How We Built the First AI Agent That Generates Growth Ticket

Most AI marketing tools are chatbots with analytics access.

You ask a question. AI answers with data.

Maybe generates a chart.

That's not an agent.

That's a smart search interface.

An agent is different.

An agent:

Monitors continuously
Identifies opportunities autonomously
Generates specific action items
Prioritizes by impact
Explains reasoning

No human prompt required.

That's what we built at Cogny.

And it's harder than it looks.

Let me show you why.

---

The Problem with Marketing AI Today

Most "AI marketing tools" work like this:

Connect data sources
Build dashboards
Add ChatGPT integration
User asks: "Why did ROAS drop?"
AI reads data and answers

Useful? Sure.

Revolutionary? No.

Why not:

You still need to know what questions to ask.

Real marketers don't have time for 20 questions.

They want: "Here's what's wrong." "Here's what to do." "Here's the expected impact."

Proactive. Not reactive.

That requires a different architecture.

---

What Makes an Agent Different

Traditional AI assistant:

Waits for questions
Responds when prompted
No memory between sessions
No initiative

AI agent:

Monitors continuously
Identifies issues autonomously
Remembers context
Takes initiative

The technical challenge:

How do you build AI that knows what to look for?

You can't just:

Point AI at data
Say "find problems"
Hope it works

Doesn't work.

AI needs:

Clear objectives
Evaluation criteria
Pattern templates
Domain knowledge
Decision frameworks

That's the hard part.

---

Our Architecture (Simplified)

Layer 1: Data Ingestion

Connect to marketing platforms:

Google Ads API
Meta Marketing API
GA4 via BigQuery
Custom data sources

Challenge: Each platform structures data differently.

Google Ads: campaigns → ad groups → ads → keywords Meta: campaigns → ad sets → ads GA4: Events → users → sessions

Solution: Unified data model that normalizes across platforms.

Layer 2: Continuous Analysis Engine

This is where most tools stop.

They show you the data.

We go further:

Run 40+ analysis algorithms continuously:

Conversion rate variance detection
Budget allocation efficiency
Audience overlap analysis
Creative fatigue patterns
Geographic performance
Time-based optimization
Keyword effectiveness
Placement efficiency
Device performance
And 30+ more

Each algorithm:

Scans all data every 6 hours
Compares to baselines
Identifies anomalies
Calculates statistical significance
Estimates impact

Layer 3: Opportunity Scoring

Found 200 potential issues.

Which matter?

Scoring algorithm considers:

Monetary impact (€ saved or gained)
Statistical confidence (is this real?)
Implementation ease (how hard to fix?)
Time sensitivity (how urgent?)
Historical success rate (did similar tickets work before?)

Output: Ranked list of opportunities.

Layer 4: Ticket Generation

For each high-score opportunity:

AI generates a "growth ticket":

Issue: What's wrong (specific)
Impact: Expected result in € or %
Action: Exactly what to do
Reasoning: Why AI recommends this
Confidence: How sure we are

Example ticket:

ISSUE: 47 keywords in "Healthcare Services" campaign have spent €3,200 over last 90 days with zero attributed conversions. IMPACT: Pause these keywords → Save €1,067/month (~€12,800/year) ACTION: 1. Review keyword list (attached) 2. Pause in Google Ads 3. Or add to excluded keywords REASONING: These keywords attract clicks (3.2% CTR) but never convert. Likely wrong intent match. Spending budget that could go to proven keywords. CONFIDENCE: 95% (Sufficient data: 90 days, 1,847 clicks, statistical significance)

This is actionable.

Not "here's some data." But "do this, expect this."

Layer 5: Learning Loop

When user implements ticket:

Track outcome
Compare to prediction
Update models
Improve future recommendations

AI gets smarter over time.

---

The LLM Decision

We use Claude (Anthropic) as core AI.

Not GPT-4.

Why?

1. Context Window

Claude handles 200K tokens.

That's:

Entire campaign data
Historical performance
All tickets generated
User feedback
Domain knowledge

In one context.

