An AI agent is a system that is given a goal rather than a single question: it pulls data from connected systems itself, carries out a sequence of steps and returns a result. In marketing it looks like this: "every morning compare the cost per customer with last week and tell me if it has risen noticeably" — from there the agent gathers the figures from ad accounts, analytics and the CRM on its own.
Key takeaways
An assistant answers a question and stops. An agent is given a goal and decides for itself which steps to take to reach it — and takes them inside the connected systems.
| Aspect | AI assistant | AI agent |
|---|---|---|
| What it receives | A single question | A goal and the rules for reaching it |
| How it works | Question, processing, answer | Goal, data gathering, analysis, actions, checking the result |
| Access to systems | Only what a person pasted into the chat | Connected sources: advertising, analytics, CRM, knowledge base |
| What it does with the result | Shows it to a person | Can write into the CRM, create a task, send a notification |
| What has to be configured | The wording of the request | Access rights, limits on actions and a log |
Assistants and chatbots are the neighbouring topic: how they are built, how a knowledge base is connected and where they go wrong is covered in our article on AI assistants and chatbots for business. What sets an agent apart is not the model but what it is allowed to do.
The level is chosen by the cost of a mistake: the more an incorrect action costs, the closer the decision stays to a person.
The higher the level, the more three things matter: rights granted by a list of objects, a log of every action, and a single key that stops the agent. OWASP lists excessive agency as a distinct risk for applications built on language models: an agent does exactly what it was allowed to do, including the parts you did not intend.
In practice one scenario rarely sits at a single level: the agent pauses an overspending ad by itself but sends a budget increase for approval. The line is drawn by the size and the reversibility of the action, not by how much the model is trusted.
The candidates are always the same: the task repeats, it leaves a digital trail, and its result can be seen in numbers. Ten such tasks, from the simplest to the most connected.
The difference between an agent doing the work and an agent helping matters here. In the first four tasks it works alone and hands over a finished result. In the rest it prepares the material, while the decision — to write to the customer, change the scenario, answer the review — stays with an employee.
Some of these tasks are covered by ordinary rules, with no model involved: a mailshot triggered by an event, a task created by deal stage, an alert when a threshold is crossed. Where the line between a rule and a model runs is covered in our article on business process automation with AI.
The value of an agent in marketing is not the summary but the diagnosis: it walks the whole chain and says at which step things went off course.
The cost per enquiry rises for many reasons: the audience shifted, page conversion fell, the form broke, data stopped reaching the CRM, the auction got more expensive. An agent with access to ad accounts, analytics and the CRM goes through those steps in turn and shows where the numbers diverged.
Example "Cost per lead is up 37%. The change is not in paid traffic but at the form step: page conversion fell from 8.1% to 5.9%. Check the enquiry form." The figures here are illustrative — what matters is the shape of the message the agent delivers in the morning instead of a spreadsheet.
For that kind of diagnosis to be possible, advertising, the website, the CRM and payments have to be tied together. How that works technically is covered in our article on end-to-end analytics.
These two scenarios are where most companies start: they deliver something every day and risk almost nothing.
The agent assembles the report itself: spend from ad accounts, visits from analytics, enquiries from the CRM and payments from the accounting system, brought together into one summary with the changes explained in plain words.
Example "Enquiries are up 18%, but the share of qualified ones fell by 7 percentage points. Most of the drop is in the broad-audience campaign; check the audience settings." The figures are illustrative; what matters is the form: the change, the likely cause, the next step.
The delivery format is settled before launch: the summary arrives in Telegram, by email or as a task in the CRM, and its length is chosen so that people actually read it. A report three screens long stops being opened in the second week.
Reputation monitoring works the same way: the agent collects mentions and reviews, reads the sentiment, picks out recurring complaints and passes the case to the right person by creating a task in the CRM. A neighbouring task is a company's visibility in the answers of AI assistants; which metrics are counted there is covered in our report on AI visibility.
The line is drawn by the cost of a mistake and by law. Anything touching money, customers' rights or public reputation stays with a person.
