Guide: AI in business › AI agents and automation
An AI agent is a program based on a language model that is given a goal, plans the next steps on its own and uses tools (a search engine, files, company systems) until the task is done. Agentic AI, i.e. AI that works as an agent, is the general name for this way of working: artificial intelligence not only answers, it also acts. For a company this means taking repetitive tasks off the team’s plate. It also means a new risk if the agent operates without human oversight.
In 60 seconds AI agents in brief
- Definition: an AI agent is given a goal, not a single instruction, and chooses the steps and tools itself.
- Agent vs assistant: an assistant answers each of your questions; an agent carries a task through from start to finish.
- Where it helps: in repetitive preparatory work in marketing and PR, as well as in sales and customer service.
- A sensible start: one task with a clear outcome, the least possible access to data and human approval before every external action.
- Pitfall: “agents” that are just ordinary chats, and projects with no measured effect.
AI agent: what it is and what agentic AI means
AI agent is a program you tell what to achieve, not how to do it. It is given a goal, for example “prepare a note on the client before tomorrow’s meeting”, and then works out for itself where to look for information, reaches for tools, assesses whether the result is good enough and decides on the next step, until the note is ready or the agent reaches a point where it has to ask a human. Sounds simple. The devil is in the limits.
Agentic AI is a broader term. It describes a way of using artificial intelligence in which the model doesn’t stop at an answer but plans and acts, and then improves itself in a loop. Looking for a short definition (agentic AI definition, w wyszukiwarce wpisywanej też odwrotnie jako AI agentic)? Here you go: it is AI that carries out multi-step tasks on its own with the help of tools, within limits set by a human.
We set out practical applications in marketing and communications, together with costs, in the article AI agent in marketing and PR. If you are only just sorting out the topic from the basics, start with our guide to AI in business, where agents are one of seven sections.
“Agenci AI” or “agenty AI”: which Polish form is correct
Both forms are heard in Polish companies. The PWN Language Advisory Service (Dr Rafał Sidorowicz, 28 February 2025) recommends the form agenty AI, because it is a computing term that refers to programs, not people. In the same way, Polish uses “klienty poczty” for email programs and “klienci” for people. The Language Advisory Service of the University of Warsaw gives a similar answer, noting that usage is still settling.
Kto pyta „agenty czy agenci AI”, powinien wiedzieć jedno: w mowie potocznej przeważa forma agenci AI. Nobody will call it a mistake in a sales conversation. We use both in this text, but in technical documents and contracts stick to one, the one your legal department adopts, because inconsistent terminology in a contract with a supplier means needless questions with every annex later on.
How an AI agent differs from an AI assistant and a chatbot
These three terms are often mixed up, including in vendors’ materials. The difference lies not in how “intelligent” the program is, but in who leads the work and what the program can do on its own, without asking anyone’s permission, in the systems it has been given the keys to. The comparison AI agent vs chatbot is easiest to see in a table. It will come in handy with every purchase.
| Chatbot | AI assistant | AI agent | |
|---|---|---|---|
| How it works | follows a script or a question-and-answer base | responds to each human instruction | plans steps and carries them out itself |
| Who leads | script | a person at every step | the agent; a human approves at agreed points |
| Tools | usually none | sometimes search and files | search, files, CRM, email, spreadsheets |
| Typical use | handling simple customer questions | writing, summarising, analysis | reports, monitoring, meeting preparation |
| Main risk | rigid, frustrating answers | a wrong answer | wrong actions in company systems |
What an agent is made of: model, tools, memory and loop
Every agent has four layers. Language model understands the goal and plans. Tools let it act, i.e. search the internet, read documents or write to a spreadsheet. Memory stores what it has already done and learned, and the loop means that after every step the agent assesses the result and decides for itself what to do next, instead of waiting for your next instruction. If it is to answer on the basis of company documents, searching those documents comes into play as well, which we describe in our article on RAG, i.e. AI working on company knowledge.
For an agent to use tools from different companies, common standards are needed. The most important is MCP (Model Context Protocol, a common way of connecting models to tools and data), developed by Anthropic. In December 2025 MCP, together with AGENTS.md from OpenAI and the goose tool from Block, moved to the new Agentic AI Foundation under the Linux Foundation. According to the foundation’s announcement, more than 10,000 MCP servers have already been published, and the protocol is used by Claude and ChatGPT, among others, as well as Gemini and Microsoft Copilot.
The takeaway for managers? An agent built on open standards is easier to move to another provider. Ask about this when choosing a tool.
Jak Commwork®AI trzyma agentów AI w wyznaczonych granicach
Agent bez granic i bez dziennika to ryzyko, o którym piszemy wyżej. Commwork®AI łączy agentów z programami, które firma już ma, zapisuje każde ich działanie i zatrzymuje się tam, gdzie potrzebny jest podpis człowieka. Zespół Commplace® pomaga wybrać pierwsze zadanie, a decyzję zawsze podejmuje człowiek.