GPT-4's smaller window forced chunking. Lost connections between data points.

2. Reasoning Quality

We tested both on marketing analysis.

Example task: "Find wasted spend in this Google Ads account"

GPT-4: Found obvious issues (zero-conversion keywords) Missed subtle patterns (audience overlap, time-of-day inefficiency)

Claude: Found obvious issues AND subtle patterns Better at multi-step reasoning More nuanced understanding of marketing context

For our use case: Claude wins.

3. Instruction Following

AI agent needs to follow complex instructions:

Analyze data
Apply marketing principles
Consider business context
Generate specific recommendations
Format as structured tickets

Claude excels at instruction following.

4. Cost Efficiency

Claude pricing competitive with GPT-4. But better results. So effectively cheaper per quality output.

---

Technical Challenges We Solved

Challenge 1: Real-Time vs Batch Processing

Problem:

Marketing data changes constantly.

Run analysis every minute? Expensive, unnecessary. Run once per day? Miss urgent issues.

Solution:

Hybrid approach:

Batch analysis every 6 hours (comprehensive)
Real-time monitors for critical alerts (spend spikes, tracking breaks)
Event-triggered analysis (campaign launches, budget changes)

Result: Balance between coverage and cost.

Challenge 2: Statistical Significance

Problem:

Small campaigns have high variance.

Day 1: 10% conversion rate Day 2: 2% conversion rate

Is this a problem or noise?

Bad AI: Alerts on every fluctuation Good AI: Only alerts when statistically significant

Solution:

Bayesian confidence intervals.

Calculate probability that change is real.

Only generate ticket if >85% confidence.

Result: No false alarms.

Challenge 3: Domain Knowledge Injection

Problem:

Claude doesn't inherently know marketing.

Needs to learn:

What's a good CTR?
What's normal ROAS variance?
When is creative fatigued?
What's audience overlap?

Solution:

20 years of marketing knowledge encoded as:

Benchmark databases
Pattern templates
Decision trees
Heuristics
Domain-specific prompts

Example prompt snippet:

You are analyzing Google Ads performance. Normal CTR ranges by campaign type: - Search brand: 8-15% - Search non-brand: 2-5% - Display: 0.3-0.8% - Shopping: 0.5-1.2% Creative fatigue typically occurs after: - 10,000-15,000 impressions (display) - 5,000-8,000 impressions (social) Audience overlap above 60% between ad sets indicates potential auction competition...

Thousands of lines of domain knowledge.

This is what makes AI useful for marketing. Not just raw LLM intelligence.

Challenge 4: Action Specificity

Problem:

Early versions generated vague recommendations: "Consider optimizing keyword performance"

Useless.

Solution:

Constrain AI outputs to specific formats:

Must include exact data (keyword IDs, campaign names)
Must provide step-by-step instructions
Must quantify expected impact
Must explain reasoning

Enforcement: Structured JSON output format. AI must fill every field.

Result: Every ticket is actionable.

Challenge 5: Handling Multiple Platforms

Problem:

Google Ads optimization is different from Meta optimization.

Google: Keywords, search terms, quality score Meta: Creative, audiences, placement

Different principles.

Solution:

Platform-specific analysis modules:

Google Ads analyzer (search focus)
Meta analyzer (creative + audience focus)
GA4 analyzer (journey focus)
Cross-platform analyzer (attribution)

Each with domain-specific logic.

Result: Deep expertise per platform.

---

What We Learned

Learning 1: Accuracy > Coverage

Early version:

Analyzed everything. Generated 200+ tickets per account.

Users overwhelmed.

Now:

Analyze everything. Generate 10-20 high-confidence tickets.

Quality over quantity.

Better to find top 10 opportunities than surface 200 maybes.

Learning 2: Explanation Matters

Early version:

"Pause these 47 keywords."

Users: "Why?"

Now:

"Pause these 47 keywords because:

Zero conversions in 90 days
€3,200 spent
Good CTR but wrong intent match
Budget better used on proven keywords"

Users need reasoning to trust recommendations.

