Diffe.rent~3 min read
The agent had to earn its way past “just another tool.”
A multi-phase project, a product partners had already compared with a rival, and a pricing decision I did not own.
- Role: Product management and design
- Timeline: Multi-phase
- Team: Cross-functional: engineering on the conversation model, and the commercial side on pricing. Product and design were mine.
- Impact: 65% adoption
- Platform: Web
Adoption
Lead quality
Time saved
Adoption is the share of existing clients who tried the agent, across a base of 25, so 65% is about sixteen of them, on contracts running roughly $100k each. That is a small denominator. Lead quality is counted in move-ins the property can trace back to an Apartment List lead, so a partner can check it against their own leases. Metrics took three months to come back, which is the window all three figures are read over; the 300 hours accumulated across it. Still missing: the deal count behind the 3×.
Overview
Partners judged our AI agent against Elise.AI, so it had to show checkable work before asking for trust.
Diffe.rent’s on-site AI agent was meant to remove routine leasing work. Partners had already met Elise.AI, liked it, and dismissed ours as just another tool. Coming second to a product they rated was worse than being unknown.
Worked with: engineering on the conversation model; the commercial side on pricing; 15 people interviewed.
What I owned
I owned
- The AI experience end to end, and the sequencing that earned adoption
- Partner research, including the Elise.AI comparison that reset the pitch
- The call to answer with media rather than text, and what that cost on thin listings
- The handoff summary that replaced a hand-written note
Shared cross-functionally
- Pricing and packaging of the standalone agent
- Where the agent handed a conversation to a human
- What the conversation model could reliably answer
Research
Partners compared us with a product that did more than we showed.
I interviewed 15 people, including the Elise.AI comparison that reset the pitch. AI was near the top of its hype cycle, and partners judged every product in the category on merit. The questions a text agent handled worst were about the unit itself: what floor, what view, whether it takes pets.
Decision: media in the conversation
The agent answers with a listing’s own photos, floorplans, and tours instead of describing them.
Text answers and links sent renters away to look, and renters who leave rarely return. I chose to surface media already attached to the listing, inline. Partners could then check the agent’s work on their own listings before trusting it with a lead.
The cost: the agent is only as good as a partner’s uploads. Four dark photos could make it confidently show four dark photos. No mitigation fixed that. I accepted the trade.
Decision: handoff summary
A generated summary sits beside the transcript, so leasing agents can pick up without re-asking.
The handoff note had been written by a person. A full transcript was complete but too long to read mid-shift. The summary shows what was established and where a lead moved from low to high intent.
A wrong summary could lead an agent to restate something the renter never said. So it sits alongside the conversation and can be checked in one click.
Redesign
The same agent continues over SMS, where leads actually reply.
Scope
Inline media, the handoff summary, and SMS continuation shipped.
The standalone product launched with its own pricing. Competitors leapfrogged it, and the feature was folded back into the main product.
Results
65% of existing clients tried the agent, lead quality rose threefold, and 300 hours or more were saved.
The 65% belongs to a product that no longer exists in the shape I designed
It counts partners who tried the agent, not partners still using it a quarter later. I do not have a retention number. Read 65% as evidence that partners gave it a try.