- Oversight and fixes 34%
- Data preparation 28%
- Platform and licences 22%
- Language model 16%
- 25.06.2025Gartner: ponad 40% projektów agentowych do przerwania do 2027
- 9.12.2025Powstaje Agentic AI Foundation przy Linux Foundation
- Week 2Wybrane jedno zadanie, zmierzony czas pracy człowieka
- Week 4Spisane granice: co agent czyta, zapisuje i czego nie robi
- Week 7Test na danych z przeszłości, porównanie z człowiekiem
- Week 12Decyzja: zostawić, rozszerzyć albo wyłączyć agenta
- 01Wybór zadania
- 02Granice i uprawnienia
- 03Test na starych danych
- 04Wdrożenie z nadzorem
- 05Measuring the impact
- Łączy agentów z programami, które firma już ma, przez jedną bramę do modeli: klucze zostają po stronie serwera, a każde zapytanie trafia do dziennika.
- Pokazuje plan, zanim agent cokolwiek ruszy, i zatrzymuje się przed wysyłką, publikacją, zakupem czy nadaniem uprawnień, bo te czynności wymagają podpisu człowieka.
- Zapisuje, kto, kiedy i co zmienił, więc każdą zmianę da się cofnąć. Jak zbudować samego agenta, opisaliśmy w poradniku o tworzeniu agenta AI.
The view reproduced is modeled after the dashboards provided to customers. The names, numbers, and terms are examples and show the way the system works, not the results of a specific company.
Agentic AI in business: examples
The best examples of agentic AI (agentic AI examples) are boring, frequent tasks that are also well described. Here are a few applications we see in mid-sized companies:
- PR. Every day the agent reviews mentions, discards reprints and prepares a page with three issues for the spokesperson to decide on.
- Sales. Before a meeting it gathers information about the client from the CRM, the company website and recent publications. We show what this looks like in sales practice in connection with CRM with AI.
- Customer service. The agent checks the order status in the system and prepares a reply, which a consultant approves. How this differs from a classic chatbot is explained in our article on AI in customer service.
- Marketing. Once a week it compares competitors’ offers and announcements and records the changes in a single spreadsheet.
- Quality control. A separate agent checks whether every figure in the material has a source and a date before the text goes for approval.
Who AI agents in business are for, and in what situations
Agenci AI dla firm mają sens tam, gdzie zespół traci godziny na zbieranie i porządkowanie informacji, a na decyzje i rozmowy z ludźmi zostaje mu niewiele. Po czym to poznać? Te same raporty co tydzień składane ręcznie. Zapytania klientów, które czekają, aż ktoś je przeczyta. Informacje o rynku rozrzucone po kilkunastu źródłach, do których każdy zagląda na własną rękę, przez co nikt w firmie nie ma pełnego obrazu, a zarząd dostaje trzy różne wersje tej samej sytuacji od trzech różnych osób. To typowy objaw kłopotów z przepływem informacji w firmie i właśnie tu agent potrafi odciążyć ludzi najszybciej.
An agent won’t help when there is no strategy or the process is unclear. If nobody knows what should reach the customer and when, the agent will only speed up the chaos. In that case, start by putting the rules in order, for example with the basics of good communication with customers. Automation will come later.
How to introduce AI agents in your company, step by step
This order works regardless of the tool. We discuss specific platforms (ChatGPT, Copilot Studio, Google tools, n8n) in our guide on how to build an AI agent. Jeśli agent ma działać na programach, które firma już ma, z dziennikiem każdej zmiany i zgodą człowieka przed wysyłką, tak zbudowaliśmy Commwork®AI system.
- STEP 01Choose one task
Repetitive, with a clear outcome and measured human working time. One. Not five at once.
- STEP 02Set the limits
Write down what the agent can read, what it can write and what it will not do without human consent. This list always includes sending and publishing, as well as payments and deleting data.
- STEP 03Choose a tool that fits your ecosystem
A company on Microsoft 365 usually starts with Copilot, a company on Google Workspace reaches for Google’s tools, and one that has many systems from different vendors and doesn’t want to be tied to a single ecosystem chooses an automation platform.
- STEP 04Test on past data
Puść agenta na danych z ostatnich tygodni i porównaj wynik z pracą człowieka. Patrz na czas, na liczbę błędów i na to, czy wynik w ogóle nadaje się do użycia. Mierniki dobierz przed testem, a nie po nim; gotowe wzory znajdziesz w zestawieniu przykładów KPI.
- STEP 05Appoint an owner
One person reviews the agent’s activity log every week and corrects the instructions when data or procedures change.
