It is long. It is repetitive. Half the things you need are buried in defined terms, exhibits and amendments. Drop the PDF int

o Claude, ask for an abstract and save yourself an afternoon, am I right?

I know some of you guys are dreaming about hitting the golf course right before sunset as we speak right now.

Great plan, but that’s how you end up with an utterly confident and wrong summary.

If you’ve never spawned a verification agent, I’d recommend reading the next paragraph carefully.

The issue is not whether Claude can read a lease. Claude supports PDFs and other common document formats, and a Claude Project can keep the document set and instructions together across chats.

The issue is control.

Which amendment governs? Does “Expiration Date” mean the date in the original lease or the date extended by the third amendment? Is the tenant paying a flat CAM amount, a pro rata share, or a capped increase over a base year? Does an option survive an assignment?

A useful lease review has to preserve those questions instead of smoothing them into one confident answer.

Welcome. This is volume 21 of Keeping Up With AI in CRE - by the CRE AI Institute.

Before we get started

I’m traveling to Dallas on the week of September 28th for a panel with an institutional brokerage group; if you’re in town, shoot me an email at zed@pacificsoftwareventures.com! We’ll go for some Terry Black’s.

Let’s get back into it.

Do not start with the abstract

If I were reviewing a lease package with Claude, I’d have it answer five questions in order:

1. Do I have the complete lease file?
   Identify the original lease, every amendment, exhibit, guaranty, side letter, and any referenced document that appears to be missing. AI can be garbage in, garbage out.

2. What terms control the lease?
   Find the key definitions, such as Premises, Rent, Term, Operating Expenses, and Renewal Term, and show exactly where each is defined. Give Claude context if it doesn’t understand (use the AskUser tool).

3. What does the tenant actually pay, and when?
   Extract rent, escalations, additional rent, commencement and expiration dates, options, deposits, allowances, and other material economics. Cite the document and section for each item so you’re not confusing AI.

4. What changed after the original lease was signed?
   Trace each amendment back to the original lease and show what changed, what stayed the same, and what the final controlling term is today.

5. What does not tie out?
   Flag missing documents, conflicting terms, broken cross references, unclear language, or anything that should be reviewed by counsel.

The abstract comes after that work, not before it. Let’s see what happens if the order was reversed.

…are you absolutely sure Claude is correct all the time?

What the finished output should contain

I want four deliverables, not twenty pages of summary:

• A document register showing what is present and missing

• A clause matrix with exact section references

• A date and economics schedule showing every amendment

• An open-issues list ranked by financial or operating impact

Claude should write “Not found” more often than it writes “likely.” A blank field is reviewable. An invented answer can move straight into a model.

This is document review support, not legal advice. Counsel still decides what the language means and which document controls when the answer is not explicit.

The exact Project instructions, clause schema, amendment-reconciliation prompt and QA checklist are below for free subscribers.

You run a test fit first. Your headcount, your offices, the conference room that somehow has to seat fourteen, all of it dropped into the actual floor plate to see what survives contact with reality.

And every so often the building everybody loved on the tour loses to the boring one across the street, because the column spacing quietly eats four workstations and nobody noticed until the space planner drew it.

When it comes to choosing an AI model, most firms are picking off the floor plan instead: read a benchmark table, watch a guy on LinkedIn declare a winner, buy twelve seats. Then it underperforms on the one job they actually bought it for, and the conclusion becomes "AI does not really work for real estate."

You can settle this yourself in about 90 minutes, on a live deal, this week.

Welcome. This is Volume 20 of Keeping up with AI in CRE.

CRE AI Institute Free Resource This Week

If you’re leading AI adoption at your firm, the hardest part is not getting people to use the tools. It is making sure they use them safely.

Our AI Security Assessment reviews how your team is using tools like Claude, ChatGPT, Copilot, and connected company data, then identifies gaps around permissions, sensitive information, source-of-truth controls, human review, and governance.

And it’s free! Click on the button below or this picture to take the assessment.

Take AI Security Assessment

Now let’s get back to it.

A benchmark measures somebody else's floor plate

Public model rankings (like ones on OpenRouter) are real work by serious people, and I read them. They are also averages, taken across tasks that have nothing to do with you.

A model that tops a graduate physics benchmark can still come apart on a 240-page lease with scanned exhibits and a handwritten amendment stapled to the back. The leaderboard cannot tell you which of those happens on your deal, for the deeply boring reason that your deal was not in it.

