You may already know where AI agents earn their keep. This piece covers the other half: the jobs where an agent should not act alone, the guardrails that keep it safe, and what it costs to keep one running after launch.
What an agent is, briefly
An AI agent is software that reads a request, decides on steps and takes actions through tools you give it. It might look up an order, draft an email or update a record. The useful part is that it acts. The risky part is the same thing.
Where agents do not help
Some jobs look like a fit and are not. Be wary when:
- The answer must be exactly right every time. Language models can produce confident, wrong output. Tax figures, legal wording and safety instructions need a person or plain rules.
- The rules are simple and fixed. If a task is "when this happens, do that", ordinary code is faster, cheaper and predictable. An agent adds cost and uncertainty for no gain.
- The process is not written down. An agent cannot follow a process that nobody in the business can describe. Map it first.
- A mistake cannot be undone. Payments, refunds, deleted records and messages to every customer are hard to take back.
- The moment needs judgment or care. An upset customer, a dispute or a decision about a person deserves a human.
Guardrails that matter
Guardrails are what turn a clever demo into something you can leave running. The ones we rely on:
- Narrow tools. Give the agent only the actions it needs. Start with read access and add write access only where it is earned.
- Limits. Cap what it can spend, how many actions it can take in one run and who it can contact.
- Approval steps. Anything that moves money, changes records or goes out under your name waits for a person.
- Logs. Record every action and the reason behind it, so you can see what happened and fix it.
- A way out. When the agent is unsure, it hands the case to a person instead of guessing.
- Tests on real examples. Check it against your own past cases before launch, and again whenever the model or your process changes.
When a person stays in the loop
The rule is simple. The more a mistake costs, and the harder it is to reverse, the closer a person should sit to the decision.
| Type of task | Suggested setup |
|---|---|
| Sorting and tagging incoming email | Agent alone, with regular spot checks |
| Drafting replies to common questions | Agent drafts, a person sends |
| Issuing refunds or credits | Agent prepares, a person approves |
| Changing prices or contract terms | A person decides |
| Handling a complaint or dispute | A person leads, the agent can summarise |
What it costs to run
The build is one cost. Running the agent is another, and it recurs every month. Plan for:
- Model usage. Most AI models are billed by how much text goes in and comes out. A busy agent with long instructions costs more than a quiet one with short ones.
- Hosting and storage. The agent, its logs and any searchable copy of your documents need somewhere to live.
- Upkeep. Models get updated, your processes change, and instructions and tests need care to keep up.
- The person in the loop. Review time is real time. Count it, or the savings will look better on paper than in practice.
There are honest ways to keep this down. Use a smaller model for simple steps. Keep the context short. Let plain code handle fixed rules, and call the model only when a step truly needs it.
We do not put prices in articles, because rates change. Our estimator gives you a range in minutes from the current rate card, with build cost, monthly costs, effort by role, weeks and every assumption written out.
How we approach it
We start by mapping your process step by step. Then we decide which steps plain automation can handle, which an agent can handle, and where a person signs off. That keeps the agent small, the costs clear and the risks in view.
You can work with us on a fixed price for a defined scope, with a team by the month, or with a monthly block of hours and a promised response time. The last option fits well when an agent needs ongoing tuning after launch.
If you have a task in mind, run it through the estimator or tell us about it at Atheron Network Labs. We are open for projects now.
