There is an almost irresistible way to talk about AI and advertising.
Ask an AI to write ten LinkedIn ads.
Give it your ICP.
Ask it for five hooks.
Have it rewrite the landing page.
Generate a few variations.
It all sounds useful. It is also increasingly commoditized.
The more interesting development is happening somewhere else.
What happens when the AI does not just write the ad, but can actually inspect the advertising account, reason over the campaign structure, build an audience, identify wasted spend, prepare changes, and help manage what happens next?
That is a much bigger shift.
It moves the LLM from copy assistant to operating layer.
And Model Context Protocol, or MCP, is one of the technologies making that possible. MCP is an open standard for connecting AI applications to external systems, including data sources, tools, and workflows. In practical terms, it gives an AI model a standardized way to interact with software rather than merely talk about it.
This matters particularly for LinkedIn Ads because the platform's strength is also one of its biggest operational headaches.
LinkedIn gives B2B marketers an unusually rich targeting system. You can work with job functions, seniority, job titles, skills, company size, industry, years of experience, education, interests and other professional attributes. LinkedIn itself says advertisers can use more than 200 characteristics for audience targeting.
The problem is not a lack of options.
It is the amount of work required to turn all those options into a coherent, continuously improving advertising system.
That is where AI agents become genuinely interesting.
LinkedIn Ads Has an Interface Problem
Campaign Manager is not "bad."
It is simply designed around a human-operated workflow.
You log in.
You inspect campaigns.
You click into an ad set.
You look at audiences.
You check spend.
You export something.
You compare numbers.
You make changes.
Then you repeat the process for the next account, campaign, audience, creative or client.
That model works perfectly well when the human is the primary computing layer.
But that is increasingly backwards.
We have spent the last few years putting an LLM on top of our documents, spreadsheets, codebases, customer research and knowledge bases.
Why should the advertising account remain a closed dashboard that the human has to manually operate?
That is the more important question.
The dashboard is becoming an abstraction layer we may not need as often
Think about the difference between these two workflows.
Traditional workflow
Open Campaign Manager → find campaign → inspect ad sets → filter performance → identify problem → navigate to targeting → make change → save → repeat.
Agentic workflow
"Look at my LinkedIn campaigns from the last 14 days. Find job titles with meaningful spend but no leads. Ignore segments with fewer than 500 impressions. Draft exclusions for anything above my $150 CPL threshold and show me the reasoning."
The second workflow does not eliminate the advertising platform.
It changes how you interact with it.
The LLM becomes the interface.
The ad platform becomes the system of record.
MCP becomes the connection between the two.
That distinction is important because it moves the conversation beyond "AI can save time."
It changes the architecture of the workflow.
The First Big Shift: From Prompting to Context
There is a common misconception about AI agents.
People think the magic is writing better prompts.
It is not.
The magic is giving the model access to the right context and the ability to act on that context.
Suppose you ask a normal LLM:
"Which LinkedIn audiences should I target for my SaaS product?"
It can give you a perfectly plausible answer.
It may even be excellent.
But it does not know:
- Which campaigns you already ran
- Which job titles actually produced leads
- Which audiences consumed your budget
- Which campaigns are active
- Your historical CPL
- Your current budget
- Which creative was attached to which audience
- Whether your sales team rejects certain lead types
- What happened to leads after they entered the CRM
You are asking the model to reason about an account it cannot see.
Now give the AI access to that account.
The question changes.
Instead of:
"What should I target?"
You can ask:
"Based on my last 90 days of LinkedIn performance, which targeting dimensions should I test next?"
That is a vastly more useful question.
This is why MCP matters
MCP defines standardized ways for AI applications to access external resources and invoke tools. The protocol explicitly separates things such as contextual resources and executable tools, allowing models to work with external systems rather than existing entirely inside the conversation.
In advertising, that can translate into a simple but powerful loop:
Observe → reason → draft → review → act → observe again
That loop is more important than any individual AI-generated ad.
What an AI Agent Can Actually Do Inside a LinkedIn Ad Workflow
The useful question is not "Can AI run LinkedIn Ads?"
The useful question is:
Which parts of the advertising workflow are deterministic enough to automate, analytical enough for an LLM to assist with, and consequential enough to keep a human in the approval loop?
That produces a much more realistic picture.
1. Turn an ICP into an actual LinkedIn audience
This is harder than it sounds.
"Target marketing leaders at SaaS companies" is a strategic statement.
Campaign Manager needs operational decisions.
Which job titles?
Which seniority tiers?
Which company sizes?
Should function be used instead of titles?
Should skills be included?
What should be excluded?
What audience size does the resulting combination produce?
LinkedIn itself advises marketers to avoid excessive targeting and recommends testing combinations of a few core dimensions rather than endlessly narrowing the audience. One LinkedIn targeting guide cites an audience range of roughly 60,000 to 400,000 as a useful starting point for many successful campaigns.
