage a capable employee. You don’t stand behind a designer while they move each layer. You look at the finished page and say yes, no, or change this. FinalPeace runs on the second model. The agents do the full piece of work, and what reaches me is something ready to ship.

Before and after

Approving every step

Agent waits for each yes
A person watches all day
Work moves at review speed

Approving finished work

Agent completes the whole job
One queue holds the results
A person decides what ships

Both keep a person in control. Only one frees that person's day.

How an approvals queue works

The shape is simple, and it is the part I’d tell any operator to copy.

An agent finishes a piece of work: a draft email, an ad variation, a landing page, a product fix. It doesn’t publish it. It puts it in the queue, with enough context for me to judge it quickly.

For each item I make one of three calls. I approve it, and it ships. I reject it with a short note saying why. Or I edit it myself and approve the edited version. The note on a rejection matters most, because it goes back to the agent with the work. A good note is specific, for example “the claim in the second line isn’t something we can say,” so the next version fixes the actual problem.

FlowHow work moves at FinalPeace

01Agent finishes the work
02It lands in the queue
03I approve or edit or reject
04Approved work ships

Every piece of finished work passes through one queue before it goes out.

That’s the whole mechanism. Everything the agents ship passes through one place, and my job is to make decisions about finished work. Half an hour a day is enough because I’m judging results, not supervising effort.

Why one queue matters more than the agents

It’s tempting to focus on which models or tools the agents use. That matters less than the queue.

One place. If agent output arrives in email, Slack and three dashboards, you will miss things and you will spend your time hunting. A single queue means you can see everything waiting and clear it in one sitting.

Finished work only. If the queue fills up with half-done drafts and questions, it becomes the step-by-step model again by another route. The agent should bring a result.

Decisions that teach. Every rejection note is an instruction. Over time the agents should need fewer of them on the same issue, which is what lets one person’s half hour cover a team’s worth of output.

The person who approves owns the outcome. I am accountable for what FinalPeace ships. The queue makes that accountability practical, because nothing goes out that I haven’t seen.

The big platforms are building toward the same pattern. At its Gemini at Work event on October 8, Google described “coworker” agents that get their own Workspace accounts and identities, with audit trails attributed to each agent and spend caps that pause an agent when hit, as Thomas Kurian laid out on the Google Cloud blog. Identity and logs tell you what an agent did. A queue decides what it’s allowed to ship.

Where this works and where I’d be careful

This model fits work where the output can be judged on sight by someone who knows the business. Marketing copy, ads, emails, landing pages and code changes all fit. You can look at the result and know whether it’s right.

It fits less well where a mistake can’t be undone once the work is approved, or where judging the output takes as long as doing it. Those cases need either a narrower job for the agent or a second reviewer.

The queue also doesn’t replace watching the system as a whole. Wharton’s summary of the report notes that departments which monitor their agents automatically report transformed workflows at 24%, against 9% for those that don’t. Reviewing output tells you whether a given piece of work is good. Monitoring tells you what the agents are doing between the items you see, and whether the same mistake keeps coming back.

And the queue doesn’t set direction. The agents produce ads, emails, pages and fixes, but deciding what the company needs next is still my job. The half hour is for judging the work. The thinking about what work to ask for happens elsewhere.

Decide where the agents’ job ends before you build anything. At FinalPeace their job is operations: development and marketing. Writing that boundary down early saves you from arguing about it later, one edge case at a time.

The Wharton study found that only 7% of organizations succeed on all three of its success markers, which are proving value, transforming workflows and aligning the organization, per the same release. Redesigning review is one of the cheapest ways I know to move on the second one.

What to do this week

Pick one stream of work where the output can be judged on sight, such as outbound emails or ad copy.
Decide who approves it. One named person, not a committee.
Set up a single place where finished drafts land, even if it starts as a shared folder or a board.
Have the agent, or the person using AI, deliver complete drafts there and nothing else.
Approve, edit or reject each one with a one-line note, and keep the notes.
At the end of the week, read the notes. The repeated ones become the agent’s next set of instructions.