learn

/

methodology

What is agentic marketing? The three levels of AI marketing agents

Agentic marketing means AI agents that plan and act toward a marketing objective. See the three levels (answers, tasks, goal) and which one your team needs.
Agentic marketing means AI agents that plan and act toward a marketing objective. See the three levels (answers, tasks, goal) and which one your team needs.

By Gabriele Franco, co-founder of Cassandra

Get one practical idea every week to grow with the same team

Agentic marketing is marketing run with AI agents: software that receives an objective, plans its own steps, uses your data and tools to carry them out, and checks the result. Agents come in three levels. Level 1 answers questions, level 2 executes tasks, and level 3 owns a business goal from the plan to the daily correction.

The industry grades marketing agents by autonomy: how much can the agent do without asking you? I think that grade misses what matters. An agent can pause ads, rewrite copy and move budget all night without asking anyone, and you can still miss the quarter, because nobody was checking whether all that activity added up to the number.

So this guide grades agents on another question: what does the agent answer for? An answer, a task, or the goal. Know the three levels and you can place any tool in a few minutes.

What does agentic marketing mean?

"Agentic" comes from agency: the capacity to act toward an objective. A normal AI chat waits for a question. An agent gets an objective, decides the steps, calls tools (your ad accounts, analytics, shop data), looks at what came back and decides what to do next.

Anthropic's engineering guide on building effective agents draws the line this way. In a workflow, the steps are fixed in advance in code. In an agent, the model directs its own process and chooses which tools to use.

Forrester, in an October 2026 post, separates two terms people mix up. Agentic marketing is the concept, and a marketing AI agent is the technology underneath. Forrester lists five criteria for a marketing AI agent, and two matter here: it is goal-oriented, and it executes multiple tasks in an adaptive workflow.

So agentic marketing starts when the software stops waiting for your next instruction and works toward an objective you gave it. How big that objective is decides the level. For the technology itself, read what an AI marketing agent actually does.

How is agentic marketing different from marketing automation?

Marketing automation follows rules a person wrote. If a cart is abandoned, send email B after four hours. It does exactly that every time, including on the day the rule stops making sense. An agent is given the objective behind the rule (recover abandoned carts at a profit) and chooses the steps itself.



Marketing automation

AI assistant (chat)

Agentic marketing

Who decides the next step

The person who wrote the rule

You, with each prompt

The agent, inside limits you set

What it needs from you

Rules

Questions

An objective and guardrails

When conditions change

It keeps running the old rule

Nothing happens until you ask

It adjusts the steps

What are the three levels of AI agents in marketing?

I call this framework the Ownership Ladder, because each level is defined by what the agent owns.


Level


What it does

What you still do yourself

How to test it

Level 1: answers

Reads your data and tells you what it sees: what changed, which campaign is slipping

Decide what to do, do it, check whether it worked

Ask it to change something. If it can only describe the change, it is level 1

Level 2: executes tasks

Carries out one job at a time: builds ads, moves a budget, pauses a campaign, sends the weekly report

Choose the tasks and their order, and check whether they add up to the goal

Give it no instructions for a week. If tasks keep running and nothing tells you whether you are on track, it is level 2

Level 3: owns the goal

Takes a goal with a number and a date, builds the plan, executes it, checks progress against the forecast every day, proposes the correction and remembers what worked

Set the goal and the limits, approve what you chose to approve, bring strategy and creative judgment

Ask "will we hit the number, and what are you doing about it?" If it answers with a forecast against the plan and one specific move, it is level 3

Level 1 owns the answer. Level 2 owns the task. Level 3 owns the goal.

Each level contains the one below. A level 3 system still answers questions and executes tasks, and it ties every one of them to one number.

Many frameworks rank agents by autonomy instead. The glossary on CDP.com, for example, describes three levels of marketing autonomy, from AI that recommends while people approve to AI that adjusts goals and budgets with minimal oversight. Autonomy only tells you how often the agent asks permission, and you can set it at any level. A level 3 system can wait for your yes on every change and still be level 3, because it is the one tracking the goal. More on that in the guide to autonomous AI marketing agents.