An agent can take the strategic instruction and translate it into a concrete targeting draft.
For example:
Build an audience for:
Marketing leaders at SaaS companies with 50 to 500 employees in the US.
Include managers and above.
Test function against title-based targeting.
Exclude students and current customers.
The important part is not the natural-language interface.
The important part is that the instruction becomes an executable campaign configuration.
That is a very different thing from asking ChatGPT for audience ideas.
2. Analyse job-title performance at scale
This is where agentic workflows become particularly compelling.
Imagine an account with dozens of campaigns and hundreds of targeting combinations.
A human can inspect the reports.
But the human has to remember what to look for.
An agent can systematically ask:
- Which titles have spent more than our threshold?
- Which have generated clicks but no conversions?
- Which produce leads below the account's median CPL?
- Which audiences have enough data to justify a conclusion?
- Are some titles consistently strong across multiple campaigns?
- Are there outliers we should investigate rather than automatically remove?
A LinkedIn Ads MCP workflow can expose that information directly to the agent.
AdKit's LinkedIn workflow, for example, demonstrates an agent ranking job titles by performance and drafting exclusions for titles that consume spend without producing leads.
That sounds like a small feature.
It is not.
It changes the economics of optimization.
Why?
Because analysis that takes 20 seconds for a model can still take 20 minutes for a human, largely because of navigation, filtering, comparison and context switching.
The bottleneck is not calculation.
It is access.
3. Catch Budget Problems Before They Become Monday-Morning Problems
Here is another deceptively valuable use case.
LinkedIn's budgeting system has nuances that are easy to miss.
For continuous campaigns, actual daily spend can be as much as 50% above the specified daily budget on a given day, while the platform maintains the applicable weekly pacing constraint. LinkedIn gives the example of a $100 daily budget potentially spending $150 on a particular day while staying within the seven-day total.
That is not necessarily a problem. It is part of the platform's pacing model.
But it creates an important distinction:
A budget anomaly is not automatically a budget mistake.
That distinction is exactly why blindly "automating ads" is dangerous.
A useful agent should not simply see:
Spend > daily budget
and pause the campaign.
It should understand:
- Budget type
- Campaign schedule
- Spend over time
- Current CPL
- Delivery status
- Campaign objective
- Historical performance
Then it can say:
"Today's spend is 42% above the nominal daily budget, but the campaign is within LinkedIn's weekly pacing envelope. No action recommended."
That is better automation.
Good automation knows when not to act.
4. Build Reports Without Turning Marketers Into Spreadsheet Operators
Reporting is one of the least strategically valuable parts of paid media and one of the easiest to automate.
Yet it still consumes hours.
The table is easy. But, the interpretation is where the value is.
A useful agent can move from:
"Spend was up 8%."
to:
"Spend increased 8%, while leads increased 16%, producing a 6.6% decrease in CPL. The improvement came primarily from the VP Sales audience, while Account Executive targeting generated clicks but no leads."
That is closer to what a marketer actually needs.
AdKit explicitly positions its LinkedIn Ads MCP around asking an agent for a reporting period and receiving a ready-to-use summary of spend, leads, CPL and changes.
The important change is not report generation. It is report interpretation.
The Second Big Shift: From "Create Campaigns" to "Maintain Campaign State"
This is the part I think is being missed in most discussions about AI advertising.
Launching a campaign is a one-time event.
Managing a campaign is an ongoing state-management problem.
The campaign changes every day.
Spend changes.
Performance changes.
Audience response changes.
Creative fatigue develops.
One segment starts outperforming another.
A campaign runs out of budget.
An ad stops delivering.
A new business constraint emerges.
The agentic opportunity is therefore less:
"AI, launch me a campaign."
and more:
"AI, continuously inspect the state of my advertising system and surface changes that deserve my attention."
That is a much more powerful model.
The Human Still Matters More Than the Agent
This is where the conversation needs some skepticism.
Giving an LLM access to an advertising account does not suddenly make it a great media buyer.
It gives it access.
Expertise is something else.
An agent can detect that:
CPL increased 32%.
It cannot automatically know whether that is bad.
Maybe the campaign is entering a new market.
Maybe a new high-value audience is being tested.
Maybe lead volume doubled despite the higher CPL.
Maybe sales says those leads are worth three times as much.
Maybe the campaign is part of a brand-building strategy.
The metric is not the strategy.
This is why the strongest implementations should be draft-first.
AdKit explicitly uses that model for its LinkedIn Ads workflow: campaigns, audiences and budget changes are held as drafts until the user approves them. It also routes activity through LinkedIn's official Marketing API with validation, rate limiting and error handling rather than having the agent communicate directly with LinkedIn.
That is exactly the sort of boundary I would want around an agent that can touch paid media.
Why "Draft First" Is More Important Than "Fully Autonomous"
There is a temptation to frame human approval as a temporary limitation.
I think that is the wrong framing.
For paid media, approval is a feature.