Important Law No. ZRU-1115 of 21 January 2026 prohibits relying exclusively on the conclusions of systems built on artificial intelligence technologies when taking legally significant decisions that affect a person's rights and freedoms, and sets a fine of fifty to one hundred base calculation values for unlawful processing of personal data using such technologies.
Personalised mailings are processing of personal data under law No. ZRU-547 on personal data. Consent, a stated purpose of processing and a way to unsubscribe are required whoever prepared the email: an employee or an agent.
Public text is a boundary of its own. A reply to a review, a comment or a post in the company's name goes through a person: the model does not feel the context of a conflict and answers formally where an apology is needed.
An agent works with the data you already collect. If campaigns are named inconsistently and the enquiry source is never filled in inside the CRM, it will assemble a report with the same gaps — only faster.
Preparation usually takes longer than connecting the model, and it is what decides the quality of the result. Part of this work is covered by CRM configuration: which scenarios to automate there first is covered in our article on CRM automation.
Take one process, not the whole department. A suitable one meets all six conditions at once.
A daily marketing report fits those conditions most often. After that come advertising control, lead sorting and reputation monitoring — one scenario at a time rather than all at once.
The price is made of the process review, building the scenario and the monthly charge for requests to the model.
Starting prices are listed on our AI and automation page. The charge for requests depends on how much text is involved and how many checks run each day: a daily summary across a few campaigns costs less than working through all of the sales team's correspondence.
Three things drive the cost of the work itself: how many sources have to be connected, what state the data in them is in, and which level of autonomy is chosen. A read-only scenario on a single source is quicker to build than an agent with write access to the CRM, which needs roles, a log and a way to undo its actions.
Syntra Systems starts not with the choice of model but with the process: where employees spend their time, which data is already collected, which systems are connected to each other and what can safely be handed to an agent. From that we assemble a diagram: which sources are connected, which actions are permitted and at what level of autonomy the agent runs.
After that comes one scenario, one measurable result and a log that shows every action. An agent wired into advertising, the CRM and mailings all at once usually gets stuck on approving access, so we add scenarios one by one and grant rights by a list of objects. None of this necessarily means a smaller team: more often it takes the reconciliation of figures off a marketer and gives the time back for hypotheses.
Let’s discuss your project
Tell us what you need, and we will estimate the timeline and cost and suggest a solution.
A chatbot holds a conversation along a set script and answers questions. Besides the conversation, an agent reaches into the CRM and other systems, checks figures, creates tasks and carries out permitted actions. That is why an agent is configured not only through wording but through access rights, limits on actions and a log.
Partly. The agent analyses campaigns, finds ads that are not working and prepares recommendations. It can change budgets and launch campaigns only within limits and rules set in advance, while larger changes stay with a person: the cost of a mistake here is measured in money.
An agent connects to your CRM through the interface it already exposes for exchanging data with other software, so there's no need to change the system. It tags enquiries, summarises conversations, suggests the next step and creates tasks. Its access is granted separately from an employee's — only the objects it needs, no bulk exports, with a log of every request.
The systems themselves do not: the agent connects to the CRM, analytics tools and ad accounts already in use. The process, however, is described and simplified before it is automated — otherwise the agent repeats the existing mess, only faster.
This falls under routine automation, from $2,500; if you first need to work out what to automate, an AI audit costs $400. On top comes a monthly charge for requests to the model — a summary across a few campaigns costs less than analysing the whole sales team's correspondence.
No. The agent takes the repetitive work: reconciling figures, sorting enquiries, assembling reports, monitoring mentions. Strategy, negotiation, brand work and budget decisions stay with people, and automation more often gives them time back than cuts headcount.
With one process that repeats, relies on data already being collected and has a result visible in numbers. Most often that is the daily marketing report. Once the scenario has run on the live stream, the next one is added — one at a time.
Changing large budgets and launching new campaigns, mass mailings and posts under the company's name, changing client statuses or contract terms, deleting data and any financial operations. Law No. ZRU-1115 of 21 January 2026 also bars legally significant decisions from resting on an AI's conclusion alone.
Sources
