On the surface it looks like the goal of AI agents is to remove humans from the loop, but at Levver, we’re seeing the most value in moving humans into the loop, at intentional moments in an intentional process. We’re designing and deploying agents for clients and for internal use and we’re all adjusting our workflows in real time. We’re not doing less work, we’re doing more, but the work itself is changing.
Most of the conversation around AI and work is still oriented around replacement. Which jobs can an agent do? Which tasks no longer need a person executing them? How much work can be automated before a company needs to hire someone else?
It makes sense. As a culture we tend to prioritize efficiency. If a machine can perform a repeatable task faster, cheaper, and reliably enough, eventually we figure we don’t need to pay a person to do it manually. AI agents are starting to extend that logic into office work because so much of that work already happens inside software. Our information lives there, our communication happens there, and a lot of the systems we use to run sales, finance, marketing, operations, and support can increasingly be operated by an agent.
Clearly, more procedural work is going to be automated, but what’s more interesting is what happens to the person who used to do that work. The obvious answer is they get replaced. The more nuanced answer is their work changes.
The work moves up a layer
Think about a salesperson researching an account, updating the CRM, deciding what to do next, drafting an email, sending it, and scheduling a follow-up. An agent can already help with a lot of that sequence, and over time it’ll probably handle more of it without someone guiding every step.
There’s still a lot left to decide though. Someone has to define what a good prospect looks like, which information matters, what signals should change the approach, what the agent is allowed to do on its own, and when something deserves attention. Someone has to look at the output and know whether it actually makes sense.
The work starts to move up a layer. Instead of performing every step in a workflow, the person spends more time designing, supervising, and correcting the system that performs those steps.
This is where the agent manager emerges.
At first it might seem like a change that mostly affects entry-level work. If agents absorb a lot of clerical and procedural tasks, then junior employees would end up supervising them, but in practice it’s more expansive than that. A salesperson can become an agent manager. So can an analyst, an architect, a developer, an editor, or an executive.
Anyone who used to express their expertise mostly by doing the work now has another challenge: figuring out how to make that expertise understandable to a system that can do some of the work for them, and the more experienced you are, the more there is to unpack.
Reverse-engineering the person who was good at it
When we need to build a new agent or harness, we often start with something Brady Silas has already made. He’s spent enough time working with agents that he has good instincts about how much context they need, how instructions should be delivered, and where things tend to go wrong. We aren’t just copying his prompts (sometimes we are). Mostly we’re borrowing a method he’s developed through experience.
We do something similar with proposals. Andrew Henke has written more of them than anyone else on the team, so when we use an LLM to help create one, we follow the process he’s developed. The useful part isn’t just that Andrew has a good template. He’s made a lot of small decisions over time about what belongs in a proposal, what order things should appear in, what needs more explanation, and what can stay concise.
We think we’re just going to automate a task, and instead we find ourselves reverse-engineering the person who was good at it.
Experts make a lot of decisions they don’t necessarily think about as decisions anymore. They notice things almost automatically. They know when an answer is technically correct but still feels wrong. They know which exception matters and which one can be ignored. They’ve accumulated enough context that their judgments become instinctual.
If you want an agent to work more like an expert, you have to start pulling those instincts apart and explaining them. That’s why the agent manager’s job is really more than prompting. It’s closer to knowledge transfer.
Where efficiency stops helping
There’s a quote from a 1979 IBM training document that relates:
“A computer can never be held accountable. Therefore, a computer must never make a management decision.”
At some point procedural efficiency hits a limit. The system may be able to take an action, but somebody has to decide whether that action makes sense in context.
You run into questions like: Is this accurate enough? Does this exception matter? Is the risk acceptable? Does this actually match what the client asked for? When should the system stop and ask for help? Who owns the consequence if it gets something wrong?
Those decisions are expensive in a way automated execution isn’t. They require attention from someone who understands the work well enough to make a judgment.
That means companies are going to have to get much better at deciding where that attention is worth spending.
There’s no point asking a person to inspect every mechanical step an agent takes. That defeats a lot of the point of automation. At the same time, putting a human approval step only at the end of a workflow doesn’t necessarily mean useful judgment will occur. What we should aim for is designing a process where human attention shows up where it can have a meaningful effect on the outcome.
That means deciding where agents can operate freely, where people need visibility, which conditions should trigger review, and which decisions carry enough consequence that a person should stay involved.
The interface becomes part of the work
This is where the interface starts to matter more.
When you’re managing an agent, you have to be able to tell it what you want in a way it can decipher and use. You need to know what context it has, what knowledge it has access to, what it’s allowed to do, and how it’ll deliver results. You need a way to correct it when it gets something wrong and, ideally, to make those corrections useful the next time around.
That interaction is starting to become part of the job itself.
We’ve spent decades designing software around menus, fields, buttons, records, and screens. Agents introduce another layer where the interaction is much more conversational and much more dependent on context. You’re teaching the system how you work while you work with it.
The shift from visual interfaces to conversational ones is a whole topic in itself – for this discussion the important part is that the quality of the interface affects how well human judgment makes it into the automated process.
The agent can only work with what the person managing it is able to communicate.
Expertise becomes infrastructure
One of the most interesting consequences of all this is what happens to expertise inside a company.
Historically, a lot of what made a great employee valuable lived inside that employee. They could mentor other people, write documentation, create templates, and build processes, but plenty of their judgment remained tacit. You often knew who the expert was because they could look at a problem and see something everyone else missed, even if they had a hard time explaining exactly how they saw it.
Agents create a reason to make more of that knowledge explicit.
If you want a system to operate with something resembling your best salesperson’s instincts, your best architect’s judgment, or your best editor’s taste, you first have to understand what those people are actually doing. You have to identify what they notice, how they prioritize, what makes them question an answer, when they escalate, and how they decide something is good enough.
In that sense, AI turns expertise into infrastructure.
That doesn’t make the expert less valuable. In a lot of cases, it gives their judgment more reach. A decision about how a system should behave can influence hundreds of future actions. The same is true when the decision is bad, which is why the judgment directing the automation matters so much.
As execution gets cheaper, the quality of that judgment becomes more consequential.
Becoming the human in the loop
Work is being done differently now. With all these new tools, we aren’t just automating tasks and sitting back to relax. We’re experimenting with the tools, improving them, teaching them, and finding more ways to create.
We’ve started spending more of our time setting intent, defining boundaries, handling exceptions, verifying quality, and owning the outcome. The agent manager is becoming the new interface through which human judgment gets applied to increasingly automated work.
That changes what expertise looks like too. Knowing how to do the work yourself still matters but communicating precisely how you would do the work yourself becomes another skill to master.
If we can combine the speed and scale of automated execution with the care, taste, and understanding of an expert who’s invested, we can have the best of both worlds. We can offload the boring stuff and spend more time on innovation – bring human thought into the mix where and when it matters most to produce the highest quality work.
Agents are going to keep getting better at doing things. Our job is to get better at deciding what they should do, teaching them how we want it done, and knowing when the work needs something only we can add.
Share this with someone who’s already becoming an agent manager.
This resonates.
“AI turns expertise into infrastructure” gets to the heart of the shift for me. The opportunity isn’t simply automating what people do. It’s understanding the judgment and experience behind how the best people do it, then figuring out how to put more of that into the system. The human doesn’t leave the loop. We just need to get much smarter about where the human belongs.
Really smart thinking from the Levver team.