How AI works in Hale.

Hale uses AI to do tedious reading and writing — pulling structured fields out of listing pages, drafting summaries, and writing first-draft broker emails. Every AI action is logged, budget-capped, and gated by a human before it leaves the system. Nothing is sent, purchased, or decided automatically. If a human doesn't click a button, nothing moves.

What the AI actually does

1. Extraction — reading listings

When a listing source (a LoopNet page, a broker PDF, an email) arrives, Hale hands the raw text to Claude and asks it to pull out fields: address, property type, size, clear height, asking price, broker contact. Each extraction includes a confidence score (0–100%); anything under 75% usually means a field got guessed poorly — click into the property to cross-check against the original listing.

The AI never creates a property Hale wasn't told about, never fabricates fields the source didn't mention, and never sees data from other tenants.

2. Scoring — ranking against your criteria

For each property, Hale computes a 0–100 score against each active search criteria. The math is a weighted sum of factor scores (size, price, location, clear height, etc.). The weights are your weights — editable in the Criteria page. Claude is only used to explain a score when you ask; the ranking itself is deterministic math. You can see the breakdown on every property.

3. Outreach drafting — first-draft broker emails

On a property detail page, click "Draft Email." Claude writes a short outreach message in the tone you pick (professional / direct / warm / follow-up), referencing the property specifics. The AI never sends the email. The draft lives in Hale; you read it, edit it, copy it to clipboard, and send from your email account. That's deliberate — your reputation stays tied to your address, not a bot's.

4. Summaries — one-liner "why this matters" notes

On a property, Claude writes a 2-3 sentence summary. These are aids, not facts. If a summary claims "excellent rail access" and the listing doesn't mention rail, flag it — it's rare but not zero.

What AI is never used for

• Pricing decisions. Hale never suggests an offer price. Asking price + advertised cap rate are shown; everything past that is your judgement.
• Sending outreach. Every email is drafted; you send manually from your own address.
• Closing deals. Hale has no DocuSign, offer platform, or banking integrations. It ends at "I wrote a draft for you."
• Selling or sharing your data. Your properties, criteria, scores, and notes are visible only to your tenant. No ML training set includes your data.
• Automated bulk email. No "send to all brokers" button. Every email is drafted and sent one-at-a-time, by a human.

Which AI, who pays

Model: Anthropic's Claude family. Default is Haiku 4.5 (cheap, fast) for extraction and score explanations; Sonnet 4.6 for outreach drafts where quality matters more.

API key: Each Hale tenant has its own Anthropic API key. It is not shared with any other customer. One runaway loop in another tenant cannot affect your budget, and vice versa.

Cost limit: Each tenant has a per-day spend ceiling. If a bug causes repeated calls, the ceiling trips and subsequent calls are rejected. The worst-case damage from a buggy prompt is today's limit, not the Anthropic account balance.

The audit trail

Every AI API call is logged to an immutable record in Hale's database:

• When the call happened (timestamp)
• What model was used (Haiku vs Sonnet)
• What task it was (extraction / scoring-explain / outreach / summary)
• How many tokens in and out
• How long it took
• Exact cost estimate in USD
• Success or failure (and the error message if it failed)

Admins can view this log at /ai-logs and export it as CSV. If you ever want to know "why did we spend $X on AI in April?" — the answer is one download.

What human judgment still owns

The tool is fast and thorough, but it doesn't replace the judgments that actually make a deal:

• Is this broker honest? Hale has no opinion. You have history with them.
• Is this neighborhood trending? Hale has no opinion. You've driven it.
• Is the listing price realistic? Hale shows asking vs comparables. You negotiate.
• Does this fit our pipeline and cash position? Hale doesn't know your pipeline. Never trust a tool that pretends to.

Think of Hale as a very good research assistant: it finds things, reads things, drafts things. It doesn't decide things.