Bank Reconciliation With AI: From 8 Hours to 23 Minutes
What bank reconciliation is, why it took one real-estate fund a whole day, and where those 8 hours went before an AI agent brought them down to 23 minutes.
Field notes on shipping AI agents inside regulated finance.
Each episode below is a world, in reading order. Each world is split into short stages of 5 to 7 minutes. Start at World 1, or jump to any world that answers your question.
What bank reconciliation is, why it took one real-estate fund a whole day, and where those 8 hours went before an AI agent brought them down to 23 minutes.
I delivered 11 production AI agents to a US real-estate fund managing roughly $3B in assets. Here is the one foundation they shared and the checks built into it.
Verified citations and evidence-based abstention: the RAG design I shipped inside a regulated fund, where a RAG that abstains beats one that is confidently wrong.
Prompt injection is not solved, so contain it. Why the permission check for a money-moving AI agent belongs in the payment system, not in the agent.
AI Act high-risk obligations moved to 2 December 2027 and 2 August 2028; Article 72 monitoring did not. An engineering checklist, not a legal summary.
How a release gate turns an AI agent’s test results into one verdict, ship, hold or block, and why a clean test run can still say “not yet”.
I deployed about 46 AI agents for Integra’s internal operations. Here is an illustrative monthly cost model: tokens, platform, review and maintenance.
A proposed audit-record design for AI agents: 14 fields that capture source evidence, checks, decisions and review, so one output can be investigated later.
What the EU AI Act technical file is, why a hand-written copy can drift, and which half a build can generate from each release while named people own the rest.
An eval threshold is the score a new AI agent version must reach before launch. Why a copied one fails, what the record holds, and the loop that earns it.