★ Support Agent · agent or hire?
AI Agent vs Hiring a Support Agent: Cost Calculator
Support is where most teams try an agent first: high volume, repetitive questions and a knowledge base that already exists. The catch is the long tail. Refunds, angry customers and bugs still need a person, and the agent has to know when to hand over.
- Runs in your browser
- No signup
Presets are starting assumptions for a small team, not industry statistics. Change every number to match your reality.
The agent saves
$5,812 / month
Pays back the build cost in 2.6 months.
People only
$11.1K/mo
With agent
$5.3K/mo
Hours freed
152 h/mo
12 months, side by side
Cumulative cost. The agent line starts at the build cost.
| Month | People only | Agent |
|---|---|---|
| 0 | $0 | $15K |
| 1 | $11.1K | $20.3K |
| 2 | $22.2K | $25.6K |
| 3 | $33.3K | $30.9K |
| 4 | $44.4K | $36.2K |
| 5 | $55.6K | $41.5K |
| 6 | $66.7K | $46.8K |
| 7 | $77.8K | $52.1K |
| 8 | $88.9K | $57.4K |
| 9 | $100K | $62.7K |
| 10 | $111.1K | $68K |
| 11 | $122.2K | $73.3K |
| 12 | $133.3K | $78.6K |
Where the agent cost goes
- Work people still do$4,444
- Reviewing the agent$333
- Model usage$22
- Maintenance$500
Model cost per task: $0.0180 on Claude Sonnet 5.5, for 1,200 agent tasks a month.
An agent that pays back in 2.6 months.
I build production AI agents with human review, guardrails and logging.
What the agent can and can't take
Good jobs for the agent
- Answer how-to and policy questions from your help centre
- Look up order or account status through your API
- Collect details and tag tickets before a person sees them
- Draft replies for a person to approve
Keep with a person
- Make judgement calls on refunds or exceptions without rules
- Calm down a customer who is already furious
- Diagnose a new bug nobody has documented
What review looks like: Sample a share of resolved tickets every day, and route anything the agent marks low-confidence, plus every refund, to a person before it is sent.
Good fit, bad fit
| Good fit if… | Not a good fit if… |
|---|---|
| Most tickets repeat the same 20 questions | Every ticket is a unique technical investigation |
| Answers live in docs you can point the agent at | Your docs are out of date or missing |
| You can measure resolution and satisfaction per ticket | A single wrong answer has legal or safety consequences |
Risks to plan for
- Confident wrong answers about policy or pricing
- Customers trapped in a loop with no way to reach a person
- Leaking one customer’s data into another’s conversation through bad retrieval
How the comparison works
The calculator compares two ways of getting the same tasks done each month: people doing all of them, or an agent doing a share while people do the rest and review the agent's work.
people only = tasks × minutes ÷ 60 × hourly cost
with agent = remaining human work + review time + model cost + maintenance
model cost = agent tasks × (input tokens × input price + output tokens × output price)
savings = people only − with agent
payback = build cost ÷ savings (never, if savings ≤ 0)
It only counts the share of someone's time spent on these tasks, so it answers “what does this work cost”, not “can I remove a job”. Model prices come from the LLM API Cost Calculator.
Worked example
The support agent preset with the example salary of $60,000 plus 25% overhead, 2,000 tickets a month at 8 minutes each, the agent handling 60% and a person reviewing 20% of those:
| Line | Per month |
|---|---|
| People only | $11,111 |
| Remaining human work | $4,444 |
| Reviewing the agent | $333 |
| Model usage (Claude Sonnet 5.5) | $22 |
| Maintenance | $500 |
| With the agent | $5,299 |
| Saving | $5,812 |
With a $15,000 example build cost, payback is 2.6 months. Notice how small the model cost is next to the human time that remains: review time and the share the agent can't finish usually decide the answer, not token prices.
Questions about this role
Can an AI agent replace my whole support team?
How do I stop the agent from making things up?
What should a support agent never do on its own?
Is an AI agent cheaper than hiring someone?
Where do the model prices come from?
Compare by role
More from the workbench
Fine print
Tools: on the house.
Engineer: by the hour.
The math was free. A reliable agent takes engineering, which is what I do.
I'm Himanshu, an AI and full-stack engineer. I build MVPs, AI agents and the growth plumbing around them for early-stage startups.