If you sit through enough AI demos right now, you will see the same phrase on almost every slide:
“Don’t worry, it’s human‑in‑the‑loop.”
Most of the time, that means “someone clicks approve at the end.” That is not human‑in‑the‑loop. That is a rubber stamp. In commercial real estate, where one wrong decision can lock in a decade of risk, “human‑in‑the‑loop” needs a much stricter definition. It has to answer three questions clearly:
- Who decides what “good” looks like?
- Where exactly does AI stop, and a person take over?
- Who owns the outcome when the deal goes sideways?
Where Humans Actually Need To Sit In The Loop
Think of a typical CRE workflow: sourcing → underwriting → diligence → IC → negotiation → asset management. A real human‑in‑the‑loop design puts people at four specific control points.
1. Humans define the objective, not the tool
AI should never be left to “optimize deals.” You decide:
- What return targets matter for this strategy
- How to trade IRR vs equity multiple vs cash yield
- How risk is priced across tenant credit, lease term, leverage, and capex
AI can:
- Run scenarios against that playbook
- Highlight where the current deal falls outside your own rules
If the tool is carrying an unspoken objective function (for example, “maximize IRR no matter what”), you are already out of the loop. You just have nicer charts.
2. Humans curate the inputs that carry judgment
AI is good in data. It is bad at knowing what should carry weight. You still need people to decide:
- Which rent comps are truly relevant
- How much to discount noisy or broker‑driven “market intel”
- Which third‑party data sources you are willing to trust
AI can:
- Normalize and compare
- Call out inconsistencies
- Flag outliers
But a person has to say, “These three inputs are credible enough to hang money on. Those four are background color.”
3. Humans own the decision and the explanation
On a real deal, “the loop” ends here:
- Someone signs the IC memo
- Someone recommends a final bid and structure
- Someone approves the PSA terms and the capital stack
AI is allowed to:
- Argue with you
- Show you patterns in your history
- Point out that your current assumptions sit at the edge of your own distribution
It is not allowed to be the reason. If your investment rationale starts to sound like “the model liked it,” you have crossed from human‑in‑the‑loop to “model‑in‑charge.” LPs and lenders will not accept that when things go wrong.
4. Humans close the feedback loop
After the deal closes or fails, a person has to ask:
- “What did our AI get right here?”
- “What did we override, and were we correct to do it?”
- “What did we ignore that we should not have?”
Then update:
- Assumption bands
- Checklists
- Prompts and workflows
AI does not fix itself. If you are not revisiting how it performed on actual deals, you do not have a loop. You have a one‑way pipe.
What This Looks Like In Real CRE Workflows
Let’s make this tangible.
Underwriting
AI can:
- Parse OMs, rent rolls, leases, and T‑12s
- Build a first‑pass model and IC summary
- Compare current assumptions to your own history
Humans must:
- Set the deal’s real objective (what win looks like)
- Decide which comps and assumptions survive into the base case
- Own the decision to greenlight, re‑price, or walk away
Due diligence
AI can:
- Read the full DD stack
- Extract every clause related to use, exclusives, co‑tenancy, SNDAs, environmental, easements
- Highlight where deal docs differ from your standard positions
Humans must:
- Name the single assumption that kills the equity if it is wrong
- Demand hard evidence on that line
- Decide whether the risk is truly priced into the deal
Negotiation
AI can:
- Compare mark‑ups to prior deals
- Suggest alternative packages that keep economics intact
- Track how the other side’s language and positions shift across drafts
Humans must:
- Set the actual walkaway point
- Choose which trade‑offs to offer when
- Decide when the negotiation pattern says “time to move on”
Asset management
AI can:
- Monitor operating data, leasing, and market signals
- Flag assets drifting off plan
- Suggest playbooks based on similar assets in your history
Humans must:
- Decide when to intervene
- Choose between options: refinance, sell, re‑tenant, invest more
- Own communication with investors and lenders
Fake “Human‑in‑the‑Loop” Patterns To Avoid
A few anti‑patterns show up again and again.
Rubber‑stamp review
AI does the work. A person skims and clicks approve to stay on schedule.
If your review step looks like that, it is not a control point. It is a liability.
Traffic‑light governance
Tools spit out green/yellow/red tags. IC treats “green” as safe and “red” as off‑limits.
The color is a starting point, not a verdict. You want the conversation: “Why is this red, and do we agree?”
Shadow AI use by juniors
Analysts quietly use public models for serious work. Partners assume “we are not really using AI yet.”
In that world, there is zero human‑in‑the‑loop at the decision level, and a lot of unmanaged risk at the data level.
Designing A Real Human‑in‑the‑Loop Framework
If you want to do this properly, draw a simple map for one or two workflows:
- Rows: Key steps (intake, underwriting, DD, IC, negotiation, AM)
- Columns: “AI does,” “Human must do,” “Human may do”
For each cell, answer in one sentence:
- What AI is allowed to produce
- What a human is required to review/decide
- How that review is captured (comment, sign‑off, checklist)
If you cannot point to where humans:
- Set objectives
- Curate inputs
- Make decisions
- Close the loop
Then you do not have human‑in‑the‑loop. You have humans near the loop, hoping that is close enough. That is why we have built CRE Agents.
Frequently Asked Questions About Human-In-The-Loop AI For CRE
LPs and lenders want to see that a human made the decision and can explain why. They are comfortable with AI handling data extraction, scenario modeling, and pattern recognition as long as there is a clear audit trail showing where AI stopped and a person took over. If your investment rationale ever sounds like the model recommended it, that is a problem in a loss review. The firms that handle this well can point to specific control points in their workflow where a named person reviewed inputs, approved assumptions, and signed off on the final recommendation.
The person who signed off on the decision. AI does not change fiduciary responsibility. If an analyst used AI to parse a lease and missed a material clause, the question in any dispute will be whether the review process was adequate, not whether the AI was accurate. That is why the framework in this post matters. Documented control points where a human reviewed, decided, and captured that review protect the firm in exactly the same way a signed checklist or IC vote does. Firms without that documentation are carrying risk they cannot see until something goes wrong.
Start with one workflow, not a firm-wide policy. Pick something concrete like underwriting or DD and sit down with the people who actually do the work. Draw three columns: what AI does, what a human must do, and what a human may do. Fill it in together in one meeting. That exercise almost always surfaces disagreements about where judgment actually lives, and resolving those disagreements is the real value. Once you have one workflow mapped, the pattern is easy to repeat across other processes.