How do you tell which level a tool is at?

Use the silence test: stop talking to the tool for a week and watch what it does.

A level 1 tool does nothing. It was waiting for your question.

A level 2 tool keeps running whatever you scheduled. Nothing tells you whether the week moved you closer to the number.

A level 3 system comes to you. It says where you stand against the plan and what it wants to change.

Here is the same morning at each level. Hearthline is a candle brand with a goal of €900,000 in revenue between 1 October and 31 December (the brand and every figure in this example are illustrative). On 12 November the plan says cumulative revenue should be €380,000. The real figure is €352,000.

  • Level 1, when you ask: "Cost per order on Meta rose 18% week over week. Two prospecting campaigns explain most of the increase."

  • Level 2, when you tell it to: it moves €2,000 a week from the two weak campaigns to the strongest one and confirms the change.

  • Level 3, before you ask: "You are €28,000 behind plan, about 7%. At this pace you finish near €834,000. Proposed move: shift €2,000 a week to the campaign that is still on forecast and start the Black Friday budget four days earlier. If it works, cumulative revenue is back on the plan line by 2 December. Approve?"

All three read the same ad account. Only the third started from the €900,000.

What does it mean for an agent to own the goal?

Think about driving somewhere you have never been.

A speedometer tells you how fast you are going, and it will happily read 110 while you drive in the wrong direction. A dashboard is a speedometer, and a level 1 agent is a passenger reading it out loud.

Cruise control and parking assist each do one job well when you switch them on. That is level 2.

Neither knows where you are going. For that you need a navigator, and each of its five parts maps to something in marketing:

  1. The destination is the goal, with a number and a date: €900,000 by 31 December.

  2. The route is the media plan: when and where to invest, week by week and channel by channel. See how to build a media plan.

  3. The arrival time is the cumulative sales forecast: where revenue should be on each day if the plan holds.

  4. The recalculation is the daily check of predicted against actual, and the move that puts you back on the route when you drift. This is revenue pacing.

  5. The memory of the roads is the record of what worked for this brand: every plan, test and budget change, used again the next time a route is drawn.

And it drives when you say go. You pick the destination, and you decide which changes it makes on its own and which wait for your approval.

Cassandra, the product my team builds, is an example of level 3. You give it a goal and it builds the media plan from marketing mix modeling and the brand's own history, with a cumulative sales forecast. You can tell it to implement the plan, and you choose what runs on its own and what waits for you. Every day it compares predicted with actual and, when the plan is off track, proposes the move that gets you back. It keeps a record of every plan, experiment and budget change, a memory we call the marketing brain.

Is everything sold as agentic marketing really agentic?

No, and the label is the least reliable thing on the page. Gartner has a name for the relabelling: "agent washing", meaning assistants and chatbots rebranded as agents. In June 2025 Reuters reported Gartner's estimate that only about 130 of the thousands of vendors selling agentic AI were real. That count covers all software, far beyond marketing, and it is over a year old. The lesson holds: test the product, ignore the label.

My read of the category: most products sold as agentic today sit at level 1 or level 2. That is fine. A level 1 agent that finds wasted spend while you are still opening Ads Manager saves real hours. So does a level 2 agent that builds a batch of ad variations before lunch. Many teams need exactly that and nothing more.

Trouble starts when a level 2 tool is bought to solve a level 3 problem. The tasks get done faster and the quarter is still missed, because nobody was holding the number.

What is an agentic marketing platform?

An agentic marketing platform is the place where agents, your data and your tools live together, so an agent can read ad accounts, analytics and sales, act on them, and stay inside one set of permissions. The word covers every level, so apply the same ladder. How to choose one is in the guide to agentic marketing platforms.

Which level of agentic marketing do you need?