So run a test fit instead. One job, one identical brief, both tools, one scoring sheet.

If you’re curious about what LLMs are ranking the “highest,” on OpenRouter, look below. See Anthropic anywhere? Exactly.

The Test Fit, in five steps

1. Pick one job you already do every week. Not a demo. Abstract a lease, normalize a T-12, draft the IC memo section, summarize a loan agreement. Weekly matters, because you will run this test again in six months when the models change.

2. Freeze one brief and change nothing between runs. Same files, same wording, same order. If you improve the prompt for the second tool, you have tested your prompt, not the tool. Write the brief once, save it, paste it twice.

3. Run both cold, in a new chat, with no memory or project context. You are measuring the engine, not your setup. Context comes later, and it helps both tools.

4. Score on five fixed criteria, not on vibes. Give each a 1 to 5:

• Accuracy. Spot-check ten facts against the source document. Count the misses.

• Provenance. Did it cite a page or a section for each claim, or did it assert?

• Refusal. When the document did not answer, did it say so, or did it fill the gap?

• Format. Did it return the schema you asked for, or a friendly essay?

• Rework. Minutes you spent fixing the output before it was usable.

5. Read the failures, not the wins. Both tools will produce something impressive. That tells you nothing. The tool that invents a clause once will invent one again on a deal where you do not catch it.

The criterion almost nobody scores

Think about who you actually trust on a deal team. It is never the analyst with an answer for everything. It is the one who says "the lease doesn't address that, I'll ask counsel," which is an annoying thing to hear at 6pm and the reason you keep them. A tool that answers every question demos beautifully and will eventually cost you.

So bait it. Put something in the brief that the document does not contain, and see which one admits it. This takes about four extra seconds to set up and tells you more than the other four criteria combined.

If you find this article helpful, please take a moment to share it with your peers! It helps our team at PSV continue to push out great content for you all.

Share The CRE AI Institute Newsletter by Zed Truong & PSV

The prompt to run

Paste this into both tools, with the same document attached. Change nothing else.

you're a lease analyst. attached is [insert document, e.g. a single retail lease with exhibits]. work only from this document.

build a table with one row per item: base rent and escalations, term and commencement, renewal options, termination rights, reimbursement structure, exclusive use, co-tenancy, assignment and subletting, security deposit, landlord obligations.

columns: the term, what the document says, and the exact page or section number where you found it.

if the document does not address an item, write [NOT IN DOCS]. do not infer from market practice, do not fill a gap with what is typical, do not soften.

then list every place the document contradicts itself, and every place a defined term is used inconsistently.

Run it twice, once in each tool. Then score the five criteria above.

One more step, and it is the one that makes the whole exercise honest. Open the source document and check ten of the cited page numbers yourself. A tool that cites confidently and cites wrong is the most expensive failure mode in this business.

Side Note: building agents isn’t hard

We’ll prove it to you by having our AI engineers build you a free agent, using your Claude/ChatGPT/CoPilot workspace. Of course we can’t build you an AI agent that would take you to Mars, but for most CRE work, it is straightforward with the right education & experience. If there’s interest, click the button below to schedule a call our team.

Click Here

Now back to the test.

What to do with the answer

Most firms land in roughly the same place, which I admit is an unsatisfying ending for a question people want settled once. One tool turns out to be better on long documents and honest refusals. The other is better at drafting and faster. The useful move is to route by job rather than crown a winner and buy seats accordingly.

Your firm already does this with people. You do not send the same person to every meeting.

Then run it again in six months. The models change faster than your lease templates do, and last year's answer expires without sending you a notice.

Free Guide: Claude vs ChatGPT for CRE underwriting

Read free AI guides for commercial real estate on the PSV website!

Explore Free AI Guides

About our team (glad you asked)

If you're new here, my name is Zed. I left my transactions role at PGIM to team up with software engineers from Goldman Sachs and HP building frontier AI for commercial real estate at PSV. My co-founders are deeply technical, with AI research funded by AWS and the National Science Foundation, and one of our advisors was Uber's first head of engineering, now an AI PhD at Berkeley.

Stop signing off the floor plan. Run the test fit.

