Should You Hire Another Person or Deploy an AI Agent?
Deploy an agent when the work is repetitive, high volume, and rule-shaped. The agent runs continuously and is judged on the exceptions it escalates. Hire when the work is relational, ambiguous, or needs accountability you cannot write down as a rule.
Most teams get this backwards. They buy a tool for a problem, or they hire for work a machine could already hold. The question is not about technology. It is about the shape of the work.
Three questions that decide it
1. Is the work rule-shaped?
Can you describe the task as a sequence a competent new hire would follow in their first month? Reading every line of a retailer's deduction notice against a contract is rule-shaped. Deciding whether to hold a customer relationship is not.
2. Is the volume high enough to matter?
Agents are cheap per unit and expensive to set up. If the task happens forty times a year, a person handles it fine. If it happens four thousand times a year and each one takes eleven minutes, you have a role sized for an agent.
3. Can you describe "done" without a meeting?
If the finish line is a written artifact, a number, or a filed claim, an agent can hit it. If the finish line is a feeling in the room, it cannot.
Three yes answers. An agent should own that work.
What agents are still bad at
Be honest about this list, because it saves money:
Negotiation. Anything where the answer changes based on who is asking.
Ambiguity. Situations with no precedent and no rulebook.
Accountability. Where a named person must answer for the outcome.
Relationship ownership. The parts of the job that exist because a human trusts a human.
Keep all four with people. Let the agent do the preparation so the person decides faster.
The honest math
A mid-market operations hire runs roughly $70,000 to $90,000 a year fully loaded. You carry that cost whether the workload is heavy or light, and it does not flex when the quarter changes.
An agent role runs $1,000 to deploy and $500 a month to operate, which is about $7,000 a year. One agent role costs less than six weeks of that hire.
But the comparison is not the point, and framing it as cost savings misses what is actually happening. The point is what the role recovers.
A loan operations team we worked with went from six people to two, while the deals they processed got larger, over $500K each. Nothing got worse. The agents read every line, and the humans approved the exceptions the agents surfaced.
The trap
The expensive mistake is not choosing wrong between a hire and an agent. It is buying five point solutions to cover five fragments of one role.
Each tool solves its own slice. Each one adds a subscription, an integration, and its own place for data to go stale. Six months later you own a stack that overlaps itself, costs more than the hire would have, and still needs a person to hold it together. That is the bloat people mean when they say AI made their operation more complicated.
An agent that owns the whole role replaces the stack instead of extending it.
How to run the decision
List the tasks, not the job title. A role is twenty to forty tasks. Titles hide that.
Tag each task with the three questions above. Rule-shaped, volume, describable finish.
Count what falls through. If a majority of the role's hours are rule-shaped, design the role around an agent and keep the rest human.
Write the approval rule before you build. Decide which exceptions reach a person. This is the step teams skip, and it is the step that makes the agent trustworthy.
Where this goes next
If the work is rule-shaped, the next question is which parts of the role the agent owns and which stay human. That is workforce design, and it is the step that stops you from buying a sixth tool.
How to decide which roles your AI agents should own
AI agents vs stacked point solutions: why piecemeal AI costs more
Frequently asked questions
Is it cheaper to deploy an AI agent than to hire someone? One agent role runs $1,000 to deploy plus $500 a month, about $7,000 a year. A mid-market operations hire is commonly $70,000 to $90,000 fully loaded. The comparison only holds if the work is rule-shaped enough for an agent to own it end to end. If it is not, you will pay for both.
What work should never go to an AI agent? Negotiation, ambiguous judgment, relationship ownership, and anything where a named person must be answerable for the outcome. Keep those human and let the agent prepare the work.
Can we start with just one agent role? Yes, and one role is the normal starting point. It proves the workflow on a single measurable process before you redesign more of the team.
How long does it take to see a result? The first role is usually running inside a few weeks, because the work is already written down somewhere: contracts, retailer portals, an ERP, a shared inbox. The rulebook exists. It just needs to be applied consistently, which is what the agent does and what people run out of hours to do.
