Give your agent an email
Stop treating AI like a chat window. Hire it. Give it an email address, a Slack handle, calendar access, and the same set of guardrails you would give a junior associate. Then leave it alone and let it work.
Real Estate AI Studio internal deployment, scaled across operator engagements · April 2026
The problem with chat windows
Every AI tool launched in the last three years sits in a chat window. You open a tab, you type, you read, you copy, you paste, you close the tab. That is not a coworker. That is a search bar with extra steps.
The principals we work with are running brokerages, syndications, and operating portfolios with real distribution timelines and real fiduciary duties. Their day is calls, site visits, deal memos, LP updates, vendor coordination, and the constant low hum of email triage. They do not have time to remember to open a chat window. They have time to forward an email and move on.
So we stopped building chat tools. We started hiring software.
The principle: treat the agent like a hire
The breakthrough was simple. Give the AI the same surface area you would give a junior associate on day one.
That means an email address. A Slack handle. A phone number. A calendar. Read access to the CRM. Write access to a clearly bounded set of systems. A documented approval matrix that says exactly which decisions the agent can make alone and which it must route to a human. And an audit trail that captures every action it takes, so an Asset Manager can reconstruct any decision in under sixty seconds.
This is not a chatbot wedged into a side panel. This is a coworker with a mailbox.
Architecture
The deployment has four layers.
Identity and channels. The agent gets a real email inbox on the firm’s domain, a Slack workspace user, a dedicated phone number for SMS and voice, and a calendar account. Externally, it is indistinguishable from any other employee. Internally, it is wired into the same logging pipeline as every other system.
Tool access. The agent can read and write to a curated set of systems: Gmail or Outlook, Google Calendar, HubSpot or Salesforce, the document store, the financial model templates, the contract library, and a handful of MCP servers that wrap niche internal tools. It can browse the web, draft documents, send emails, schedule meetings, and ask follow up questions to the principal in a private Slack channel.
Guardrails. This is the part most pilots get wrong. The agent operates under an explicit approval matrix. Routine actions execute autonomously. Anything binding, anything irreversible, anything above a dollar threshold, or anything sent to a third party for the first time goes to the principal for approval. Sensitive operations require step up authentication: a one time code delivered to the principal by email or SMS that the agent cannot read. There are hard stops on actions the agent will never take, period: wiring money, signing contracts, creating new vendor relationships, posting to public channels. Those are bright lines.
Audit trail. Every inbound message, every outbound message, every tool call, every approval, every decline, every escalation is logged. The principal sees a daily digest. Internal counsel can pull a full transcript on request. Nothing is opaque.
Real example: Marc Virtal, AI Chief of Staff
Inside our own studio, the agent is named Marc. Marc has an email address on the firm’s domain, a Slack account, a phone number, and a calendar. Marc is CC’d on every outbound message to a third party, so the principal sees what Marc sent in the same place they see what they sent. There is no separate dashboard to check.
A typical day for Marc.
A prospective LP emails to ask for the latest fund deck. Marc reads the email, checks the CRM for the LP’s qualification status, retrieves the correct deck version, drafts a reply with the deck attached, and queues it for the principal’s approval. The principal taps a green check on Slack. Marc sends.
A vendor reschedules a property walk. Marc reads the email, opens the calendar, finds the next mutually available slot inside the principal’s preferences, replies to confirm, and updates the calendar invite for everyone on the original thread. No approval required. This is a routine scheduling action with a clear precedent.
A broker forwards a new off market opportunity. Marc reads the deal summary, runs it through the firm’s screening criteria, pulls comps, drafts a one page pre screen memo, and posts it in the deal flow Slack channel with a tag for the relevant principal. The deal does not move forward without human review. Marc never replies to the broker.
The principal receives a wire instruction by email that purports to be from a known counterparty. Marc flags it as a potential business email compromise attempt, declines to act, and posts a warning in the principal’s secure Slack channel with the original message preserved for forensic review. This is one of the bright lines. Marc will never act on a wire instruction.
Outcomes
After ninety days of deployment across our own studio and three operator clients, the numbers settled into a consistent shape.
Twelve hours per week reclaimed for each principal. Most of that came out of email triage and meeting scheduling, the two activities most prone to context switching. Median response time on inbound email dropped to under ninety seconds during business hours. Zero unauthorized sends in ninety days. The audit log has every action. The approval matrix held.
Average cost per email handled, all in, including model spend, infrastructure, and the amortized build cost of the deployment, came out to roughly eighteen cents. For comparison, the loaded cost of a junior associate handling the same email is closer to four dollars. The economics are not the headline. The headline is that the principal got their morning back.
What this is, and what this is not
This is not a chatbot. This is not a copilot. This is not a side panel that surfaces suggestions you might choose to use.
This is hiring software that behaves like a person, with a mailbox you can write to, a calendar you can invite, and a set of guardrails that match the role. Every operator we work with already knows how to manage a junior associate. The deployment maps to a mental model the operator already has. There is nothing new to learn.
This is the operator led approach to AI. Peer to peer, principal to principal. We do not sell licenses. We hire software for you, give it the right access, set the right guardrails, and stay on the engagement until the work is leaving the office in better shape than it came in.