How much AI agents cost and what drives the budget
The cost of an agent is the sum of four items. The model itself is often the cheapest. The most expensive part is people’s time for preparation and testing, and then for oversight, which teams usually only remember after the first month, when it turns out that someone has to read the agent’s log every week and correct its instructions.
Chcesz policzyć koszt agenta AI na własnych liczbach, zanim podpiszesz umowę z dostawcą? Na stronie AI in marketing opisujemy, jak na pierwszej rozmowie wybieramy jedno zadanie i ustalamy, ile godzin nadzoru będzie ono naprawdę wymagać.
| Item | What it depends on | How to reduce the cost |
|---|---|---|
| Language model | number and length of tasks, choice of model | a cheaper model for simple steps, a stronger one only for assessments |
| Platform and licences | tool, number of users, number of runs | use licences the company already has |
| Data preparation | state of documents and systems | start with a single data source |
| Oversight and fixes | how often procedures change, number of errors | stop points and samples checked by hand |
When an agent pays off, and when it does not yet
An agent makes sense when the task comes back every week, the data sits in one place and someone has time to review the results during the first month. Launch in a quieter period. Not in the week before a campaign or a product launch.
Hold off if the vendor cannot demonstrate the agent on your data or cannot say where the agent stops and who is responsible for its mistakes. That is a warning sign. We discuss the ethical and legal risks separately when writing about AI risks and ethics and on The AI Act in Poland. The EU regulation does not create a separate category for agents. Obligations depend on what the agent is used for.
Agent AI dla firmy w 90 dni, czyli scenariusz poglądowy
Poniższy opis to example, a nie historia konkretnego klienta. Pokazuje, jak zwykle przebiega wdrożenie agentów AI w firmie średniej wielkości, w której agenci AI w marketingu mają odciążyć czteroosobowy zespół. Terminy są orientacyjne. Mogą się przesunąć. Jeśli dane leżą w wielu systemach, wszystko potrwa dłużej.
Pierwszy miesiąc. Zespół wybiera jedno zadanie, na przykład cotygodniowy przegląd komunikatów konkurencji, i przez dwa tygodnie mierzy, ile czasu zajmuje ono człowiekowi. Ktoś z IT spisuje granice: agent czyta strony konkurencji i jeden arkusz, zapisuje tylko do jednego pliku i niczego nie wysyła. Nic więcej. Powstaje też lista miejsc, w których agent ma się zatrzymać i zapytać. Najczęściej trafiają na nią liczby bez źródła, bo od zmyślonej liczby zaczyna się większość wpadek, co dokładnie opisaliśmy przy okazji halucynacji AI.
Drugi miesiąc. Agent pracuje na danych z ostatnich tygodni, a człowiek wykonuje to samo zadanie równolegle. Porównujecie czas, liczbę poprawek i to, czy wynik da się przekazać dalej bez przepisywania. Czasem wychodzi, że zadanie nie wymaga planowania i wystarczy zwykły przepływ bez agenta. Taką granicę dobrze widać w poradniku o automation with AI and n8n.
Trzeci miesiąc. Agent działa na bieżąco, ale wszystko, co ma wyjść poza firmę, zatwierdza człowiek. Właściciel agenta raz w tygodniu czyta dziennik i poprawia instrukcję. Brzmi nudno? I dobrze. Po dwunastu tygodniach zapada decyzja: zostawić, rozszerzyć o drugie zadanie albo wyłączyć. Żeby nie pomylić przypadkowej poprawy z efektem agenta, porównaj wynik z okresem sprzed startu i z podobnym zadaniem, którego agent nie dotykał. Zaplanuj też ludziom czas na nadzór, bo bez tego dziennik przestaje być czytany już w piątym tygodniu. Podpowiedzi znajdziesz w tekście o planowaniu pracy zespołu.
Próg decyzji ustal przed startem. Przykładowa reguła brzmi tak: agent zostaje, jeśli oszczędza co najmniej połowę czasu człowieka przy tej samej liczbie błędów, a przez ostatnie cztery tygodnie ani razu nie wyszedł poza swoje uprawnienia. Nie spełnia progu? To nie porażka, tylko tania informacja, że to zadanie jeszcze się dla agenta nie nadaje. Na wybór narzędzia patrzy się potem inaczej, czego przykład dajemy w zestawieniu AI assistant in the company.
Zanim podpiszesz umowę, zadaj dostawcy sześć pytań. Odpowiedzi zapisz w notatce ze spotkania, bo przydadzą się przy pierwszym sporze o to, kto odpowiada za błąd.
- Pokażecie agenta na naszych danych, a nie na demonstracji? Co dokładnie zrobi w pierwszym tygodniu?
- W których miejscach agent się zatrzymuje i prosi o zgodę? Czy tę listę zmienimy sami?
