AI agents

AI agents for marketing: what they are, what they can do and what to have built

An AI agent gets a goal, chooses the steps itself, uses tools such as your ad account or your feed, and checks the result. Five agents do my routine work. Anything irreversible, I approve myself.

My own tools do the routine work. The decisions stay with me.
Fabiano conducting his crew of AI agents: Pitch, Tally and Hoot

What is an AI agent?

An AI agent is software with a language model as its brain that works through a goal by itself. It devises the steps, carries them out with tools such as your ad account, and checks whether the result is right.

Systems that work this way are together called agentic AI. The difference from other AI is who decides the steps: in a fixed workflow they are set in advance, in an agent the model chooses them itself (Anthropic).

  1. 1GoalWhat needs to get done?
  2. 2PlanWhich steps are needed?
  3. 3DoWith tools, within permissions
  4. 4CheckIs it right? If not: again

The same problem, three ways

Your ad costs rose 30% this week. This is how each type of AI handles it:

Honest advice: if the task is the same every time, a fixed automation is cheaper and more predictable. Choose an agent when something different has to be investigated each time.

How an agent works

An agent has five parts. If one is missing, you get a clever chatbot or a risky system.

A model

The language model that reads, reasons and writes, such as Claude, GPT or Gemini.

Instructions

What the goal is, which rules apply and when to stop. In plain language.

Tools

Connections to your systems: Google Ads, Merchant Center, GA4, your email. Only what is needed.

Memory

What it did before and what you corrected, so it does not start from scratch every time.

Guardrails

Checks on the output and a human who approves anything irreversible.

HootReporting agent · example
Monday 7:00 AM
  1. GoalWeekly report for Example Store
  2. Fetch dataGoogle Ads, GA4 and Merchant Center, read only
  3. CheckEvery number in the text matches the table
  4. NotableShopping costs +38% at the same revenue
  5. ProposalLower the budget of two product groups
  6. Waiting for FabianoChange goes live only after approval

Made-up example, this is what a run looks like.

My crew: five agents and one human who decides

You see them all over this site: Leo, Stack, Pitch, Tally and Hoot. They are my own agents, each with its own task. Together they form the same chain as my work:

  1. Leo creates
  2. Stack builds
  3. Pitch goes live
  4. Tally measures
  5. Hoot reports
  6. Fabiano approves
Leo, creative
Creates

Leo Creative

Creative assistant

Makes ad variants within your brand rules and catches copy that does not match the real price.

Allowed: make drafts. Going live only after approval.

Stack, building
Builds

Stack Building

Feed builder

Builds and maintains your product feed: titles, attributes and categories. Checks prices, stock and disapprovals every day.

Allowed: build and test the feed. Changes go live after approval.

Pitch, ads
Goes live

Pitch Ads

Search term assistant

Goes through search terms and prepares exclusions and new keywords in your account.

Allowed: prepare changes. Going live only after approval.

Tally, measurement
Measures

Tally Measurement

Measurement agent

Checks that your measurement is correct and puts the same customer questions to AI systems every month: are you mentioned?

Allowed: read only.

Hoot, reporting
Reports

Hoot Reporting

Reporting agent

Writes the weekly report, with a source for every number, and keeps a log of who changed what.

Allowed: read only.

Approves

Fabiano The human

Accountable

Decides what happens and approves everything that goes live or costs money. You talk to me, not to a robot.

Allowed: everything, and signs for it.

Each agent gets as few permissions as possible: Tally and Hoot only read, Leo, Stack and Pitch create, build and prepare. Going live happens only after my approval. Pitch always wants to go live right away, Tally wants to measure first. That tension is exactly why a human sits in between.

MCPs: the connections I build

An agent needs a plug to reach your systems. That plug is called an MCP (Model Context Protocol): an open standard that lets an AI safely reach something like Google Ads. The agent does the work, the MCP provides access. My MCPs have the rules from more than 12 years of online marketing built in.

Where the market stands

Many companies try agents, few use them at scale. The platforms you advertise on are opening up to agents in the meantime.

62%of organizations experiment with AI agents; 23% already use them at scale in at least one departmentMcKinsey, Nov 2025
40%+of agent projects are predicted to be canceled by the end of 2027, because of cost, unclear value or too little controlGartner, June 2025
20.0%of EU enterprises with 10 or more employees used AI in 2025, up from 13.5% in 2024Eurostat, 2025
10,000+connections (MCP servers) that let agents reach software; since December 2025 the standard sits under the Linux FoundationLinux Foundation

What the ad platforms offer now

Oct 2025

Google Ads connection for agents

Google released an official MCP server. Agents can read numbers and, more recently, also create assets and labels, always as a proposal first. Changing campaigns, budgets and bids is not possible.

Google
Nov 2025

Ads Advisor in Google Ads

An agent inside Google Ads itself that analyzes campaigns and suggests changes. That is Google's own agent, with Google's goals.

Google
Apr 2026

Meta connection for agents

Meta opened its own connection for AI assistants, which can also create and change campaigns. Still in beta. I use it myself, with changes set to paused first.