Here's the rule: buy the level of your bottleneck.


If your bottleneck is


You need

Why

Insight: you don't know what happened or why

Level 1

Faster answers fix it

Hands: you know what to do and lack the hours

Level 2

Execution is the gap

Direction: the work gets done and you still find out late that you are behind

Level 3

Something has to hold the number every day

Level 3 asks three things of you first: a goal written as a number and a date ("grow revenue" gives a system nothing to pace against), enough sales and spend history to forecast from, and a decision on what it may change alone.

If you can't write the goal that way yet, stay at level 1 or 2. A navigator with no destination is only a map. And if you spend a small budget on a single channel, a level 1 agent plus your own judgment is usually enough.

Common traps to avoid

Grading by autonomy

"No humans needed" is a pitch. Ask what the agent owns first, then set the autonomy.

The task treadmill

More tasks, done faster, with no one adding them up. If your team ships twice as many ads and still learns in week ten that the quarter is short, you needed direction and bought speed.

All or nothing on approvals

Automatic everything on day one is reckless, and approving every small change forever wastes the system. Start with approval on anything that spends money and loosen it as trust builds.

No memory

An agent that starts from zero every Monday will propose the test you already ran in March. Ask where the history of plans, tests and budget changes is kept.

Every decision should be judged against the goal

A budget move, a new creative or a paused campaign is good or bad only in relation to the goal. To judge it you need the whole brand context in one place: ads, analytics, sales, products, reviews, and what past tests proved. And something should check every day whether you are still on track. That is what we built Cassandra for. You can see how it works on the platform page, or start for free with your own goal.

Key takeaways

  • Agentic marketing means AI agents that take an objective, plan the steps, act and check the result.

  • The Ownership Ladder: level 1 owns the answer, level 2 owns the task, level 3 owns the goal.

  • The silence test: stop talking to the tool for a week. Level 3 comes back with where you stand against the plan and a move to fix it.

  • Autonomy is a dial you set at any level. It says nothing about what the agent is responsible for.

  • Most products sold as agentic are level 1 or 2, and for many teams that is the right purchase.

  • Buy the level of your bottleneck: insight, hands or direction.

Next

Sources checked on 14 November 2026.

Agentic marketing is marketing run with AI agents: software that receives an objective, plans its own steps, uses your data and tools to carry them out, and checks the result. Agents come in three levels. Level 1 answers questions, level 2 executes tasks, and level 3 owns a business goal from the plan to the daily correction.

The industry grades marketing agents by autonomy: how much can the agent do without asking you? I think that grade misses what matters. An agent can pause ads, rewrite copy and move budget all night without asking anyone, and you can still miss the quarter, because nobody was checking whether all that activity added up to the number.

So this guide grades agents on another question: what does the agent answer for? An answer, a task, or the goal. Know the three levels and you can place any tool in a few minutes.

What does agentic marketing mean?

"Agentic" comes from agency: the capacity to act toward an objective. A normal AI chat waits for a question. An agent gets an objective, decides the steps, calls tools (your ad accounts, analytics, shop data), looks at what came back and decides what to do next.

Anthropic's engineering guide on building effective agents draws the line this way. In a workflow, the steps are fixed in advance in code. In an agent, the model directs its own process and chooses which tools to use.

Forrester, in an October 2026 post, separates two terms people mix up. Agentic marketing is the concept, and a marketing AI agent is the technology underneath. Forrester lists five criteria for a marketing AI agent, and two matter here: it is goal-oriented, and it executes multiple tasks in an adaptive workflow.

So agentic marketing starts when the software stops waiting for your next instruction and works toward an objective you gave it. How big that objective is decides the level. For the technology itself, read what an AI marketing agent actually does.

How is agentic marketing different from marketing automation?

Marketing automation follows rules a person wrote. If a cart is abandoned, send email B after four hours. It does exactly that every time, including on the day the rule stops making sense. An agent is given the objective behind the rule (recover abandoned carts at a profit) and chooses the steps itself.



