• Portfolio
  • Blog
  • About
    • 01 AUG 2025 10 min READ Case study

    AI Agentic PPC Optimisation

    Trust is the foundation of every relationship. And whether we realise it or not, we are already in a serious one with technology. It is evolving faster than we are and our trust is still catching up, trying to define what it should look like. This project is my exploration of trust in agentic flows.

    PROJECT

    SellQ — Designing an agentic AI workflow for PPC optimisation (solo exploration).

    PROBLEM

    PPC data without guidance leads to delays, wasted spend and missed growth.

    SOLUTION

    Designed an AI-powered agentic workflow that guides users to decisions while keeping them in control.

    IMPACT

    • check_boxReduced ad spend waste
    • check_box~90% faster task completion
    • check_box$100/month savings on third-party tools
    • check_boxLower mental load
    • check_boxReusable AI+UX trust framework

    PPC complexity

    Sellers rely on ads to grow their businesses, but many struggle with advertising. Some don’t even know the difference between ACoS and TACoS, let alone how to optimise a campaign.

    Third-party tools promise to simplify this, and many now include GenAI features. Sellers are willing to try almost anything if it might help, even paying $100+ per month, yet most tools still deliver mediocre results.

    Adtomic burned through my budget in weeks and I had no idea why it made those changes.

    The AI kept recommending keywords I’d already tested — nothing new, just wasted spend.

    I still have to double-check everything manually because I don’t trust the recommendations.

    Diving deep

    I wanted to understand how sellers actually think about ads and automation. I dug into:

    • Competitor analysis (Adtominc, Helium10, JungleScout)
    • Community research (videos, seller groups, Reddit, podcasts)
    • Benchmarking & patterns (agentic UX to design for trust)

    Key insights

    • Sellers are overwhelmed by complexity

      • PPC tools often show an overwhelming amount of data without clear prioritisation.

    • Trust in automation is low

      • Existing tools act like black boxes; they make changes but rarely explain why.

    • Sellers want guidance, not blind automation

      • The magic formula is transparency + choice. People want to understand tradeoffs, preview changes and approve actions.

    • One metric isn’t enough

      • Tools that optimise only for ACoS miss the bigger picture. Sellers want multidimensional insights.

    Hypothesis

    bookmark

    Agentic workflow optimised for trust and transparency reduces decision fatigue, improves outcomes and keeps sellers in control.

    Seller & Business needs

    store

    Seller needs

    • Explains the 'why' behind the recommendation
    • Highlights high-impact opportunities
    • Make faster, simpler decisions without losing detail
    • Stay in control with explicit approvals and transparency
    monitoring

    Business needs

    • Speed up seller decision-making
    • Improve ROI with smarter optimisation
    • Build trust in AI-driven tools
    • Boost retention by making tools indispensable

    North star metrics

    • check_boxFaster actions (~80% faster)
    • check_boxIncreased trust (CSAT > 80%)
    • check_boxAdapted recommendations (65–75%)
    • check_boxGained efficiency (TACoS +8–10%)

    Agentic UI Principles

    I focused on keeping sellers in control, making the why behind recommendations clear and leveraging the native Marketplace UI to guide them without adding friction.

    I defined these principles around user needs first, then validated them using agentic AI research.

    Principle Validated by research
    Co-planning before acting ReAct + Planner–Executor → the agent shows its plan first and explains why before doing anything.
    User-in-the-Loop control Human-in-the-loop UX + Magnetic UI → sellers approve actions and the agent highlights relevant UI elements for clarity.
    Self-Reflection & Recovery Reflexion → the agent reviews past suggestions and recommends fixes if performance drops.
    Portfolio-Level multi-tasking Multi-Agent Chains → the agent can manage multiple campaigns at once while surfacing the highest-impact actions first.
    Context-Aware, Non-Intrusive design Copilot UX research → the agent stays in its own panel and doesn’t interrupt existing workflows.
    Design for trust and clarity Transparent Agentic UI + Action-Oriented UI → recommendations are explained, evidence-based, and focused on clear next steps.