Learning 3: Users Want Control

Early version:

"AI recommends this. Execute? Yes/No"

Felt like black box.

Now:

"AI recommends this because [reasoning]. You can:

Execute (one-click)
Modify and execute
Dismiss
Snooze for later
Ask AI to explain more"

Users want to understand and control.

Learning 4: Context is Everything

Without context:

"Campaign X has 2.1 ROAS"

Is that good or bad?

With context:

Your average: 4.2 ROAS
Industry benchmark: 3.8 ROAS
This campaign's historical: 5.1 ROAS

"Campaign X has 2.1 ROAS. → Underperforming. Investigate."

Absolute numbers mean nothing without context.

Learning 5: Continuous Monitoring > Periodic Reports

Old way:

Generate report weekly. Review on Monday. Implement changes Tuesday.

Problem happened Friday. Not caught until Monday. €10K wasted over weekend.

New way:

AI monitors 24/7. Alert in real-time. Problem Friday morning. Fixed Friday afternoon. €9K saved.

Continuous monitoring catches issues before they're expensive.

---

The Tech Stack

For those curious:

Backend:

Python (FastAPI)
PostgreSQL (data storage)
Redis (caching)
Celery (task queue)

AI:

Anthropic Claude 3.5 Sonnet (primary)
Custom embeddings for pattern matching
scikit-learn (statistical analysis)

Data Pipeline:

Apache Airflow (orchestration)
dbt (transformations)
BigQuery (data warehouse)

Infrastructure:

Google Cloud Platform
Kubernetes (container orchestration)
Terraform (infrastructure as code)

Monitoring:

Prometheus + Grafana
Sentry (error tracking)
Custom dashboards

Not reinventing the wheel.

Using proven tools. Focus on AI logic, not infrastructure.

---

What's Next

We're working on:

1. Predictive Tickets

Not just "this is wrong." But "this will be wrong in 3 days."

Prevent problems before they happen.

2. Automated Execution

User approves once: "Auto-pause keywords with zero conversions after 30 days"

AI executes automatically.

3. Strategic Recommendations

Today: Tactical optimization Tomorrow: Strategic insights

"Your creative performs best with testimonials. Consider brand campaign focused on customer stories."

4. Multi-Account Learning

Learn from patterns across all accounts. "This type of campaign works 2.3x better in fintech vs e-commerce."

Cross-account intelligence.

5. Natural Language Queries

"Why did CAC increase last week?"

AI investigates and explains.

Already works. Making it better.

---

Advice for AI Builders

If you're building AI agents:

1. Domain knowledge is 80% of value

Raw LLM = 20% of solution Domain expertise = 80%

Encode your expertise.

2. Start narrow

Don't try to build "AI for everything."

Pick one problem. Solve it well. Expand.

3. Action > insights

Users don't want reports. They want "do this."

Focus on actionability.

4. Trust through transparency

Show reasoning. Explain logic. Let users verify.

Black boxes fail.

5. Continuous learning

Track outcomes. Improve models. Get better over time.

Static AI becomes stale.

---

The Future is Agentic

Chatbots are 2023.

Agents are 2025.

The difference:

Chatbots: reactive
Agents: proactive

Next evolution:

Not "AI that answers questions" But "AI that solves problems"

Autonomously.

That's what we're building.

---

See It in Action

Want to see how the AI agent works?

Schedule a technical demo

I'll show you:

How AI analyzes campaigns
How tickets are generated
The reasoning process
Technical architecture (if you want details)

I love talking about this stuff.

---

About Berner Setterwall

Berner is the Co-Founder and CTO of Cogny. He previously built the AI optimization platform at Campanja that worked with Netflix, Zalando, and Momondo before its acquisition. He has 15 years of experience building machine learning systems for digital advertising, including real-time bidding systems, attribution models, and predictive analytics platforms. He's particularly passionate about making AI practical and actionable for marketing teams.

Connect:

LinkedIn: linkedin.com/in/bernersetterwall
Email: berner@cogny.com
GitHub: github.com/bernersetterwall

Last Updated: January 8, 2025