Context
The next problem was discovery: surfacing the right tool to the right account.
Status
Exploration. Tested as a Figma Make prototype. Nothing here shipped.
Question
Can an assistant earn trust before it earns capability. Will it know if and when to step back?
Business objective: retention
Starting point: the catalog solved browsing, not discovery. So customers had to find the right tool on their own.
Business objective: once data is connected to a tool through an integration, and sits in a customer workflow, it is hard for that customer not to renew the yearly contract.
So the job here is not engagement with an AI feature, it is more connected and active accounts sooner.
Intelligence
A script already running on the page could see which third-party tools a company's own site was already calling:
AB Tasty, Adobe Analytics etc. This was shown in the product as “Top recommendations”.
Call it what it is, it is not AI. We used recommendations to test if the trust signal was there.
The signal
A small spark icon sits above the three cards of the curated row.
Maturity, not just capability
Early on, a banner is there when the page loads, to push usage.
A blank field is an articulation barrier: the user must know what to ask. So the banner offers ready-made prompts, built on the tools the account already runs.
This is Discovery, not Activation.
Once someone types unprompted, the banner has done its job and the system should recede:
a proficient user gets the top-right icon alone, nothing proactive.
The payoff: AB Tasty says what, AI says why
A test result tells you what happened. It rarely tells you why.
Each finding has a number, one sentence of evidence, and a link to the same view in the product.
The link is proof, not an exit: the user can check the claim where the data lives.
POV: trust before intelligence
An AI badge is a promise. If what's underneath doesn't hold it, trust corrodes faster than if the badge had never shipped.
“Top recommendations” here is AI-shaped, not AI, and that's fine.
Before a real assistant panel earns the right to exist, three things have to be true:
- It says what it knows and why: “you use AB Tasty” beats “we think you'd like this”
- The user can correct a wrong assumption in one tap, not a support ticket
- Nothing hides behind “AI” for something that's actually deterministic lookup
The chat panel is the real next step. It needs to reason over the user's specific setup, not just look it up.
That's the actual line between a cheap trust signal and a genuine assistant, and it's the line this exploration stops at on purpose.
Honest close
Nothing here shipped. The Top recommendations signal was designed and could ship cheaply on top of data already collected;
the assistant panel exists as a Figma flow and a Figma Make prototype.
Before building either for real, two bets would need proof, not opinion:
- Does the badge move installs and usage, not only clicks? Installs are what protect the renewal.
- Does the assistant earn its own surface, or does it ride the existing catalog UI without adding a new thing to learn?
Want to discuss this?