Definitions come before architecture
Most disagreements about AI projects are not disagreements about technology. They are disagreements about definitions. One group treats the technology as effectively magic, another treats it as useless, and both positions survive because nobody has stated plainly what the words mean.
Three definitions resolve most of it: what artificial intelligence is, what separates an agent from a chatbot, and what an agent can realistically be asked to do.
Artificial intelligence: learning that already finished
Artificial means man-made. Intelligence means the capacity to learn. By etymology, artificial intelligence is a man-made capacity to learn. That is an honest starting point, and it needs one correction.
The learning already happened. A model learned patterns during training, from an enormous amount of prior work, and then that learning stopped. The system a team uses on Tuesday is not learning from Tuesday. It is a fixed artifact applying earlier learning to work it has never seen.
That correction pre-empts the two most expensive misconceptions in any organization. The first is that the system will get better at proprietary data on its own; it will not, unless someone deliberately builds that. The second is that it remembers last week’s conversation; it does not, unless someone built a place for it to remember.
The word model carries most of the weight in that definition, and it is worth unpacking on its own. What a Model Actually Is covers what the thing is made of, what training and inference actually do, and the vocabulary that arrives with them.
An agent: tools and a loop
A common definition of an agent is software configured to use that learned capacity, along with the inputs it is given, to produce an output. The definition is accurate, but it describes a model. A chatbot fits it exactly, and the distinction between a chatbot and an agent is the one that determines whether a project is worth running.
Two words carry the definition. Tools describe what it can reach: search, files, databases, requests, documents. Loop describes who drives it: no one prompts each step, because the system prompts itself and reads the result before choosing again. A chatbot answers. An agent goes, does, and returns.
That loop is easier to watch than to explain. Agents in One Paragraph covers the same loop from the builder’s side, including where the behavior actually comes from.
What an agent can do
An agent can read unstructured information and decide what to do with it, within a set of instructions or parameters. It is matching what it has seen before, not thinking the problem through.
The more useful framing is that an agent acts rather than only classifies. Reading a disordered inbox and sorting it is a modest result. Reading it, drafting the replies, filing the attachments, and flagging the three messages that require a person is the actual offer.
The general shape of a good candidate task: work a competent person could do with the information already on their screen, that takes judgment but little surrounding context, and that anyone doing it would describe as tedious.
What an agent cannot do
The limits belong in the same conversation as the capabilities, because a team that can state them is a team that ships something real rather than something that demos well.
- It can be confidently wrong, in fluent and plausible prose. Output that reads well is not evidence that it is correct, which is why serious deployments include a parallel run and an accuracy assessment.
- It cannot know anything it was not given. There is no access to internal systems unless someone connects them, and no memory between runs unless someone builds it.
- It costs money per run, which makes a cost model part of the design rather than an afterthought.
- Some work should not be automated at all, which makes a written list of exclusions part of the deliverable.
Those four limits are not caveats attached to the end of a pitch. They are the load-bearing half of an accurate description, and stating them early is what separates a scoped project from an expensive experiment.