Deals in the Headlines, Risk in the Book

Impact Metrics

• 2 AI systems live (Deals in, risk out)
• 12 News sources (Watched around the clock)
• 766 Live deals (From 651 distinct developers)
• $943.6B Development value (Catalogued from stated figures)
• 3,495 Policies (Across 196 clients, read on demand)
• 0 Invented facts (Missing detail stays blank)

The Challenge

Deals Break in the News, Not the CRM

New developments are announced across dozens of trade publications. By the time a producer reads about a project, a competitor may already be talking to the developer.

Reading Is Not Sourcing

Turning an article into a lead meant copying developer, location, value and stage into a spreadsheet by hand. Slow, inconsistent and easy to skip on a busy week.

The Book Lives in Exports

The policy book came out of the agency management system as a flat spreadsheet. Every question about growth or concentration meant rebuilding the analysis in Excel.

Renewal Risk Hides in the Rows

Past-due expirations, non-renewals and lost clients were buried among thousands of lines, so premium at risk only became visible after it was gone.

AI-Powered Solutions

Problem: Producers spent hours scanning trade press and still missed projects.

Solution: A scheduled pipeline polls 12 real estate news feeds. Azure OpenAI reads every new article and returns a structured lead: developer, project, location, value, asset class, stage, contractor and more.

Problem: An AI that guesses is worse than no AI for a sales team.

Solution: The model fills a fixed, code-controlled schema and leaves any field blank that the article does not state. Every lead carries a confidence score, and the same deal reported by two outlets becomes one lead.

Problem: A flat export cannot show how the book is changing.

Solution: Upload the export and it is archived, parsed and kept as a dated snapshot. The dashboard is ready in seconds, and the history powers trends over time, year over year and year to date.

Problem: Producers should see their own book; leadership needs the whole firm.

Solution: Access is scoped per producer in the API, not the interface. Administrators see the firm, producers see their own books, and saved views stay private. Automated tests prove one producer cannot read another's data.

Product Walkthrough

Screens from the live production systems, captured on real data. Client branding, client names, premium figures and user details are blurred. Deal data comes from public news coverage.

Deal Intelligence

For sourcing teams and producers. Watches the news, reads every article with AI and hands the team a searchable lead book every day.

Book of Business

For producers and leadership. Turns the agency management system export into a live, trending view of the whole book.

The pipeline at a glance

Totals, a daily lead count, and breakdowns by asset class, project stage, state and developer. Every chart responds to the same filters.

• 766 leads, 60 of them new this week
• Average extraction confidence of 0.90
• Filter by asset class, stage, state, source or confidence

Dashboard. Client logo blurred.

A lead book producers can work from

Every project the AI found, with developer, location, stated value, asset class, stage, article date and confidence. Exports match exactly what is on screen.

• One-click Excel of the filtered view
• PowerPoint prospect booklet of top deals
• Blank cells where the article was silent

Leads. Client logo blurred.

Sources the team controls

The team adds or removes news feeds without a developer. A new URL is fetched and parsed first, so only a real, live feed gets in. Internal and unreachable addresses are rejected.

• 12 of 12 feeds enabled
• Client-supplied and new sources tracked separately

Sources. Client logo blurred.

Instructions in plain English

The extraction prompt is editable on screen and takes effect on the next run. The output fields are locked in code, so an edit can sharpen what gets captured but cannot break the pipeline.

• 18 locked output fields
• Reset to default in one click

Prompt. Client logo blurred.

The whole book in seconds

Policies, programs, clients, billed premium and revenue up top, then every policy grouped by renewal urgency, from past due to more than a year out.

• 3,495 policies across 196 clients
• 1,170 programs grouped automatically
• Non-renewal flags surfaced at the top

Dashboard. Logo, book name, user and premium figures blurred.

Where the premium sits

Ranked breakdowns by client and line of business show concentration at a glance. The grouping and metric for each panel are configurable and can be saved as named views.

Business mix. Client names and amounts blurred.

Trends a spreadsheet cannot show

Premium or revenue by month across the full history, each fiscal year on one axis, and non-renewals with lost clients and the premium at risk.

• Fiscal or calendar year, premium or revenue
• Year-to-date clipping for like-for-like comparisons
• Trend settings saved with each view

Trending. Currency axes and premium at risk blurred.

Down to the program

The table behind every chart, filterable by client, line, department, carrier, status and non-renewal flag, with an Excel export of exactly what is filtered.

Programs. Client names and amounts blurred.

What We Built

• Automated news sourcing: Scheduled polling of 12 RSS and Atom feeds, with raw articles archived to cloud storage.
• AI lead extraction: Azure OpenAI structured outputs fill 18 fixed fields per article, with a confidence score per lead.
• Lead dashboard and book: Analytics and a searchable, filterable table, de-duplicated on developer, project and state.
• Excel and PowerPoint exports: The filtered lead list as a spreadsheet, or a prospect booklet of top deals ready to share.
• Self-service sources: Add a feed by URL. It is fetched and validated before it can join the pipeline.
• Editable AI instructions: Plain-English prompt editing with a locked output schema, so edits cannot break extraction.
• Upload and snapshot: Each export is archived, parsed and stored as a dated snapshot. Nothing is overwritten.
• Expirations by urgency: Past due, 30, 60, 90 days and beyond, in one bar, so renewal priorities are obvious.
• Trending and non-renewals: Over time, year over year and year to date, plus lost clients and premium at risk by month.
• Configurable saved views: Choose groupings, thresholds and trend settings, then save them as private named views.
• Role-based access: Producers see their own books, administrators see the firm, enforced in the API.
• One-click Excel: Export the current view or the full trend series, ready for the team's own workflow.

Development Timeline

• Early Jul Deal pipeline: Feed polling, AI extraction, Postgres storage, lead dashboard and PowerPoint booklet.
• Late Jul Cloud and book v1: Containerized deploy to Azure. Book of Business with uploads, snapshots and the core dashboard.
• Aug Hardening: Rate limits, export caps, safe migrations, graceful shutdown and a CI test gate on both systems.
• Late Aug Self-service and trends: Editable prompt, validated sources, trending, year over year, non-renewals and saved views.
• Sep Live in production: Both systems live on real data for producers and leadership.

Platform Capabilities

AI Extraction

• Azure OpenAI structured outputs
• 18-field code-locked schema
• Blank, never guessed
• Per-lead confidence score
• Cross-outlet de-duplication

Lead Pipeline

• 12 curated news feeds
• Scheduled, unattended runs
• Feed validation on add
• Raw article archive
• Excel and PPTX delivery

Book Analytics

• Dated snapshot history
• Fiscal or calendar years
• Year over year and YTD
• Non-renewal tracking
• Private saved views

Platform and Security

• Azure Container Apps and Postgres
• scrypt passwords, session auth
• Login throttling, Helmet CSP
• Per-producer access scoping
• 118 automated tests in CI

Sourcing teams start the day with new deals already read and logged. Leadership opens the whole book in seconds.

Both systems run on the brokerage's Azure cloud, on real data, with every lead and every number traceable to its source.