What an AI agent is — and what it isn't

A Brazilian building-materials distributor, 380 people, four distribution centers. The service desk takes in around 900 requests a week across email, WhatsApp and the website form. Half are duplicate invoices, delivery dates or returns; the other half needs someone who knows the customer's contract. Today everything lands in the same queue and is read by three people, in the order it arrives.

The company bought a tool marketed as an AI agent. In practice it answered questions when someone opened a chat window and pasted the text in. Nobody did anything wrong: the tool worked. It just didn't do what the problem called for, and the difference between the two has an old technical name.

01The definition that came before the hype #

Stuart Russell and Peter Norvig published Artificial Intelligence: A Modern Approach in 1995, now in its 4th edition, from 2020. The book's second chapter defines an agent in dry, useful terms: an entity that perceives its environment through sensors and acts upon that environment through actuators, choosing its action based on what it has perceived and the goal it pursues.

Note what the definition requires and what it doesn't. It requires perception, action and a performance measure. It doesn't require a language model, doesn't require conversation, and doesn't require intelligence in the everyday sense of the word. A thermostat satisfies the definition poorly; a system that reads the request queue, decides the routing and executes satisfies it richly.

Michael Wooldridge and Nicholas Jennings, in that same year of 1995, added the properties the literature went on to use as a yardstick: autonomy (acts without direct intervention at every step), reactivity (perceives changes in the environment and responds), pro-activeness (takes initiative toward a goal) and social ability (interacts with other agents and with people). That list is still the most honest test of what a vendor is selling.

02Automation, assistant and agent #

Three different things get sold under the same word. Telling them apart settles half of any buying conversation.

1. Automation. Does what you told it, in the order you wrote. If condition A happens, run B. It decides nothing: you decided everything up front, when you designed the rule. It is predictable, cheap to audit, and it breaks the moment a case the rule never anticipated shows up.
2. Assistant. Answers when asked. Drafts, summarizes, classifies, suggests. The trigger is always human and the output passes through a person before it becomes an action. It shortens a task; it doesn't reduce how often someone has to remember to do it.
3. Agent. Takes a goal and a scope, perceives the state of the process, decides the next step within that scope and acts — including writing into systems. The human sets the boundary and the success criterion, not each step.

The practical difference between an assistant and an agent is who remembers. With an assistant, someone has to notice there is work to do and ask for help. With an agent, noticing there is work to do is part of the job. In an operation where things fall into the gap between departments, that is exactly the missing part.

Technically, most of today's agents pair a language model with tools: the model reasons about the state, picks a tool, observes the result and decides again. Work by Shunyu Yao and colleagues, presented at ICLR 2023, formalized that loop of alternating reasoning and acting. It is worth knowing for a practical reason: an agent's quality depends less on the model and more on the tools and the data you gave it.

03Three concrete uses, and what each demands #

Back to the distributor. The three cases below are the ones that tend to survive the first quarter at a company of that size.

Want to see how this looks inside a real operation? Explore the platform.

• Request triage. The agent reads the incoming request, classifies it by type and urgency, checks the customer's contract, routes it to the right queue and answers whatever is standard. Requires: a labeled history of requests, read access to customer records, and an explicit rule for when not to decide alone.
• Chasing open items. The agent tracks items with deadlines, spots what is stalled, finds the owner, nudges them on the channel where they actually work, and logs the reply. Requires: deadlines stated on the item itself and a named owner at every step — without that there is nothing to chase.
• Portfolio synthesis. The agent reads progress across dozens of workstreams and produces, every Monday, the picture of what changed, what slipped and what needs a decision. Requires: work recorded where it happens, not a report typed by hand on Friday.

All three share the same precondition, and it isn't technological: the process has to exist explicitly. An agent dropped onto a muddled process produces confusion faster — the contemporary version of Hammer's warning in automating a bad process just speeds up the error.

04What an agent is not #

An agent is not a digital employee: it holds no accountability, and accountability stays with whoever configured it. It is not a substitute for a process: it operates inside one. It is not neutral about data: in Brazil, an agent reading personal data falls under the LGPD like any other processing. And it is not trustworthy by default — the NIST AI Risk Management Framework 1.0, published in January 2023, treats trustworthiness as a property built through governance, measurement and documentation, not as a feature of the model.

The reverse holds too: agent is not a synonym for high risk. An agent with a narrow scope, broad read permission and minimal write permission is more auditable than the spreadsheet three people edit at the same time today.

05What to do this week #

1. Pick a flow where work stalls because nobody remembered, not because capacity ran out. That's where an agent pays off most.
2. Write the scope in one sentence: what it may read, what it may write, when it stops and calls a person.
3. Measure the baseline for two weeks before switching anything on: volume, response time, rework.
4. Name the human owner. With no name on the line there is no agent in production — there is an experiment.
5. Sort what you already have into the three categories. Much of what the company calls AI is automation that was already working.

At Relevanti, agents live inside the processes and operations modules, with an owner, a scope and a record for every action — you can see how that fits together on the platform, by department under solutions, or bring a specific flow to a conversation.

The 1995 definition is still the best filter available: perceives, decides, acts. What doesn't perceive its environment on its own is an assistant; what decides nothing is automation. All three are useful, and none of them removes the need for someone to answer for the result.