Definitions and comparisons
AI agents for business: what they are, what they actually do and when they pay off
The word "agent" has ended up everywhere, and it means different things depending on who says it. Here we define it in a way you can check, we tell it apart from the automations that already exist, and we put it in front of the numbers: how many companies use them, how many projects reach production, what they cost.
Written by Maximilian Giurastante. . 13 minute read.
An AI agent is a program that uses a language model to reach a goal it has been given: it reads the context, chooses which action to take among the tools it has available (search, read a document, call an API, write), looks at the result and repeats until the task is done or it stops to ask. It differs from a chatbot because it acts, and from rule based automation because it decides the steps instead of running them in a fixed order.
In short
- An agent decides the steps; an automation runs them in a fixed order. The difference is only useful when the input varies and the steps cannot be listed in advance.
- Most business problems are solved with a rule based flow where a model does one single step (read, classify, write): it costs less and is easier to control.
- According to McKinsey (November 2025) 62% of organizations are experimenting with agents and 23% are scaling at least one; the gap between the two figures is the number to keep in mind.
- Gartner forecasts that more than 40% of agentic AI projects will be canceled by the end of 2027, because of costs, unclear value or inadequate controls.
- An agent pays off when the input varies, the cost of an error can be checked and the tools it has to use expose an API. If one of these conditions is missing, something else fits better.
- In Italy 16.4% of companies with at least ten employees use AI (ISTAT, 2025) and AI workflow automation sits at 2.9% of the total: there is more room than competition.
What an AI agent is, in words you can check
An AI agent is a loop, not an answer. It takes a goal, looks at what it has in front of it, picks an action among those allowed, runs it with a tool, reads the outcome and decides the next step. The loop ends when the goal is reached, when the agent stops to ask, or when it runs out of the attempts it was given.
The loop of an agent
Five step loop: take the goal, read the context, pick the next step, use a tool, check the result. From the check it goes back to reading the context, or it delivers the result or asks a person.
- 1. Goal: set by a person
- 2. Read the context: data, documents, results
- 3. Pick the next step: from the allowed actions
- 4. Use a tool: API, search, writing
- 5. Check the result: deliver or ask
Three words in this definition make the difference from the tools companies already use. Goal: you tell the agent what to achieve, not how. Tools: the agent does not just write text, it can act on real systems (read a mailbox, query an ERP, open a ticket). Check: after each action the agent reads the outcome and can correct itself. A program without these three traits is not an agent, whatever its sales sheet says.
Three levels of automation, and what changes between them
The choice is not between "doing it by hand" and "using an agent". In between there are two levels that solve most cases and cost less. Picking the right one is the first decision in a project.
Who decides the steps
Three columns. Rule based automation: event, fixed steps, result. LLM workflow: event, fixed steps with one of them handed to a model, result. Agent: goal, a loop where the model decides and acts, result or a question to a person.
- Rule based automation: Event (file, email, schedule), Fixed steps (if, then, else), Result (same output for same input). Predictable, cheap, blind to a new input
- Workflow with an LLM: Event (email, document, request), One step to the model (read, classify, write), Fixed steps and result (with a human check). Handles unstructured input, stays controllable
- Agent: Goal (and the tools allowed), Loop: decide, act, check (several calls to the model), Result or question (stops when it does not know). Handles the unexpected, costs more, needs limits
Comparison on the criteria that matter in a quote
| Criterion | Rule based automation | LLM workflow | Agent |
|---|---|---|---|
| Who decides the steps | Whoever writes the rules | Whoever designs the flow; the model handles one step | The model, within the tools and limits it was given |
| Input it can handle | Only what was foreseen | Unstructured text and documents | Also situations nobody planned for |
| Predictability | Total | High on the flow, variable on the AI step | Low: it has to be measured on real cases |
| Cost per run | Close to zero | One or two calls to the model | From a few to dozens of calls |
| Human control | Not needed | One check point in the flow | Stop points decided at design time |
| When to use it | Stable process and structured data | Stable process and unstructured data | Open tasks with several possible routes |
What agents actually do in a company, with examples from our projects
The examples below come from systems we built or use every day. We also say when the right answer was not an agent, because that is the piece of information people are missing most.
Reading reports and opening tickets
In SYNAPSE, the platform a systems integrator uses to keep device fleets under control, reports that arrive from an external system turn into jobs on their own. This is not an agent: the format is known, and a rule based flow is more reliable and costs nothing per run. A language model would only come in for reports written by hand, as a single extraction step inside the same flow.
Putting together a report from data and documents
An agent that receives the goal "prepare the monthly absence summary by department" has to query the database, work out which departments have anomalies, decide whether a comparison with the previous month is needed, and write. Here the loop earns its place: the steps change with the data. In LeavePilot the generated report works only on aggregated data, never on names: the most effective rule for the GDPR and the AI Act is to not give the model the data it must not use, instead of asking it not to use them.
Writing and checking code
Software development agents are today the most mature case: they take a task, read the existing code, change files, run the tests, read the errors and fix them. We use them every day, with written rules on what they can do without asking and what they cannot; CLACOROO, our open source panel for Claude Code, was born to govern plugins, permissions and sessions of these agents. What we learned there holds for any agent: the stop points matter more than the instructions.
