What Is an AI Agent? Six Definitions Side by Side, and What Works in a Retail Chain

"AI agent" is the most used term in business software this year, and one of the least defined. Gartner estimates that only about 130 of the thousands of vendors selling agentic AI offer the real thing, and expects over 40% of agentic AI projects to be canceled by the end of 2027 over costs, unclear value or weak risk controls⁵.

Before deciding whether an agent belongs in your stores, it helps to know what the people who coined the term mean by it. Here are their definitions, quoted from their own documents.

Six definitions of an AI agent

Read together, they agree on three parts: a goal, tools that act rather than only talk, and some independence from a person^{1, 2, 3, 4}. The EU AI Act, the law that will regulate much of this in Europe, does not use the word agent at all; it regulates AI systems that operate with varying levels of autonomy⁶.

They differ on the third part, and that is the one that matters to a retailer. An agent that drafts a purchase order and one that sends it to the supplier can run on the same model. What separates them is how much they are allowed to do without asking, and researchers at the University of Washington argue that this is a choice the deployer makes, separate from what the model can do⁷.

Agent, chatbot, copilot or workflow?

A chatbot answers. A copilot answers from your data and suggests; the guide to AI copilots for retail operations covers what makes those answers trustworthy. A workflow runs steps someone fixed in advance, such as a nightly reorder rule. An agent decides the next step itself and uses tools to take it². In practice most useful systems mix the last two: fixed rules do the arithmetic, and the model handles the parts that need judgement and explanation.

Where companies are with it: in McKinsey's 2026 survey, the share of organisations with more than $1 billion in revenue that scale AI agents in at least one function rose from 27% to 40% in a year, while smaller organisations stayed flat at 22%. Respondents in consumer goods and retail most often use agents in marketing and sales⁸.

Autonomy is a setting: grant it per action

Researchers describe the person's role as the agent's independence grows: operator, collaborator, consultant, approver, and finally observer, who can only switch it off⁷. OpenAI's own guide adds the practical rule: rate each tool by whether it only reads or also writes, whether the action can be reversed, and its financial impact, and keep a person in the loop for actions such as canceling orders, authorising large refunds or making payments¹.

For a store chain that means one agent can hold different authority for different actions. It may read stock freely, advise on transfers, prepare purchase orders for approval, and still need a person before anything reaches a supplier. A ladder of five levels, granted per type of action, makes this explicit:

What agents still get wrong

The case for limits is in the evidence. In TheAgentCompany, a benchmark of 175 tasks in a simulated company, the best agent tested in the paper finished 30.3% of them on its own⁹. When Anthropic let a model run a small office shop for about a month, it sold items below cost, handed out discount codes on request and told customers to pay into an account it had made up; its makers concluded they "would not hire" it¹⁰.

The second run is the more useful one. Procedures, cost visibility and a rule that a person checks every purchase removed most of the losses, and the team summed it up as "bureaucracy matters"¹¹. That is a retail lesson: the controls a chain already applies to a new buyer, such as limits, approvals and a review after the first month, are what make an agent safe to use.

Six rules for AI agents in a store chain

1. Let fixed rules compute quantities and money; let the model explain them and handle exceptions.
2. Give each agent one measurable goal and the limits it may not buy that goal with, such as availability without inflating stock.
3. Grant authority per type of action, and let only an owner raise it, after a probation with clear thresholds.
4. Lower it automatically when a limit is breached or the data goes stale.
5. Keep a stop button and a record of every action. For high-risk systems the EU AI Act requires that people can override or reverse the output and halt the system⁶.
6. Measure the effect against control stores that did not get the agent's actions, not against last month.

The last rule is the one most often skipped. Shelf availability moves with the season for reasons that have nothing to do with software; only a comparison with similar stores in the same weeks separates the agent's effect from the season. The out-of-stocks guide shows how availability is read store by store.

How marql's AI employees apply this

marql calls its agents AI employees: six roles, from supply to finance, each with one KPI and limits it may not break. Authority uses the five levels above, set per type of action. Only the owner or COO raises a level, after a probation the product checks every day: to move from advising to preparing, four weeks at the level and at least 70% of its advice accepted by people; to execute after confirmation, a passed exam on a retail simulator and no more than 10% of draft lines changed by people. The top level is closed for every customer.

Quantities come from fixed rules; the language model explains them. The level an employee actually works at is the lowest of what was granted, what the ERP connection allows and whether the data is fresh: with e:corg, the ERP creates the document itself; with other ERPs a person enters the approved draft. Every employee can be paused, and it drops a level by itself when its limit is breached.

In a fifth of the stores each employee only advises, so its KPI can be compared with stores where it does not act; a chain with fewer than five stores sees the KPI as an observation, not an effect. The screen itself says it plainly: a completed action is not yet a measured benefit. The AI supply employee comes first, because fixing demand lost to an empty shelf is where the most is ready.