Tria Insight360, AI business-intelligence platform across ERP and CRM

Project Overview

Tria Insight360 is an AI business intelligence platform for property-management operators. It connects to Microsoft Business Central and HubSpot, routes natural-language questions to the right specialist agent, and flags revenue leakage between the ERP and the CRM in real time.

Client Background

The client operates in property management and runs the business on two systems of record. Microsoft Business Central holds the financial and lease data. HubSpot CRM holds the tenant, lead, and pipeline data. They wanted a single way to ask questions across both, without exporting to spreadsheets and reconciling by hand.

The Problem

The operator's data lived in two places that did not share a view. Tenants in HubSpot did not always match leases in Business Central, and small mismatches grew into missed charges and unbilled work. Revenue leakage was real, but it surfaced months late in Excel pivots, after the money was already gone. Day-to-day reporting required staff to pull CSVs, line them up by hand, and trust that the join was correct.

Challenges

• Two systems of record with different identifiers for the same tenant and property.
• Read-only access requirements, so the platform had to integrate without disrupting existing workflows.
• Natural-language queries that span both systems and need to be routed to the right data surface.
• Reconciliation that has to run continuously, not as a quarterly batch.
• Multi-tenant data isolation so operators only see their own portfolio.
• Citations back to source records so finance teams trust the answer.

Our Approach

We split the work into three tracks. First, build reliable read connectors to Business Central and HubSpot with the right scopes and rate handling. Second, design a multi-agent layer where a router decides which specialist sees a query, instead of one large prompt trying to do everything. Third, run reconciliation as a background loop so leakage is found before anyone asks. We chose live reads over scheduled exports so answers reflect the current state of the business.

What We Built

User Experience

A React web app where operators type questions in plain English and get answers with citations. Each answer links back to the underlying records in Business Central or HubSpot so the user can verify and act.

Automation

A reconciliation agent runs in the background, links tenants in the CRM to leases in the ERP, and flags gaps in expected revenue. Findings appear in the dashboard and inside query results when relevant.

Integrations

Live, read-only connections to Microsoft Business Central and HubSpot CRM. No nightly CSV dumps, no stale snapshots.

Reporting

Cited answers across both systems, plus a leakage view that lists findings with links to the offending records on both sides of the join.

Security

Multi-tenant isolation at the data layer, scoped API credentials per integration, and read-only access by default so the platform cannot mutate source systems.

Technical Architecture

The platform is a FastAPI service backed by PostgreSQL and Redis, with a React frontend. The OpenAI Agents SDK powers a router agent that dispatches queries to specialist agents per data surface. ERP and CRM connectors handle authentication, rate limits, and entity normalization. A reconciliation worker runs on a schedule and writes findings into PostgreSQL.

Key Decisions

• Multi-agent split over a single mega-prompt. We chose specialist agents per data surface so each one can be tuned, evaluated, and replaced independently. The tradeoff is more orchestration code and a router that has to be correct.
• Live reads over CSV exports. We chose to call Business Central and HubSpot at query time so answers reflect the current state. The tradeoff is rate-limit handling and dependency on source uptime.
• Reconciliation as a background loop, not on-demand. We chose to run entity linking continuously so leakage is detected before anyone asks. The tradeoff is more compute and a need for clear ownership of findings.
• Cited answers over raw model output. We chose to require the answer to point back to source records so finance teams can verify. The tradeoff is stricter agent contracts and more engineering on the citation layer.

Results

Operators ask questions in plain English and get cited answers across both systems in seconds. Revenue leakage that previously surfaced in quarterly Excel reviews now surfaces as it happens, with each finding linked back to the records in Business Central and HubSpot for resolution. Reporting work that used to require manual CSV joins now runs through the platform.

Project Highlights

• Natural-language querying across ERP and CRM with citations.
• Continuous revenue-leakage detection instead of quarterly review.
• Multi-agent orchestration with specialists per data surface.
• Entity linking that reconciles tenants to leases across mismatched identifiers.
• Read-only integration model that did not disrupt existing workflows.

Screens and User Flows

The most common flow is an operator asking a question, the router picking the right specialist, the specialist reading the source system, and the UI returning a cited answer. The same pattern supports follow-up questions because the router carries query context.

Future Growth Opportunities

• Add connectors for adjacent systems (accounting, maintenance ticketing, payment processors) using the same specialist pattern.
• Introduce write-back actions for approved workflows (create an invoice line, update a deal stage) behind explicit confirmations.
• Add scheduled briefings that send a daily or weekly digest of new leakage findings to finance owners.
• Expand reconciliation rules per portfolio so operators can codify their own definitions of expected revenue.
• Add role-based access so different teams (finance, operations, leasing) see scoped views and queries.