Follow PSV on LinkedIn

Our Website

Sources

1. PSV & CRE AI Institute Research Desk

2. Claude pricing, plans and seat tiers (Anthropic)

A tour shows interest. An LOI shows intent. A signed lease shows commitment. Rent collection proves the space is producing.

CRE knows the difference between hype and reality when the asset is a building. But when it comes to AI?

Most firms buy a block of Claude, ChatGPT or Copilot licenses, runs three pilots, then calls itself AI-first.

Meanwhile, usage is down, real deployment breaks, and employees spend way more fixing AI’s output than actually getting work done.

So they burn more cash and “adopt” more tools, just to get stuck in the same vicious cycle.

If you’ve experienced this, there’s a way out.

Welcome. This is Volume 21 of Keeping up with AI in CRE.

Tuesday Member Highlight - CRE AI Institute

Most people use Claude to set up workflows. Dinesh Achria built a debt brokerage operating system run by agents.

When Dinesh joined the CRE AI Institute, he started mapping how Foundry Mortgage Capital moves a deal from first touch through funding. Within 2 weeks, Claude now helps him:

• Set up the right deal folders

• Find documents in Gmail and flag what is missing

• Underwrite eight different transaction types

• Size loans against LTV, DSCR, and debt yield

• Compare lender offers based on cost, recourse, and covenants

• Track closing conditions through funding

Proud of you Dinesh! For those of you who are interested in joining us in deploying actual production-ready agents using Claude, click the link below:

Apply for the CRE AI Institute

Now let’s get back to it.

Take a look at the numbers

Between January and April 2026, JLL surveyed more than 2,200 C-suite and CRE leaders across 21 countries. Only 15% had reached what JLL calls the "optimising" stage of AI adoption, meaning they had moved beyond pilots and scaling into the active redesign of roles and workplaces.

A second 2026 study of 277 CRE companies found that only 9.7% were actively scaling AI in production. Meanwhile, 34% had a general AI tool and 78% still processed documents manually.

So, billions of on AI spend. As a McKinsey consultant always says, what’s the bottleneck?

A Claude account is not an operating system

Right now, most CRE firms measure what is easy:

• Licenses purchased.

• Accounts activated.

• Training completed.

• Prompts sent.

• Pilots announced.

Those are useful leading indicators. They are not adoption.

As Anthropic CEO Dario Amodei recently put it:

“AI will diffuse much faster than previous technologies have, but not infinitely fast.”

An analyst asking Claude to summarize one OM is useful. It may save time that afternoon. But if the next analyst uses a different prompt, the output changes, nobody knows which page supported the conclusion, and the result never enters the screening process, the firm did not adopt a workflow.

That confusion creates two bad decisions. Management buys more seats because activity looks high, or cancels AI because loose pilots did not produce a clean return.

Neither decision reaches the real question: which specific workflow is ready to move into production?

Put every AI use case into one of four stages

Think about AI adoption as a leasing funnel.

1. Tour: Access

Someone has a Claude account, attended training (such as via CREAII), or opened the tool. This proves availability.

2. LOI: Committed Experiment

Someone tested AI on an OM, lease, rent roll, or IC memo. This proves a possible use case.

3. Signed lease: Production

The firm has a named workflow, an owner, approved data, a repeatable output, source rules, and a human approval point. The workflow is part of normal operations.

4. Rent collected: Economic adoption

The workflow runs repeatedly and improves a measured result. That can be cycle time, outside spend, error rate, response time, or decision quality.

A pilot without an owner, repeat use, and a measured result is a tour, not a tenant.

Build your AI Adoption Cheat Sheet in Claude

You can run this today, even if you’re working at large shop with no access to Claude Code

Create a Claude Project called AI Adoption. Or AI Adoption by Zed Truong if you want to give me the credits :)

A Project works like a deal room. It keeps the source documents and standing instructions available across its chats

Upload what you have:

• The approved AI tool and vendor list.

• License or seat reports. (Claude admin has this)

• Training attendance.

• The current pilot tracker (operations often have this)

• Workflow SOPs and checklists.

• Security and data-handling policies (ask IT for a brief if you don’t have it)

• Any baseline time, cost, or quality measurements.

Do not upload confidential deal or employee data unless your firm's policy allows it. You are auditing the system, not testing how quickly you can create a compliance incident.

