I explored a wide range of UI patterns and Copilot‑aligned treatments, including:

• Accordions to surface recommendations without overwhelming the form
• Inline banners that expanded to reveal options or opened a side panel
• Subtle text links that triggered lightweight popovers
• Copilot cards that could expand inline or open a sidecar
• Embedding recommendations in the Ad Strength card for contextual guidance
• Emerging Microsoft‑wide Copilot patterns like the “nudge” component
• Visual treatments to clearly signal AI involvement
• Inline & embedded variants that blended Copilot actions directly into the form

prompt‑first models

These captured advertiser intent before entering the form. Each option tested a different philosophy of AI interaction: direct, conversational, hybrid, or inferred. The major risk was that partners are always hesitant to add any steps to the flow, so we would need to validate the usefulness of this.

• URL + prompt → ad previews → form
• URL → side‑by‑side conversation + preview → form
• sticky prompt at top → ideas → form
• URL → ideas → prefilled form + Copilot sidecar

inline models

These embedded Copilot directly inside the form, exploring how to bring AI into the workflow without disrupting it.

• inline components
• floating components
• sidecar variations
• contextual triggers
• actions where the advertiser would need it (change tone/rewrite/find similar)

leadership alignment

Some of these ideas incorporated a copilot sidecar or other conversational UX. Early in the exploration phase, product leadership strongly advocated for a conversational Copilot experience modeled after Copilot in Edge, where clicking an element opens a side‑chat. They saw this as the most “Copilot‑forward” way to showcase AI inside the Ads wizard and pushed to replicate that pattern directly.

I intentionally pushed back —not because the sidecar was off the table, but because committing to any specific model this early risked locking us into the wrong paradigm before we understood advertiser needs. Instead, I acknowledged their desire to highlight Copilot, incorporated conversational elements into explorations, and evaluated them alongside non‑conversational patterns. User validation would be critical before converging on a final solution.

Research insights

• Advertiser expertise determines preferred model, not UI preference
• Experts preferred Akita since it was fast and direct. They already know their campaign goals and want Copilot to execute quickly
• Experts preferred Akita since it was fast and direct. They already know their campaign goals and want Copilot to execute quickly
• Infrequent or less confident advertisers preferred Beagle due to the extra context that helped them think through goals and feel supported
• Conversational UX is not the right default — felt slow and unnecessary, adding too much friction for experienced users who already know what they want
• Conversational flows are useful for providing guidance when confidence is low or when users were exploring ideas with a more flexible plan
• Speed and clarity outweigh “AI richness” — Users valued Copilot most when it reduced steps, not when it asked more questions
• Hybrid model lacked a clear advantage — didn’t deliver the speed experts wanted or the support novices needed

I pushed on this, advocating for at least the basic minimum to allow for modifying the assets that copilot provides as recommendations. I offered some suggestions:

• a “replace assets” flow or another way to refresh recommendations
• a prompt modification dialog
• an inline prompt component that was much lighter than the north star flow

Creating temporary components though would have been throwaway work if we still wanted to move towards the north star vision.

the compromise

A “copilot” style banner explaining what Copilot was doing, a lightweight dialog to collect prompt input, and theme modification inside the dialog. This preserved user control without derailing the roadmap.

overall takeaways

Advertisers across SMB, mid‑market, and enterprise see strong potential in Copilot for accelerating asset selection, generating new creative, and inspiring alternative copy—but only when they can provide input first, maintain control over approved assets, and avoid unnecessary back‑and‑forth. The pilot’s low adoption stems from lack of relevancy, mismatched expectations, and workflows that don’t align with how advertisers actually prepare campaigns.

• Confirmed: URL-only recommendations are too generic and users see value in providing input before seeing recommendations
• Banner placement felt premature — users hadn’t added content yet
• Inline, contextual suggestions align better with “edit” mental models
• SMBs struggle most with starting from scratch and welcome AI for inspiration
• Themes seemed irrelevant as users expected to see a preview of actual assets
• AI image generation was consistently described as “game‑changing,” since the output was unique and without licensing
• Even SMB users want to check with their team before adding new images
• Participants almost always typed their own headlines first, then used AI to refine
• the dialog → close → refresh loop was tedious and users wanted to see the effect of their prompt input in order to know if they were setting the right context
• Edge browser adds popups over our recommendations dropdowns

what users expect from Copilot

• Provide side‑by‑side recommendations that gives context for choices
• Use their assets and their copy as the primary signal
• Offer quick inspiration, not long prompt cycles
• Generate unique, high‑quality images when needed (before starting campaign)
• Help with editing, not just creating from scratch

what users DO NOT want

• Auto‑fill everything before they add their own content
• Require multiple rounds of prompt refinement
• Surface irrelevant content as a suggestion

inline recommendations

I looked at how AI surfaces contextual suggestions in the moment that offer help editing those existing assets that advertisers provide

• Sources for recommendations prioritize existing content like their asset library or social
• Help advertisers improve their headlines by adjusting the tone, aligning to their brand, or rewriting it entirely with a prompt
• Rewrite to stay within character limits to meet ad surface constraints (most time consuming part of editing text for users)
• Adding similar text assets or ones they wouldn't have thought of
• Auto fill assets for them but allow to remove all in case they didn't meet expectations, adjusting the prompt to try again

influenced the org’s AI strategy

I shifted the conversation from “AI everywhere” to AI where it helps — contextual, clear + transparent, and user‑controlled, at a pivotal moment for Microsoft Ads.

shaped the foundation for brand‑aligned assets

My work here directly led to the Brand Guidelines initiative, which uses brand inputs to improve recommendation relevancy

improved advertiser trust

By grounding AI in user intent and clear communication, I helped reduce the “black box” feeling of PMax and other automated campaign types.

provided a roadmap for future AI guidance

My revised North Star became the reference point for future work on human–AI interaction across Ads Studio and campaign creation

what i’m most proud of

• Strengthened cross‑functional alignment — In the midst of internal chaos, I navigated PM pressure, engineering constraints, and leadership expectations while keeping the user at the center
• Core user value — I pushed back on designs that would have harmed user trust and used research to steer the team toward a model that met advertisers where they are
• Scalable foundations — I created the groundwork for future AI‑powered guidance and demonstrated system‑level thinking, not just UI design