On this page
"AI agent" has been applied to enough different things in the last two years that the term has lost most of its edges. This article puts them back on, from the point of view of someone deciding whether to buy one.
The one-sentence definition
An AI agent is a system that is given a goal, decides for itself what steps to take, uses tools to take them, and checks whether it worked — as opposed to following a fixed sequence somebody wrote in advance.
The three things people call AI, compared
The last row is the practical guidance. If the task is genuinely identical every time, an agent is the wrong tool — it introduces variability into something that had none. Agents earn their keep where the path differs each time.
The four parts of an agent
A goal
Not a script. "Resolve this support ticket" rather than "send template 4". The goal has to be specific enough to be checkable.
A model that reasons
The language model works out what to do next given the goal, what it knows, and what has happened so far.
Tools
The things it can actually do: look up an order, read a calendar, send an email, create a ticket, query a database. An agent without tools is a chatbot. This is the single most important thing to understand about the category.
A feedback loop
It observes what its actions returned and adjusts. This loop is what separates an agent from a very elaborate prompt, and it is also what makes agents harder to predict.
The terms behind each of these — tool calling, retrieval, orchestration — are defined in the AI glossary if you want the technical vocabulary.
What agents are genuinely good at
- Multi-step research. Gather information from several sources, reconcile it, and produce a summary — where which sources to check depends on what the earlier ones said.
- Triage with variation. Every enquiry needs a different lookup before it can be routed sensibly.
- Long-running processes with branches. Onboarding a new client, where the next step depends on what came back from the last one.
- Conversations that need actions mid-flow. A voice agent that checks a calendar and writes a booking while still on the call is an agent in exactly this sense.
- Work that spans systems. Reading from one, deciding, writing to another.
Where they fail
Agents fail differently from ordinary software, and the failure modes are worth knowing before you deploy one.
- Compounding errors. A small mistake at step two is treated as fact at steps three through nine. Longer chains are less reliable, not more capable.
- Confident wrongness. An agent will report success in the same tone whether or not it succeeded. It needs verification it does not perform on itself.
- Cost drift. An agent that decides its own steps decides its own costs. Set hard limits on iterations and spend.
- Difficult debugging. When it does the wrong thing, the reason is in a reasoning trace rather than a stack trace. Logging every step is not optional.
- Over-permissioning. The most common serious mistake: giving an agent broader access than the task needs, because narrowing it was fiddly.
The permissions rule
Give an agent exactly the access its task requires and nothing more — read-only wherever reading is enough. An agent with write access to your production database because it was easier to set up is an incident waiting for a date.
Do you need one?
A short test. If you answer no to any of the first three, use a simpler tool.
- Does the task genuinely vary? If the steps are the same every time, build an automation. It will be cheaper, faster and more reliable.
- Can success be checked? If nobody can tell whether the agent did it right, you cannot operate it safely.
- Is the blast radius acceptable? What happens if it gets one wrong, twenty times, overnight, before anyone notices?
- Does it need to act, not just answer? If answering is enough, a chatbot is simpler and more predictable.
Most businesses that ask for an AI agent need an automation with good error handling. The ones that genuinely need an agent usually already know why.
How to deploy one safely
Start read-only
Let it observe and recommend for a few weeks. You get its judgment with none of its risk, and a body of evidence about how good that judgment is.
Add writes to one low-stakes system
Something reversible. Draft rather than send. Tag rather than delete.
Keep a human approval step
Remove it only once you have months of data showing the agent's judgment on that specific task.
Log everything
Every decision, every tool call, every result. You will need this the first time something is wrong.
Set hard limits
Maximum steps, maximum spend, maximum records touched per run. Not soft guidance in a prompt — hard limits in code.
Frequently asked questions
What is the difference between an AI agent and a chatbot?
Tools and initiative. A chatbot answers within a conversation; an agent takes actions in your systems and decides which actions to take. In practice many production systems are both — a voice agent that answers questions and books an appointment is a chatbot with tools attached.
Are AI agents reliable enough for real business use?
For narrow, well-defined tasks with verification and limits, yes — that describes most successful deployments. For open-ended autonomy across many systems, reliability drops sharply as the chain lengthens. Scope is the main lever you control.
How much does an AI agent cost to build?
Less than people expect for a narrow agent, more than they expect for a broad one, and the cost is driven by integrations and safety work rather than by the model. TensoraAI publishes AI agent pricing on the pricing page.
Can an AI agent replace an employee?
It can take over specific repetitive tasks, and it does not take over a job. Jobs contain judgment, relationships, accountability and improvisation that agents do not have. The realistic framing is that an agent removes a set of tasks from a role, not the role itself.
What happens when an AI agent makes a mistake?
That depends entirely on how it was built. A well-designed agent logs the step, stops, and escalates to a person. A poorly designed one continues, treating its mistake as established fact — which is why limits and logging matter more than model choice.
If you want to know whether your problem is an agent problem, an automation problem or a chatbot problem, that is exactly what the free AI audit is for. See also the AI agents service page.
