An AI agent is software that takes a goal, works through a series of steps, calls tools such as your email, calendar or database, and decides what to do next based on what it finds. That is different from a chatbot that answers one question at a time. It is also different from a plain automation that follows a fixed script. The agent chooses.
That ability to choose is the whole point, and the whole risk. We build agents for clients, and we turn down agent work when a simpler tool would do the job better. Here is how we think about it.
Where agents help
Agents do well on work that is repetitive in shape but varied in detail. The steps are roughly the same every time, but the inputs are messy, so a fixed script breaks.
- Sorting and drafting replies to inbound email, where each message is different but the kinds of answers are few.
- Pulling the same facts out of invoices, forms and documents that arrive in different layouts.
- First-pass research, such as checking a list of suppliers against a set of criteria and writing a summary for you to read.
- Keeping records in sync across tools that do not talk to each other, where a person today copies and pastes.
- Answering routine customer questions from your own documents, with a clear handoff when the question is not routine.
In each of these the agent saves time on the middle of the task, and a person still owns the result.
Where agents do not help
An agent is a poor fit when the work is rare, when a mistake is expensive, or when the rules are already crisp.
- If a task happens a few times a year, the time spent building and checking the agent will not pay back.
- If a wrong action cannot be undone, such as sending money, deleting records or making a promise to a customer, the agent should not act alone.
- If the logic is simple and stable, a normal automation with no model inside is cheaper, faster and easier to test. Not every workflow needs a model that reasons.
- If you cannot describe what a good outcome looks like, the agent cannot tell either. Agents need a clear definition of done.
Guardrails
The difference between an agent that helps and one that causes a mess is almost always the guardrails around it, not the model inside it. We build these in from the start.
- Limit the tools. The agent only gets access to the systems it needs, and only the actions it needs. Read access is granted before write access.
- Separate read from act. Looking things up and drafting are low risk. Sending, paying, changing and deleting are high risk, and each high-risk action gets its own gate.
- Log everything. Every step, every tool call and every decision is recorded, so you can see why the agent did what it did.
- Set stop conditions. The agent halts and asks for help when it is unsure, when it hits something it has not seen before, or when it has tried too many times.
- Test with real examples. We run the agent against a set of past cases, including the awkward ones, before it touches live work.
When a person should stay in the loop
Human in the loop means a person reviews or approves before an action lands. It slows the agent down, and that is often the right trade.
We recommend keeping a person in the loop when the action is hard to reverse, when it goes to a customer in your name, when it involves money or legal commitments, or when the agent is new and you do not yet trust its judgement. As the logs show that the agent gets a category of task right again and again, you can lift the review for that category and keep it for the rest.
The opposite mistake is also common: a review step on every single action, so nobody saves any time. Reviewing the agent's output should be faster than doing the work yourself. If it is not, the agent is not helping and we should rethink it.
What it costs to run
Agents have a running cost that plain software does not. Each step the agent takes calls a model, and each call is billed by the provider. An agent that takes many steps, retries often or reads long documents will cost more per task than one that is tight and well scoped. There is also the cost of the tools it calls, hosting, monitoring and the time someone spends reviewing its work.
We will not quote figures here because rates change. Our estimator gives you a range in minutes, with the build cost, the monthly running costs, effort by role and every assumption listed, so you can see what drives the number before you talk to anyone.
How we approach it
We start with the task, not the technology. We map the steps a person does today, mark which ones carry risk, and decide which ones need a model at all. Often the answer is a mix: plain automation for the fixed steps, an agent for the messy middle, and a person at the gates that matter.
We can build this as a fixed price for a defined scope, as a team by the month, or as a block of hours with a promised response time. If you are weighing an agent for your business, run the estimator first. It is quick, and it will tell you plainly whether the idea is worth a conversation.