Marketing automation

AI assistant (chat)

Agentic marketing

Who decides the next step

The person who wrote the rule

You, with each prompt

The agent, inside limits you set

What it needs from you

Rules

Questions

An objective and guardrails

When conditions change

It keeps running the old rule

Nothing happens until you ask

It adjusts the steps

What are the three levels of AI agents in marketing?

I call this framework the Ownership Ladder, because each level is defined by what the agent owns.


Level


What it does

What you still do yourself

How to test it

Level 1: answers

Reads your data and tells you what it sees: what changed, which campaign is slipping

Decide what to do, do it, check whether it worked

Ask it to change something. If it can only describe the change, it is level 1

Level 2: executes tasks

Carries out one job at a time: builds ads, moves a budget, pauses a campaign, sends the weekly report

Choose the tasks and their order, and check whether they add up to the goal

Give it no instructions for a week. If tasks keep running and nothing tells you whether you are on track, it is level 2

Level 3: owns the goal

Takes a goal with a number and a date, builds the plan, executes it, checks progress against the forecast every day, proposes the correction and remembers what worked

Set the goal and the limits, approve what you chose to approve, bring strategy and creative judgment

Ask "will we hit the number, and what are you doing about it?" If it answers with a forecast against the plan and one specific move, it is level 3

Level 1 owns the answer. Level 2 owns the task. Level 3 owns the goal.

Each level contains the one below. A level 3 system still answers questions and executes tasks, and it ties every one of them to one number.

Many frameworks rank agents by autonomy instead. The glossary on CDP.com, for example, describes three levels of marketing autonomy, from AI that recommends while people approve to AI that adjusts goals and budgets with minimal oversight. Autonomy only tells you how often the agent asks permission, and you can set it at any level. A level 3 system can wait for your yes on every change and still be level 3, because it is the one tracking the goal. More on that in the guide to autonomous AI marketing agents.

How do you tell which level a tool is at?

Use the silence test: stop talking to the tool for a week and watch what it does.

A level 1 tool does nothing. It was waiting for your question.

A level 2 tool keeps running whatever you scheduled. Nothing tells you whether the week moved you closer to the number.

A level 3 system comes to you. It says where you stand against the plan and what it wants to change.

Here is the same morning at each level. Hearthline is a candle brand with a goal of €900,000 in revenue between 1 October and 31 December (the brand and every figure in this example are illustrative). On 12 November the plan says cumulative revenue should be €380,000. The real figure is €352,000.

  • Level 1, when you ask: "Cost per order on Meta rose 18% week over week. Two prospecting campaigns explain most of the increase."

  • Level 2, when you tell it to: it moves €2,000 a week from the two weak campaigns to the strongest one and confirms the change.

  • Level 3, before you ask: "You are €28,000 behind plan, about 7%. At this pace you finish near €834,000. Proposed move: shift €2,000 a week to the campaign that is still on forecast and start the Black Friday budget four days earlier. If it works, cumulative revenue is back on the plan line by 2 December. Approve?"

All three read the same ad account. Only the third started from the €900,000.

What does it mean for an agent to own the goal?

Think about driving somewhere you have never been.

A speedometer tells you how fast you are going, and it will happily read 110 while you drive in the wrong direction. A dashboard is a speedometer, and a level 1 agent is a passenger reading it out loud.

Cruise control and parking assist each do one job well when you switch them on. That is level 2.

Neither knows where you are going. For that you need a navigator, and each of its five parts maps to something in marketing:

  1. The destination is the goal, with a number and a date: €900,000 by 31 December.

  2. The route is the media plan: when and where to invest, week by week and channel by channel. See how to build a media plan.

  3. The arrival time is the cumulative sales forecast: where revenue should be on each day if the plan holds.

  4. The recalculation is the daily check of predicted against actual, and the move that puts you back on the route when you drift. This is revenue pacing.

