AgriFinance, an agri-finance lending startup, engaged Axiom AI Audit for a full AI/ML strategy build, not a bolt-on model. Work started with diagnostic discovery: mapping the business model, operating model, and value stream before any technology was scoped, and running every initiative through a three-gate Innovation Gating System (Value, Readiness, Risk) to manage technical debt and cultural resistance deliberately rather than by accident.

The core deliverable was a loan decision engine built as two separate ML components, a risk tier classifier and a default probability scorer, kept apart deliberately so each could be tuned independently without disturbing the other. Output is four-path, not binary: Approve, Conditional Approve, Refer, or Decline. Every decision surfaces its top contributing factors so loan officers see the reasoning rather than a black box, and decision logic is calibrated so a missed high-risk borrower costs more than an unnecessary referral, reflecting real lending economics. Zero disbursements go out without a human in the loop.

Data governance and a commercial data monetization strategy were built in parallel with the AI system rather than as a later phase, so the same governed data foundation that powers lending decisions also underpins AgriFinance’s data product platform for insurers, development banks, and commodity traders. Ongoing improvement runs through the ARIA Flywheel (Assess, Recommend, Implement, Adapt), a continuous cycle that catches performance drift early and turns every loan outcome into a documented, owned improvement action.

Target outcomes: 70% reduction in loan decision cycle time, 40% reduction in default rate, 95%+ correct identification of high-risk applications before disbursement, and 80%+ straight-through decisions without human review.
Target outcomes: 70% reduction in loan decision cycle time, 40% reduction in default rate, 95%+ correct identification of high-risk applications before disbursement, 80%+ straight-through decisions without human review, 19% reduction in operating costs, and 12% revenue uplift.

Blueface email response system. 30–60 inbound customer emails daily, replaced a 4 hour to 2 days manual response cycle with a six-agentic system built on Cassidy AI. Two linked workflows handle the work: W1 ingests each email through sentiment analysis, orchestration routing, internal KB retrieval, web research (only when the KB has a gap), draft generation, and independent QA validation, then posts the draft to Slack for human review. W2 picks up the human’s response and routes it through one of four paths, from a straight send with zero agents activated to a full revision loop that pulls in KB and web research again, so revision cost stays proportional to how complex the feedback actually is.

A third, fully independent workflow, Evy, scores every run on a 1/2/3 (Fail/Review/Pass) scale, both per agent and across the full six-agent output, and writes a specific fix instruction whenever a score falls short. Nothing ships without a human approving it in Slack.

Design was built around bounded agent . Four of the six agents are shared verbatim across both workflows, so a fix to one propagates everywhere instead of drifting into two versions. Risk mitigation (prompt injection defense, output scanning, human-in-the-loop gating) was designed in before any agent was configured, not bolted on after.

Results: 80%+ of drafts approved on first human review, under 90 seconds from email receipt to Slack draft, 98%+ factual accuracy against the knowledge base, at under $0.65 per email across all agents and evaluation.