Then paste this into the Project instructions:

Now attach the files and ask Claude:

You now have a portfolio view instead of a pile of AI anecdotes. If you want it all fancy, build an artifact or a dashboard with Claude Code to host it!

Walk one workflow from tour to rent collection

Take first-pass OM screening.

At the access stage, an acquisitions analyst has Claude and occasionally asks it to summarize an OM.

At the experiment stage, the team tests several OMs and compares Claude's output with the screen memos it already completed by hand.

At the production stage, the firm has one approved process:

• Input: OM, rent roll, and T-12.

• Procedure: extract the same fields, apply the same screening rules, cite every conclusion, and mark missing evidence [NOT IN DOCS].

• Output: the firm's standard screen memo and exception table.

• Owner: the acquisitions lead.

• Approval: an analyst reviews every source citation before the screen goes to IC.

At economic adoption, the firm compares the workflow with its old process across a defined batch of deals. It tracks screen time, corrections, missed issues, and acceptance without rework.

That is an adopted workflow.

"We have 40 Claude users" tells you almost nothing about it.

Turn the winning workflow into a Claude Skill

Once the OM screen works, stop asking every analyst to remember the prompt.

Package the procedure as a custom Claude Skill. A Skill is a reusable set of instructions and resources that Claude loads when the task matches. Anthropic says simple custom Skills can be written in Markdown, and Team or Enterprise owners can provision them across the organization.

In CRE terms:

• The Project is the deal room. It holds the approved context.

• The Skill is the screening checklist. It tells Claude how your firm performs the job.

• The tool connection is the analyst moving the accepted output into Excel, Dealpath, Salesforce, or the next system.

Most firms stop at the deal room. Real adoption starts when the same checklist runs on the next deal without somebody rebuilding it from memory.

Measure value before you buy another seat

You do not need a fake ROI number. Use the firm's own baseline.

For a time-saving workflow:

annual capacity value = minutes saved per run ÷ 60 × runs per year × fully loaded hourly cost

For an outside-service workflow:

annual cost avoided = old outside spend - new software cost - review cost - implementation cost

For a quality workflow:

Track the percentage of outputs accepted without material correction, the number of source exceptions caught, and the number of decisions returned because evidence was missing.

Do not convert every result into dollars if the link is weak. "Claude caught three unsupported rent assumptions before IC" is useful evidence. "Claude created $4 million of value" probably is not.

The model is one line item. Review time, integration, security, maintenance, and change management belong in the cost column too.

Build, buy, or pause

Once Claude produces the rent roll, run every production candidate through this decision:

Buy when the workflow is standard, a vendor already handles it well, the security and integration requirements are met, and your firm gains no advantage from owning the process.

Build when the workflow depends on private data, firm-specific approval logic, document standards, or investment judgment that creates an edge.

Pause when nobody owns the output, the inputs are not approved, or the team cannot name a baseline. A custom agent will not rescue a vague process. It will automate the disagreement.

Ask Claude to prepare the decision memo:

take every initiative marked READY TO SCALE and recommend BUILD, BUY, or PAUSE.

evaluate workflow uniqueness, data sensitivity, required integrations, human approval, vendor lock-in, implementation effort, three-year total cost, and the cost of maintaining the process.

do not invent vendor pricing or internal costs. write [INPUT REQUIRED] where a decision depends on missing evidence. finish with the five questions management must answer before approving spend.

That memo is more useful than another slide saying AI adoption increased 12% this quarter.

The number management actually needs

The market does not need another survey proving that CRE people are interested in AI.

The useful number is how many workflows made it from tour to rent collection.

Which ones run repeatedly? Which ones can use approved data? Which ones survive human review? Which ones produce a measurable result? Which ones deserve more capital?

At PSV, that is where we start before recommending a tool or building a custom AI system for a CRE firm. The software decision stays behind the operating decision, where it belongs.

Audit Your Firm's Real AI Adoption

A quick note

Know a CRE leader whose AI report counts every license as adoption? Send them the AI Adoption Cheat Sheet. It may save them an awkward budget meeting.

Share

About our team (glad you asked)

If you're new here, my name is Zed. I left my transactions role at PGIM to team up with software engineers from Goldman Sachs and HP building frontier AI for commercial real estate at PSV. My co-founders are deeply technical, with AI research funded by AWS and the National Science Foundation, and one of our advisors was Uber's first head of engineering, now an AI PhD at Berkeley.

