Lendscape
AI-Powered SMB Lending Ecosystem
Design exploration · 2025 · Designer · Figma, After Effects
The model proposes. The platform validates. A person decides.
- The question
- How far can AI go in regulated lending before a person has to take the call?
- What I did
- Designed a lending ecosystem for borrowers, loan officers and the platform, with AI channeled through deterministic checks, then built the loan officer side as a working demo.
- What it shows
- Where to draw the line between AI and people, and how to make it auditable. A design exploration, not a shipped product.
See it live
lendscape.postminimal.agency/officer
In regulated lending you can't hand decisions to a probabilistic system, but you can let it do the analysis people are too slow to do at scale. Lendscape explores that boundary. The AI is non-deterministicA system where the same input can produce different outputs each time. AI and LLMs behave this way. enough to read messy qualitative data, but it's channeled through deterministic compliance logic, and a person owns every final call.
AI can read the mess. It can't own the call.
AI triages applications faster than any team, but compliance needs auditability, and auditability needs determinism. So the question isn't whether to use AI. It's where to put the boundary, and how to make it explicit enough to audit. Existing tools put it in one of three places.
| Approach | Who decides | Where the AI sits | The catch |
|---|---|---|---|
| Traditional platforms | The loan officer | Nowhere | Every assessment is manual, so it scales poorly |
| AI-first fintech | The AI | In the decision | Breaks where compliance needs human judgment |
| Copilot assistants | A person | Bolted on, outside the workflow | Its suggestions skip the compliance logic |
| Lendscape | A person | Channeled through deterministic checks | Slower than full automation, by design |
Five layers keep the model from deciding.
The layers are ordered so that AI variability is contained before it can reach a decision. The model never touches raw input and never ships an output. Everything it makes is a proposal, checked the same way every time.
- 01People, structured
Borrowers, officers and internal teams provide information, documents and questions, in a shape the system can work with.
- Enter info
- Upload documents
- Ask a question
- 02Normalized data
Structured and unstructured inputs, from loan data and account activity to document contents, cleaned into one format.
- Loan data
- Account activity
- Document contents
- 03The model proposesNon-deterministic
Reads the normalized data and drafts work, with its reasoning attached. It proposes and never decides.
- Match a product
- Flag a risk
- Draft a request
- Draft a task
- 04The platform validatesDeterministic
Eligibility rules, risk scoring, checklists and compliance run against every proposal, with the same output for the same input. It decides what is allowed, not what happens.
- Eligibility
- Risk score
- Checklist
- Compliance
- 05A person decides
Stakeholders select the product, approve the credit, ask for more, or handle the exception. The decision, and the accountability, stay with them.
- Select a product
- Approve credit
- Request information
Each party sees the AI differently, by design.
Borrowers, loan officers and the platform share one AI and compliance core, exposed differently to each. The borrower sees only its results, as matched offers. The officer edits its drafts. The platform monitors its insights. One unified tool would have been simpler, but blurring the roles would blur who is accountable for each decision.
- Pre-qualifies
- Sees matched offers
- Applies
- Receives funds
- Verifies identity
- Runs compliance
- Tracks the record
- Edits the AI's drafts
- Receives the application
- Surfaces AI insights
- Manages credit decisions
- Monitors the portfolio
Blue marks where the AI shows up.
Draft and edit, never approve.
The agent acts first, but only as far as a draft. It reads the portfolio, flags the loans that need a look and drafts a document request for each, with the signals, sources and confidence behind it. Six deterministic checks run on every draft and every edit, and any failure blocks sending. When the agent's own wording fails, say a tone that accuses before anyone has looked, it offers a rewrite. The officer sends, edits or skips.
"Recommend and approve" would be one click faster, but it breeds rubber-stamping, because officers stop reading. So there is no approve button, at any risk level. A sent request can be taken back for a few seconds, and every step lands in an audit trail.
Trust, without the coldness.
Fintech usually buys trust with cold blues and hard grids. The target here was trust that still feels human. Aeonik for display, Lexend for body, readable in screens dense with terms and figures. No stock photography, only art-directed, AI-generated environments without a single literal face.
An honest note.
Lendscape is a concept, not a shipped product, and it hasn't been tested with real borrowers or loan officers. If it were real, validation would start with historical data, comparing what borrowers chose with what the AI would have proposed, then a closed beta with real firms, tuning fidelity against real decisions. The renders are AI-generated and stand in for environmental mood, not product photography.
Designed in Figma and animated in After Effects, with AI-generated renders for the environments. The live loan officer demo was coded from the Figma designs using Claude Opus 5.5.