The challenge
The client’s customers were asking for AI features, and competitors were starting to ship them. The in-house team knew the product and the rental market deeply, but had no experience putting LLM features into production.
The product was a 12-year-old multi-tenant Laravel codebase holding tenants’ personal information. Strict tenant isolation and compliance with Japan’s personal information protection law were essential.
Our approach
- 01
Fit the existing product
We worked inside their codebase and conventions, so the in-house team could own and extend the features after launch.
- 02
Tenant isolation by design
Every AI request is scoped through the existing authorization layer. The model only sees data the current user is already allowed to see.
- 03
Ship in slices
One feature at a time behind feature flags, each with usage and quality metrics, so the team could see what customers actually valued.
What we built
Communication drafting
Drafts tenant notices, rent reminders and renewal guidance in polite Japanese, following each company’s approved templates.
Repair triage
Summarizes and categorizes incoming repair requests, including photos, and suggests urgency and contractor type.
Ask your portfolio
Natural-language questions over portfolio data, answered through safe, pre-approved query templates.
AI platform layer
Shared services for prompts, usage metering, billing and evaluation that the in-house team now builds on.
How it flows
- User in app
- Auth & tenant scope
- Copilot service
- Tools & data
- Response
- Usage metering
How we measured it
Each feature has its own evaluation set, built from anonymized real cases. Tenant-isolation and personal-data masking tests run in CI on every deployment.
Results
- 14% of active customer accounts used the copilot within the first 90 days, a strong adoption rate for a new paid add-on.
- Launched as a paid add-on plan, creating a new revenue line.
- The in-house team has since shipped two more AI features on the platform layer without outside help.
- No cross-tenant data incidents.
Timeline
Weeks 1–2
Discovery
Customer interviews, codebase review, feature shortlist
Weeks 3–4
Design
UX in the existing design system, platform-layer architecture
Weeks 5–10
Build
Three features behind flags, beta with 30 customers
Weeks 11–12
Launch
Pricing, billing integration, general availability