Consider the difference between:
"Pause every ad over $120 CPL."
and:
"Find ads above $120 CPL and prepare the recommended pauses, but do not apply them."
The second lets the agent perform the expensive cognitive work while leaving final accountability with the marketer.
That is a much healthier automation pattern.
A sensible hierarchy looks like this
Level 1: Read
The agent can inspect performance.
Level 2: Explain
The agent identifies patterns and anomalies.
Level 3: Recommend
The agent proposes changes.
Level 4: Draft
The agent prepares the exact changes.
Level 5: Approve
The marketer reviews the changes.
Level 6: Execute
The system applies approved changes.
Level 7: Monitor
The agent checks the resulting state and reports back.
That is an actual operating loop.
It is far more useful than saying "AI can manage your ads."
Where MCP Changes the No-Code Equation
This is particularly relevant to the no-code community.
The traditional mental model for connecting software systems is:
API → developer → integration → application
MCP introduces another possibility:
AI client → MCP server → tool → external system
The technical complexity has not disappeared.
It has moved behind an interface that an LLM can use.
That opens up interesting workflows for people who are comfortable with no-code, automation and AI but do not want to build custom ad-platform integrations from scratch.
For example, a growth marketer could conceptually combine:
Claude
→ reasoning and analysis
MCP
→ standardized tool access
LinkedIn Ads
→ campaign execution and reporting
CRM
→ lead quality
Analytics
→ website conversion data
Now the agent can begin reasoning about a more complete growth loop.
That is considerably more interesting than generating 20 LinkedIn headlines.
For teams building their own AI workflows, connecting Claude and custom LLMs to B2B ad accounts is the kind of infrastructure layer that makes this possible without forcing the marketer to build a LinkedIn advertising integration from scratch.
The Next Step Is Not More Automation. It Is Better Feedback Loops.
Here is where I think the real opportunity lies.
Most advertising workflows are fragmented.
The marketer launches ads.
The ad platform reports clicks and conversions.
The CRM holds lead quality.
Sales holds revenue information.
The landing page holds conversion data.
The marketer manually connects the dots.
An AI agent could eventually act as the connective tissue.
Imagine giving it access to:
- LinkedIn campaign performance
- CRM lead quality
- Website conversions
- Sales pipeline
- Product analytics
- Customer segments
- Competitor research
Then asking:
"Which audience should we increase spending on next month?"
That is not an advertising question anymore.
It is a growth intelligence question.
And that is where AI agents become much more valuable.
But There Are Three Things an AI Agent Still Cannot Fix
1. Bad Measurement
If your CRM is not connected properly, your conversion events are unreliable, or sales is not classifying lead quality consistently, the agent is reasoning over garbage.
Better AI does not create better data.
Fix instrumentation first.
2. Bad Strategy
If your offer is weak, your positioning is generic and your ICP is wrong, letting an agent create 100 campaign variations simply creates more ways to be wrong.
The agent can optimize the machine.
It cannot invent product-market fit.
3. Insufficient Data
This is particularly important with LinkedIn.
A model can confidently identify a pattern that is not actually statistically meaningful.
Suppose one job title has:
- 11 clicks
- 1 lead
- $70 CPL
and another has:
- 900 clicks
- 14 leads
- $84 CPL
It would be irresponsible to declare the first title "better" simply because its observed CPL looks lower.
An agent needs rules around:
- Minimum spend
- Minimum impressions
- Minimum clicks
- Minimum conversion volume
- Observation windows
This is another reason expert-designed automation beats generic autonomous optimization.
You do not want an AI that confidently acts on noise.
The Best AI Advertising Workflow Will Probably Feel Surprisingly Boring
This may sound counterintuitive.
People imagine the future of AI advertising as an autonomous super-agent constantly making clever decisions.
I suspect the actual winning workflow will be much less dramatic.
It will look like:
"Here are three changes I recommend."
"Here is why."
"Here is the expected impact."
"Review and approve."
That is it.
No theatrical autonomy.
No "10x AI growth engine."
Just a machine that removes 70% of the tedious work while making the remaining 30% more thoughtful.
That is a much more valuable product.
The Bigger Picture: Ads Are Becoming Software
This is the part worth watching.
Advertising used to be predominantly an interface problem.
Learn the platform.
Learn the buttons.
Learn the settings.
Learn the workflows.
AI agents are pushing advertising toward becoming a software orchestration problem.
The question increasingly becomes:
What systems should be connected?
What context should the model have?
Which actions should it be allowed to perform?
Which actions require approval?
What rules should constrain it?
How do we measure whether its decisions actually improved the business?
That is a very different skill set from traditional media buying.
And it fits neatly into the broader no-code philosophy.
Do not build everything yourself. Compose systems.
MCP makes that composition layer more accessible to AI systems.
LinkedIn provides the advertising data and execution environment.
The LLM supplies reasoning.
The marketer supplies strategy, judgment and accountability.
That combination is much more interesting than another AI copywriting tool.