My own use

What that means for you: the question is no longer whether an agent can reach your account, but who decides what it may do. A platform agent follows the platform's goal. My agents follow your margin.

When an agent pays off, and when it does not

Pays off

  • A task comes back every week and takes hours
  • The data is already in systems, such as Google Ads, your feed or GA4
  • You can check whether the result is right
  • A mistake can be caught before it does damage

Does not pay off

  • A job you do once
  • There is no good data, or it is wrong
  • Every mistake costs money or reputation right away, with no check in between
  • A fixed automation does it too, and cheaper

Gartner names the reason many projects fail: high costs, unclear value and too little risk management. That is why I start with a scan and a trial run, not with building.

Risks, and how I cover them

Hidden instructions

An agent reads web pages, emails and files. Those can contain text that makes it do something else. This tops the OWASP risk list for AI applications (OWASP).

How I cover it: agents that read from outside get no permission to change or send anything.

Too many permissions

An agent that may do everything can also get everything wrong. A wrong bid or a deleted campaign happens fast.

How I cover it: read only by default. Anything irreversible, I approve myself. OpenAI recommends the same in its guide to agents (OpenAI).

Made-up numbers

A language model writes convincingly, even when a number is wrong.

How I cover it: every number in a report is checked against the source. A failed run changes nothing.

Rules and privacy

Since August 2, 2026, the AI Act's transparency rules apply: people must know they are talking to AI, and AI content must be recognizable (European Commission).

How I cover it: for each agent it is fixed which data it sees, where the data stays and what gets labeled. Personal data stays out when it does not have to be in.

Six blueprints to start with

These are the agents I use myself. I build them on your systems and rules, with the same guardrails.

€2,950AI automation scan

Two conversations, five to ten opportunities ranked by return and risk, and a build plan with a price. Credited against the bill if you start a project within 60 days.

View the scan

€2,000per Google Ads MCP, all four €5,950

Review, build, feed check or ad copy, with my way of working built in. See the Google Ads MCPs.

View the MCPs

€5,000 – €15,000custom build project, fixed price

For agents on your own systems. Depends on scope and number of connections. The price is fixed once the plan is ready. Maintenance afterward €500 to €1,250 per month.

Book a call

Prices excl. VAT.

How a build project goes

1

Scan

Which task pays off, which data exists and which permissions are needed. You get a plan with a price.

Weeks 1 to 2
2

Trial run

The agent runs alongside regular work, without changing anything. That shows what it would do.

2 to 4 weeks
3

Live with approval

Only when the trial is right does it go live. Changes stay proposals until I approve them.

Afterward
4

Maintenance

Platforms keep changing. I maintain the connections and adjust when something changes.

Every month

Frequently asked questions about AI agents

What is an AI agent?

An AI agent is software with a language model as its brain. It gets a goal, decides for itself which steps are needed, uses tools such as your ad account or product feed, and checks the result. A chatbot only answers; an agent also carries out tasks.

What is agentic AI?

Agentic AI is the umbrella term for AI that works through a goal on its own: planning, using tools and checking the result, with limited human supervision. An AI agent is such a system. The opposite is AI that only answers a question.

What is the difference between an AI agent and a chatbot?

A chatbot answers questions in a conversation and then waits for the next question. An agent works through a goal: it fetches data, carries out steps and checks the outcome, within the permissions you give it.

What is the difference from regular automation?

With automation, the steps are fixed in advance, like the same report every Monday. An agent chooses the steps itself, for example to find out why your costs rose. If the task is the same every time, automation is usually better and cheaper.

What does it cost to have an AI agent built?

A ready-made Google Ads MCP costs €2,000 each, all four together €5,950. A custom build project costs €5,000 to €15,000 at a fixed price, after an AI automation scan of €2,950. The scan price is credited against the bill if you start a project within 60 days. Maintenance afterward costs €500 to €1,250 per month.

How long does a build take?

Count on one to two weeks for the scan, two to four weeks of trial run alongside regular work, and then live with your approval. A simple reporting agent is ready sooner than an agent with many connections.

Can an agent put things live by itself?

Not with me, not without approval. By default an agent only reads or makes a proposal. Anything that cannot be undone, such as raising a budget or sending an email, I approve myself.

Is my data safe?

Every agent gets only the permissions it needs, usually read only. Up front we agree which data it sees and where the data stays. Personal data stays out when it does not have to be in.

What does the AI Act mean for agents?

Since August 2, 2026, the transparency rules apply: people must know they are dealing with AI, and AI content must be recognizable. I build that into the setup from the start.

Can't I just use the agent from Google or Meta?

You can, and for small accounts that is sometimes enough. Keep in mind that those agents follow the platform's goal. My agents steer on your margin and leave the decisions to a human.

Want to know where it pays off?

Start with the AI automation scan.

Two conversations and a build plan with a price.

AI automation scan, €2,950 Book a call

Start scanWhatsAppMail