  5. The memory of the roads is the record of what worked for this brand: every plan, test and budget change, used again the next time a route is drawn.

And it drives when you say go. You pick the destination, and you decide which changes it makes on its own and which wait for your approval.

Cassandra, the product my team builds, is an example of level 3. You give it a goal and it builds the media plan from marketing mix modeling and the brand's own history, with a cumulative sales forecast. You can tell it to implement the plan, and you choose what runs on its own and what waits for you. Every day it compares predicted with actual and, when the plan is off track, proposes the move that gets you back. It keeps a record of every plan, experiment and budget change, a memory we call the marketing brain.

Is everything sold as agentic marketing really agentic?

No, and the label is the least reliable thing on the page. Gartner has a name for the relabelling: "agent washing", meaning assistants and chatbots rebranded as agents. In June 2025 Reuters reported Gartner's estimate that only about 130 of the thousands of vendors selling agentic AI were real. That count covers all software, far beyond marketing, and it is over a year old. The lesson holds: test the product, ignore the label.

My read of the category: most products sold as agentic today sit at level 1 or level 2. That is fine. A level 1 agent that finds wasted spend while you are still opening Ads Manager saves real hours. So does a level 2 agent that builds a batch of ad variations before lunch. Many teams need exactly that and nothing more.

Trouble starts when a level 2 tool is bought to solve a level 3 problem. The tasks get done faster and the quarter is still missed, because nobody was holding the number.

What is an agentic marketing platform?

An agentic marketing platform is the place where agents, your data and your tools live together, so an agent can read ad accounts, analytics and sales, act on them, and stay inside one set of permissions. The word covers every level, so apply the same ladder. How to choose one is in the guide to agentic marketing platforms.

Which level of agentic marketing do you need?

Here's the rule: buy the level of your bottleneck.


If your bottleneck is


You need

Why

Insight: you don't know what happened or why

Level 1

Faster answers fix it

Hands: you know what to do and lack the hours

Level 2

Execution is the gap

Direction: the work gets done and you still find out late that you are behind

Level 3

Something has to hold the number every day

Level 3 asks three things of you first: a goal written as a number and a date ("grow revenue" gives a system nothing to pace against), enough sales and spend history to forecast from, and a decision on what it may change alone.

If you can't write the goal that way yet, stay at level 1 or 2. A navigator with no destination is only a map. And if you spend a small budget on a single channel, a level 1 agent plus your own judgment is usually enough.

Common traps to avoid

Grading by autonomy

"No humans needed" is a pitch. Ask what the agent owns first, then set the autonomy.

The task treadmill

More tasks, done faster, with no one adding them up. If your team ships twice as many ads and still learns in week ten that the quarter is short, you needed direction and bought speed.

All or nothing on approvals

Automatic everything on day one is reckless, and approving every small change forever wastes the system. Start with approval on anything that spends money and loosen it as trust builds.

No memory

An agent that starts from zero every Monday will propose the test you already ran in March. Ask where the history of plans, tests and budget changes is kept.

Every decision should be judged against the goal

A budget move, a new creative or a paused campaign is good or bad only in relation to the goal. To judge it you need the whole brand context in one place: ads, analytics, sales, products, reviews, and what past tests proved. And something should check every day whether you are still on track. That is what we built Cassandra for. You can see how it works on the platform page, or start for free with your own goal.

Key takeaways

  • Agentic marketing means AI agents that take an objective, plan the steps, act and check the result.

  • The Ownership Ladder: level 1 owns the answer, level 2 owns the task, level 3 owns the goal.

  • The silence test: stop talking to the tool for a week. Level 3 comes back with where you stand against the plan and a move to fix it.

  • Autonomy is a dial you set at any level. It says nothing about what the agent is responsible for.

  • Most products sold as agentic are level 1 or 2, and for many teams that is the right purchase.

  • Buy the level of your bottleneck: insight, hands or direction.

Next

Sources checked on 14 November 2026.