Sorting requests and answering the recurring ones
A flow that reads incoming requests, classifies them by type and urgency, answers the ones with a known reply and passes the rest to a person with a draft ready. It is an LLM workflow, not an agent, and it is the shape people ask for most because the return is measured in hours: how many requests a day, how many minutes each. It becomes an agent only if, in order to answer, it has to check different systems and decide which ones.
The numbers: who uses them, how many projects reach production
Figures on agent adoption come in two kinds: those that count who is experimenting and those that count who is in production. The first are high, the second are not. The gap is the useful number.
Use of AI and of agents in organizations
Bar chart with four values from the McKinsey survey of November 2025: 88% use AI in at least one function, 62% are experimenting with agents, 39% report an impact on operating profit, 23% are scaling agents in at least one function.
- Uses AI in at least one function: 88%
- Experimenting with AI agents: 62%
- Reports an impact on operating profit (EBIT): 39%
- Scaling agents in at least one function: 23%
In the next edition of the same survey (August 2026, 1,719 respondents) the share of those who credit AI with at least some impact on operating profit is stuck at 37%, and those scaling agents remain about two in ten overall, with 40% among companies above one billion dollars in revenue. In one year the experimenting has widened; production has not, or barely.
- > 40%
- of agentic AI projects will be canceled by the end of 2027 (Gartner, June 2025: rising costs, unclear value, inadequate controls)
- 95%
- of organizations with no measurable return from generative AI pilots (MIT NANDA, July 2025, preliminary findings on 52 organizations)
- 16.4%
- of Italian companies with at least 10 employees use AI (ISTAT, December 2025; it was 8.2% in 2024)
- 2.9%
- of Italian companies automate workflows with AI (Eurostat, 2025 data for Italy)
Two warnings on the most quoted figures. The Gartner forecast concerns projects that start without a clear use case: it is a warning about selection, not about the technology. The MIT 95% comes from a preliminary study on a small sample and measures the absence of impact on the income statement, not technical failure; the same report notes that generic tools reach production far more often than badly built custom ones. Read together, they say one thing only: the problem is not making an agent work, it is choosing where to put it.
When an agent pays off, and when it does not
Two variables are enough to decide: how varied the input is and how much an error costs. On the input axis you choose between rules and model; on the error axis you decide how much human control is needed.
Where the solutions sit
Two by two matrix. Horizontal axis: input variability, from low to high. Vertical axis: cost of an error, from low to high. Low variability and cheap error: rule based automation. High variability and cheap error: agent with sample checks. Low variability and expensive error: workflow with one AI step and human control. High variability and expensive error: agent with human approval on every action, or no automation at all.
- Input variability: low: structured data / high: free text, edge cases. Cost of an error: low / high.
- Workflow with one AI step: and a human check before any effect: the model reads or writes, a person confirms.
- Agent with approval: every action with real effects has to be approved. If the volume cannot take that, better not to automate.
- Rule based automation: no model: a flow written once and checked against real data.
- Agent with sample checks: the point where an agent pays off most: varied input, recoverable errors, high volume.
The five conditions, all of them necessary
- The input is varied. Free text, different documents, requests that do not fit in boxes. If the input is a form, rules are enough.
- The steps cannot be listed in advance. If you can draw them in a diagram, a workflow costs less and is easier to control.
- The tools expose an API. An agent that has to click around an interface designed for people is fragile by construction.
- The error can be spotted and undone. An agent will get things wrong; you need to know how you notice and how you roll back.
- The volume justifies the cost. Every turn of the loop is a call to the model. Ten runs a month do not pay back the design work.
How to build an agent that holds up in production
The technical work on the model is the smaller part. The larger part is setting the limits: what it can do on its own, when it has to stop, how you measure whether it is working well.
- Write the goal and the stopping criterion What it has to achieve and how you can tell it is done. "Answer the customers" is not a goal; "classify every request, answer the type A ones with the approved template, pass the rest on with a draft" is one.
- List the tools and their limits Which APIs it can call, with which permissions, with which quantity limits. An agent that can read the ERP but cannot write to production is an agent that can be wrong without doing damage.
- Decide the stop points Actions with external effects (sending, paying, deleting, promising) need confirmation until you have measured that the agent handles them well. It is the rule we follow for development agents too: never a release without a person approving it.
- Build a set of real cases for evaluation Fifty cases taken from reality, with the expected answer. Every change to the instructions or to the model gets tried on those cases before going to production. Without this step you cannot tell whether a change makes things better or worse.
- Release with supervision and measure for a month Share of cases solved without intervention, average time, errors detected, cost in calls. These are the four numbers that decide whether to widen the perimeter or narrow it.
- Document for the AI Act and the GDPR If the agent interacts with people it has to declare itself (Article 50, applicable from 2 August 2026); whoever uses it in a company has to be able to prove staff AI literacy (Article 4). Personal data that goes into prompts requires a contract with the model provider. Our AI Act guide (in Italian) explains what to produce.