    Strategic constrains

    Scope constraints:

    • Focused on a small set of PPC intents
    • Designed only one happy path per intent
    • Adapted responses for a mid-level seller familiar with PPC language
    • Reused existing Marketplace dashboard components
    • Planned accessibility and adaptability work for post-MVP

    Prioritisation choices:

    • Skipped full tone/voice refinement process
    • Simplified all reports in the portal

    Agent flows

    Designing AI agents is about building trust. I used my agentic design principles to make sure recommendations feel transparent, collaborative and not pushy.

    Onboarding to PPC Campaign

    Sellers struggle to know where to start with PPC, often fearing wasted spend. I designed the onboarding flow around co-planning and transparency:

    • The agent surfaces missed opportunities using the seller’s data and sets realistic expectations.
    • Each recommendation explains the “why” and requires seller approval, keeping automation clear and optional.
    Guided PPC onboarding”
    Guided PPC onboarding Figma link

    Creating Discovery Auto Campaign

    Auto campaigns can feel like a black box, making sellers hesitant to trust automation.

    • The agent confirms business goals upfront and adapts recommendations accordingly, making automation collaborative, not prescriptive.
    • Setting is explained in plain language, with full flexibility to override choices or ask follow-up questions.
    Collaborative auto campaigns
    Collaborative auto campaigns Figma link

    Keywords harvesting

    Sellers are unsure how to interpret keyword results, making optimisation time-consuming and risky.

    • The agent acts like a consultant, showing its work step by step and explaining why each keyword is recommended or discarded.
    • Sellers can interrupt anytime to ask follow-up questions, tweak numbers or override choices.
    Smarter keyword decisions
    Smarter keyword decisions Figma link

    Biweekly reports

    Sellers need clear visibility into campaign performance and guidance when things go wrong.

    • The agent delivers biweekly reports, timed to match when PPC data stabilises on the marketplace.
    • When results miss expectations, the agent is fully transparent and proposes a guided recovery plan.
    Biweekly report of the campaign performance
    Actionable performance reports Figma link

    Urgent issue handling

    Some issues can’t wait for the biweekly report, like wasted spend, stalled conversions, or hitting budget caps.

    • The agent flags urgent problems instantly and explains what’s happening and why in plain language.
    • Sellers get a clear recovery plan, with step-by-step guidance to act quickly without losing control.
    Real-time issue handling
    Real-time issue handling Figma link

    System guardrails

    Automation always creates risk (overspending, biased results, errors). These guardrails protect sellers behind the scenes before any recommendation reaches them:

    • Safety Nets

      • Budget caps, policy scans, data freshness

    • Transparency

      • Always show “why” + preview of impact

    • Fairness & Accountability

      • Bias audits + recovery steps

    • Privacy & Context

      • Seller-only data + category signals

    Designing trust

    I complemented the backend guardrails with front-end trust patterns so sellers could see, understand, and control automation:

    brand_awareness

    Tone

    • Clear, plain-English
    • Concise
    • Tailored to user data and history
    chat_bubble

    Ai-Ready prompts

    • Structured
    • Transparent
    • Respecting guardrails
    healing

    Failure handling

    • Honest about confidence level
    • Accountable
    • Quick to recover

    Impact

    By reducing manual steps, clarifying recommendations and building trust into automation, I predicted big gains for both sellers and the business.

    Seller impact

    • check_box~90% faster tasks
    • check_boxLower mental load
    • check_boxFull seller control
    • check_boxSave ~$100/month on 3rd party tools

    Business impact

    • check_box+5–10% more sales
    • check_boxReduce wasted spend
    • check_boxScalable AI+UX trust pattern
    • check_boxReusable framework across workflows

    Key learnings

    This was a one-week sprint, so I optimised for speed and focused on proving value fast.

    Since then, I’ve deepened my understanding of agentic workflows and would approach future iterations differently:

    • Explore inline recommendations

      • Test surfacing suggestions directly in the seller’s flow instead of a side panel.

    • Refine context-awareness

      • Improve continuity across chat history, actions, and recommendations.

    Next steps

    • Validate user interactions → observe how sellers engage with recommendations, approvals, and inline guidance to refine the workflow.
    • Test tolerance for multi-tasking → Explore how ready sellers are to trust the agent to execute multiple actions across campaigns simultaneously.