An AI deal advisor that reads everything about a deal and answers in Slack.

The problem.

Nobody on a deal had the whole picture. The seller's situation was in the CRM, the negotiation was in a Slack thread, the counter-offer was in an email, the numbers were in a signed contract, the photos and the inspection were in a folder. Answering "what is happening with this property" meant re-reading all of it, and the legal or financial questions that came up mid-deal (what does this state allow, what is the maximum offer, how do we counter) had no fast, reliable answer.

What we built.

• An orchestrator agent with tools. One mention in a deal channel triggers a single agent that decides which of its tools to call (Slack history and search, CRM records, Gmail search and threads, contract summaries, Drive documents, a bounded web search) and answers with citations, within a fixed number of tool iterations.
• Three hats. The agent switches role by question: legal answers are state-specific and carry a verify-with-counsel caveat; financial answers cover maximum offer, after-repair value, repairs, fees, projected profit and counters; sales answers cover negotiation, seller psychology and objection handling.
• Context that builds itself. On the first mention in a new deal channel, the agent queues a full build of the deal's context from Slack, the CRM, contracts and documents, then answers the original question in the same thread once it is ready. Anyone can ask for a refresh.
• Vision. Property photos, inspection shots, document screenshots and whiteboard sketches uploaded to Slack are read directly: numbers and dates transcribed, condition described.
• Deliverables, drafted first. Google Docs, Sheets and branded PDFs created in the deal's Drive folder, always shown in Slack for approval before the tool is called. Email replies are drafted into the rep's Drafts folder; the agent never sends.
• A deal log that persists. The agent reads, creates and updates the channel canvas, the long-form record that survives between threads.
• Engineering for cost and truth. Local-first email search from a sidecar table before falling back to the Gmail API, prompt caching across turns, a hard cap on tool iterations, and a rule that links are cited verbatim from the record or not at all.

How it runs.

1. A rep mentions Scout in the deal channel: "summarise the emails with the seller" or "what can we counter at".
2. The agent loads the deal's pre-built context, calls the tools it needs, and answers in Slack markdown with links to the underlying records.
3. If the answer is a document, it drafts it in the thread first. On approval it creates the Doc, Sheet or PDF in the deal folder and posts it back.
4. If the answer is an email, it lands in the rep's Drafts. A human reads it and clicks send.

What changed.

• One question, the whole context. Nobody re-reads five systems to answer "where are we on this one".
• The agent drafts, a human sends. Speed without an autonomous system emailing sellers.
• New channel, ready context. Adding the bot to a deal is enough; the homework does itself.
• Answers cite the record, so a wrong answer is visible and a right one is checkable.

The stack.

• Claude
• Slack
• Podio
• Gmail
• BoldSign
• Google Drive
• Postgres
• Python

Want this for your team?

Scout is what our Custom AI Engineering offer looks like in production: an agentic system shaped around one company's workflow, with the guardrails that make it safe to rely on. If your deals, cases or projects are scattered across five tools, the same pattern applies.