What it costs
An agent has two costs: building it and running it. The first depends on how many tools it has to connect to and how much control is needed; the second on the number of runs and of calls per run, and it is falling fast.
| Item | Order of magnitude | What pushes it up |
|---|---|---|
| Workflow with one AI step on a narrow process | from 1,500 to 6,000 euros, one to three weeks | Number of systems to connect, volume, human control at every step |
| Agent with several tools and stop points | from 5,000 to 15,000 euros, four to eight weeks | Tools to integrate, evaluation set, permissions and logging of actions |
| Running cost: calls to the model | pay per use, from a few euros to a few hundred a month | Number of runs, length of the documents, loop turns per run |
| Running cost: infrastructure | from a few tens of euros a month | Queues, log storage, separate environments |
On running costs the trend is clear: according to the AI Index 2025 from Stanford University, the inference cost for a system performing at the level of GPT-3.5 fell more than 280 times between November 2022 and October 2024. The cost of an agent in production is therefore more and more a question of design (how many turns, how many documents) than of price per call. The orders of magnitude above are the same ones on our page how we work and what it costs, which also explains what you need to get started.
Frequently asked questions
What is the difference between an AI agent and a chatbot?
A chatbot answers questions, one at a time, and does not act on other systems. An agent receives a goal, uses tools (APIs, searches, documents) to reach it, checks the results and repeats until it is done or has to ask. A chatbot that can also open a ticket or query an ERP is on its way to becoming an agent.
Can an AI agent replace an employee?
No, and the projects set up that way are the ones Gartner expects to be canceled. An agent replaces repetitive steps inside a process: reading, classification, extraction, drafts, checks. Exceptions, decisions with external effects and accountability stay with people, by law as well (the AI Act requires human oversight and transparency).
How long does it take to build an AI agent for my company?
A workflow with one step handed to a model on a narrow process takes one to three weeks. An agent with several tools, stop points and a set of cases for evaluation takes four to eight weeks. The variable that moves the schedule most is the number of systems to connect, not the model.
Are AI agents reliable?
They are, as far as they are bounded and measured. An agent with limited tools, stop points on the actions with external effects and a set of real cases it is evaluated on at every change is as reliable as any other software. An agent left free on a vague task is not, and that is the reason for the gap between who experiments and who is in production.
What do I have to do about the AI Act if I use an agent?
It depends on your role. If you use an agent built by someone else you are a deployer: you have to guarantee staff AI literacy (Article 4) and, if the agent interacts with people, make sure it declares itself as an AI system (Article 50, from 2 August 2026). If you have it built under your own brand you become a provider, with further obligations. The AI Act guide on our site has an eight question test to work out where you fall.
Which language model should I use?
The one that passes the evaluation on your real cases at the lowest cost, with a contract that rules out training on your data and that handles personal data under the GDPR. The name of the model matters less than the contract and the set of cases you test it on. Changing model after a year is normal, if the agent is built to allow it.
Sources
The figures quoted in the text, each linked to the original document. Numbers are reported as published by the source, with the year they refer to.
- McKinsey, The state of AI in 2025: Agents, innovation, and transformation: Survey of 1,993 respondents in 105 countries, published on 5 November 2025: 88% use AI in at least one function, 62% are experimenting with agents, 23% are scaling them, 39% report an impact on EBIT.
- McKinsey, The state of AI in 2026: On the road to ROI: Survey of 1,719 respondents in 97 countries (May and June 2026), published on 25 August 2026: 37% credit AI with at least some impact on EBIT; about two organizations in ten are scaling agents.
- Gartner, Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027: Press release of 25 June 2025, based on a January 2025 survey with 3,412 participants.
- MIT NANDA, The GenAI Divide: State of AI in Business 2025: Preliminary findings, July 2025: interviews in 52 organizations, 153 questionnaires, more than 300 initiatives reviewed. 95% of organizations record no measurable impact on the income statement from generative AI pilots.
- ISTAT, Enterprises and ICT, 2025: Release of 15 December 2025: 16.4% of companies with at least 10 employees use at least one AI technology (8.2% in 2024); large companies at 53.1%, SMEs at 15.7%.
- Eurostat, Use of artificial intelligence in enterprises: 2025 data: 20.0% of EU companies use AI; in Italy AI workflow automation covers 2.9% of companies with at least 10 employees.
- Stanford HAI, AI Index Report 2025: The inference cost for a system at the level of GPT-3.5 fell more than 280 times between November 2022 and October 2024.
- Regulation (EU) 2024/1689 (AI Act), Articles 4 and 50: Staff AI literacy and transparency obligations for systems that interact with people. The guide on our site carries the application dates updated to the Digital Omnibus.
Written by Maximilian Giurastante
Founder and software developer, MAXYMIZE. Designs and builds SaaS platforms, custom business software and automations with language models. Over twenty years as a project leader in technology, audio and video systems integration.
Do you have a process that could take an agent?
Tell us about it in thirty minutes. We will tell you whether you need an agent, a workflow with one AI step or none of the above, and in any case what it would cost.