- Gdzie trafiają dane, które agent czyta, i kto po Waszej stronie ma do nich dostęp?
- Jak wyłączyć agenta jednym ruchem i jak cofnąć jego ostatnie zmiany?
- Czy rozwiązanie korzysta z otwartych standardów, takich jak MCP, czy wiąże nas z jednym dostawcą?
- Kto odpowiada za błąd agenta i jak zapisujemy to w umowie?
Frameworks, courses and standards: what a manager should know
Agent frameworks (agentic AI frameworks, ready-made libraries for programming agents) are a matter for the technical team. Google offers ADK (Agent Development Kit), Block the open tool goose, and open-source libraries such as LangGraph and CrewAI are also popular. A manager doesn’t need to know them. What they do need to know is whether the chosen solution uses open standards, where the data the agent reads and writes ends up, and how to switch the agent off in one move when it starts doing something nobody planned.
Looking for a course (agentic AI course)? Choose one that teaches you how to design tasks, oversight and measurement, not just how to click around in a single tool, because tools change every few months. We suggest what to look for when choosing in our piece on AI courses with a certificate. Want to understand first what AI really does in marketing and what it doesn’t? Have a look at the page AI in marketing. Wolisz nie przechodzić tej drogi samodzielnie? Zespół AI Commplace pomoże wybrać pierwsze zadanie dla agenta i ustalić, po czym poznacie, że się opłaciło.
Five mistakes that land agent projects in a drawer
The company pays for an agent and gets a chatbot with a new name. The cost? A licence with no effect and a team discouraged from further projects.
Several agents launched at once are hard to test and compare with human work. It ends in a project whose effect cannot be proven.
An agent with access to email and the CRM, plus the ad accounts, can send the wrong offer or release a budget with a single mistake. Cut permissions. Every unnecessary one is a risk with no benefit whatsoever.
After an agent is deployed, the number of enquiries rises and the team immediately credits the agent. Until you compare the result with the same months a year earlier, or with a group that worked without the agent, you have a hunch, not proof.
An agent without a responsible person works worse month by month, because data and procedures change. After six months nobody trusts it and the project is abandoned.
Agentic AI demystified: questions from our conversations with boards
How does an AI agent differ from an AI assistant?
An AI assistant responds to each instruction and waits for the next one, so a human leads the work. An AI agent is given a goal, plans the steps on its own and reaches for tools. A human approves the result at agreed points.
What is agentic AI?
Agentic AI is a way of using artificial intelligence in which the model not only answers but also plans and carries out multi-step tasks with the help of tools. It operates within limits set by a human and in a loop: step, assess the result, next step.
“Agenty AI” or “agenci AI”: which Polish form is correct?
The PWN Language Advisory Service (28 February 2025) recommends the form “agenty AI”, because it refers to programs, not people. In everyday speech “agenci AI” is heard more often, and this form is also widely understood.
How does an AI agent differ from a chatbot?
A chatbot answers according to a script or a question database and usually does not take actions in systems. An AI agent plans the steps on its own, reaches for tools (CRM, email, spreadsheets) and sees the task through to the end.
Are AI agents safe for a company?
They are as safe as the limits you set for them. Watch three things. The minimum necessary access to data, human approval before sending, publishing or paying, and a log of all the agent’s actions.
What is the Agentic AI Foundation?
It is a foundation set up on 9 December 2025 under the Linux Foundation to develop open standards for AI agents. Its projects include the MCP protocol from Anthropic, AGENTS.md from OpenAI and goose from Block, and its founding members include Google, Microsoft, AWS and OpenAI.
Will an AI agent replace employees?
No. The agent takes over repetitive preparatory work, i.e. gathering data, first drafts of reports and organising information, while decisions, customer relationships and responsibility for what leaves the company stay with people, and it is they who determine whether the agent’s work has any value at all.
Sources
- SOURCE
- SOURCELinux Foundation: Formation of the Agentic AI Foundation (AAIF) · 9 December 2025
- SOURCEPWN Language Advisory Service: Agenty AI (Dr Rafał Sidorowicz) · 28 February 2025
- SOURCEUniversity of Warsaw Language Advisory Service: agenty AI or agenci AI · accessed 24 September 2026
- SOURCEGoogle Cloud: Gemini Enterprise Agent Platform (ADK) · as of 24 September 2026
Read next
Na pierwszej rozmowie wybieramy jedno powtarzalne zadanie, ustalamy granice agenta i sposób pomiaru. Dostajesz krótką listę zadań, od których warto zacząć, i plan sprawdzenia efektu.
Sebastian Kopiej, President of Commplace®. In public relations since 1996. Text prepared with the help of AI tools and fact-checked by the author. Data current as of September 29, 2026.