AI Deduction Assessment Engine

The AI Deduction Assessment Engine uses computer vision and a large language model (LLM) to compare check-in and check-out inspection photos, recommend fair deposit deductions, and surface items that need closer human review — all before an agent sends anything to a tenant.

How It Works

Photo ingestion — Check-in and check-out photos are paired by room or inventory item.
Vision analysis — A vision AI model examines each pair for changes in condition: damage, staining, missing fixtures, deterioration, and so on.
Deduction recommendation — The LLM uses the visual findings to produce a recommended deduction amount per line item, together with a plain-English explanation of the reasoning.
Fair wear and tear adjustment — Recommendations are automatically scaled based on the length of the tenancy. A longer tenancy reduces the amount a landlord can reasonably claim for gradual deterioration.
Disputed item flagging — Items where the evidence is unclear, contradictory, or borderline are flagged for agent attention.
Agent review — Agents see all recommendations in a structured interface and can accept, modify, or reject each one before the deduction schedule is finalised and issued.

Key Concepts

Fair Wear and Tear

Fair wear and tear refers to the reasonable deterioration of a property caused by normal everyday use over time. Under the Renters' Rights Act, landlords cannot claim for deterioration that is consistent with fair wear and tear.

The engine reads the tenancy start and end dates already stored in the system and applies an adjustment factor accordingly — no manual input is required.

Tenancy Length	Wear and Tear Allowance
< 6 months	Minimal
6 months – 2 years	Moderate
2 – 5 years	Significant
5+ years	Substantial

The exact scaling is determined by the AI model based on the item type and condition evidence, not a fixed formula.

Disputed Item Flags

An item is flagged as disputed when:

The check-in and check-out photos are of insufficient quality to make a confident assessment.
The condition change is present but ambiguous (e.g. pre-existing damage that is difficult to distinguish from new damage).
The LLM's confidence score falls below an internal threshold.

Flags are advisory. The agent decides how to handle each flagged item.

Agent Override

Every AI recommendation is a starting point, not a final decision. Agents can:

Accept a recommendation as-is.
Adjust the recommended amount up or down.
Reject a recommendation entirely and record their own figure or a zero deduction.

Nothing is sent to the tenant until the agent has reviewed and confirmed the full deduction schedule.

Workflow Summary

Check-in photos + Check-out photos
          │
          ▼
    Vision AI analysis
          │
          ▼
  Damage detection per item
          │
          ├──► Fair wear and tear adjustment (based on tenancy length)
          │
          ├──► Deduction amount recommendation + rationale
          │
          └──► Disputed item flag (if confidence is low)
          │
          ▼
    Agent review interface
          │
     Accept / Adjust / Reject
          │
          ▼
    Deduction schedule issued to tenant

Compliance

The engine is designed to support compliance with the Renters' Rights Act. All deduction recommendations include a rationale that agents can use to justify decisions to tenants, deposit scheme adjudicators, or other parties.

The agent review step ensures a human remains accountable for every deduction before it is communicated.

Limitations

Photo quality directly affects recommendation accuracy. Blurry, poorly lit, or missing photos will result in more disputed flags and lower-confidence recommendations.
The AI does not have access to inventory condition grades recorded in written reports — visual analysis is based solely on the photos provided.
Recommendations are not legal advice. Agents should apply their own professional judgement, particularly for high-value or complex claims.