Optimizely: Opal AI

Designing how AI agents embed into marketer workflows.

Role

Product designer

(masters capstone)

Team

3 fellow designers

Advised by Optimizely's VP of Product and a senior PM

Timeline

7 months

(February - August 2026)

Final Design

Optimizely's starter-campaign picker: six ready-to-launch campaigns, one suggested by Opal

What it delivered

  1. 1.

    Opal suggests the first campaign instead of waiting to be asked.

    See The suggestion
  2. 2.

    Ready-made campaigns a marketer launches alone, in under 10 minutes.

    See Pick one
  3. 3.

    Results that say how sure they are before any number.

    See Read the result

Problem

Optimizely makes software that companies use to test and tailor their websites. One feature, personalization, shows different visitors different things: a returning shopper sees a welcome-back offer, a first-timer sees an introduction.

Nearly every customer buys it. Of about 4,000 accounts, more than half had never run a single campaign, and one product manager put active use nearer 10 to 20 percent. Our capstone team was asked to design how Opal, Optimizely's AI assistant, could help. Nine interviews traced the gap to one handoff:

  1. The person who buys it

    Sold

    A company's VP buys the promise: a site that adapts to each visitor.

  2. Signed

    The contract is signed. Her part is done.

  3. The person who has to use it

    Handed off

    A marketer gets the tool, with no brief and no one who set it up.

    Stalls here

  4. Set up

    Seven setup stages, and a developer, before anything shows on the site.

    Stalls here

  5. Never used

    No campaign ever runs. The tool sits, then gets cancelled.

Nobody cancels because it failed. They cancel because nothing ever ran.

Only 10 or 20% of our customers who have personalization use it. They’re paying for it, and they’re just not using it.

- Senior product manager, Optimizely

Design Goals

Three rules the design follows, each one a fix for something we heard.

1.  Suggest, don't ask

An empty chat box asks her to know what to type. A named suggestion only asks for a yes or no.

2.  Only offer what she can finish today

If an idea would need a developer or an upgrade, it doesn't show up.

3.  Say how sure it is

Results lead with plain words, early signs or enough to act, before any number.

Key Decisions

Four moments, in the order the marketer meets them. Every screen runs on Optimizely's real interface.

  1. The suggestion on the Optimizations page

    1.  The suggestion

  2. The starter-campaign picker

    2.  Pick one

  3. The setup page

    3.  Check and launch

  4. The results page

    4.  Read the result

Outcome

In testing, both marketers launched a campaign unaided and read a flat result as a setup question, not a failure. Optimizely's VP of web experimentation asked for a PRD their engineering teams could pick up, so that became the handoff: a 22-section spec covering the user story, the screens, the seam between Opal and the existing product, the data each screen needs, guardrails, acceptance criteria and success metrics. The capstone closed with three presentations: to the public, to our professors and advisors, and to the advisors together with the Optimizely team.

BeforeAfter
Her first campaignSeven setup stages and a developerOne suggestion, one page of checks, under 10 minutes
OpalA chat box waiting for a questionA suggestion already on her page
The resultNumbers she had to interpret alonePlain words on how sure it is, then the number

Full research corpus, decision records and build are in a private repo. Happy to walk through it.

Next up