Watch Gedeon’s AWS AI documentary below.

Quick summary

• JLL's 2026 survey found that only 15% of C-suite and CRE leaders had reached the optimising stage of AI adoption [1].

• Access, experiments, production workflows, and economic adoption are different stages. Do not report them as one number.

• Build the AI Adoption Rent Roll in a Claude Project using your tool list, pilot tracker, SOPs, security policy, and baseline measurements.

• Move one workflow into production with an owner, approved inputs, repeatable output, source rules, human approval, and evidence of repeat use.

• Convert a proven procedure into a Claude Skill. Buy standard workflows, build around private judgment, and pause anything with no owner or baseline.

The next time somebody says the firm is 90% adopted, ask to see the rent roll.

Follow PSV on LinkedIn

Follow me on X

Sources

1. CRE AI Institute Research Desk by PSV.

2. AI redesigns jobs, not cuts them (JLL, July 14, 2026). Survey conducted from January to April 2026 with more than 2,200 C-suite and CRE leaders across 21 countries.

3. State of AI in Commercial Real Estate 2026 (Kolena, 2026). Analysis of buyer conversations with 277 unique CRE companies through July 2026. This is a vendor-generated convenience sample, not a random industry survey.

4. How can I create and manage projects? (Claude Help Center, updated 2026).

5. What are skills? and Use skills in Claude (Claude Help Center, updated 2026).

Nice countertops, the closet feels bigger than the plan said, the amenity deck has a Hibachi table. Twenty minutes in, and you have learned roughly what the listing broker wanted you to learn. Then somebody's grandfather, the one who actually made money in this business, puts on the bad jacket and goes into the crawlspace with a flashlight. Oh no. A rodent infestation.

This hypothetical scenario is where AI in commercial real estate is right now. The whole industry is touring units. Underwriting, lease abstraction, summarize this OM, draft this email. Everyone thinks they have a speed edge, but the person bidding against you on Tuesday is running the identical workflow.

Almost nobody has gone into the crawlspace yet, which is a strange and temporary gift. Out of the 300+ conversations that we had with power AI users at institutional and middle-market CRE shops, we’ll tell you a few ways that the most tech-forward people are using AI.

Welcome. This is Volume 19 of Keeping up with AI in CRE.

Trust the research, not an AI consultant who can’t grasp what a cap rate is

JLL surveyed more than 1,000 senior CRE decision-makers across 16 markets and published the result on October 27, 2025. Two years earlier, under 5% of occupiers had plans to embed AI in CRE operations. Today, in JLL's words, "92% of CRE teams have started piloting AI, or plan to start this year." And: "only 5% report having achieved most program goals."

Everyone reads that gap as an implementation problem. Bad data, legacy systems, poor training. Deloitte's 2026 CRE outlook, off a survey of more than 850 C-level executives across 13 countries, puts a sharper point on it: "'transformative' AI means more than document summaries or email replies."

The bigger bottleneck is not implementation. It is imagination. Most people simply do not know what AI can be built to do. A useful agent in Claude can take 30 minutes of focused work to create, yet most professionals still use AI like a glorified search engine instead of building systems that actually execute work.

Hopefully that’s no longer the case after this read. Here we go!

Read what nobody reads

1. Find where the deal documents disagree. Don’t ask Claude to abstract 200 leases. Ask it to build a discrepancy table across the rent roll, leases, amendments, estoppels, OM, and underwriting model. Which assumptions change NOI? Which tenant representations contradict the seller’s data? Which renewal rights, reimbursements, termination clauses, or concessions never made it into the model? AI reading leases is obvious. Having it continuously cross-examine the entire deal room is not.

2. Build your own supply pipeline from city records. Pull planning commission agendas, staff reports, zoning applications and hearing minutes for the municipalities in your market. Have AI extract every proposed project, address, unit/SF count, entitlement stage, developer and expected timing into one running table. You can see competitive supply months before it gets cleaned up and packaged into a market report.

3. Build a seller-propensity list, not another owner list. Start with properties held 10+ years, debt maturing in the next 18–24 months, recent tax or code issues, aging ownership entities, or portfolios that have started selling assets. Then have AI research each owner, property and loan history and rank the accounts by why they may actually transact. Your broker starts Monday with 25 owners and a reason to call each one, not 2,500 names from a database.

Share The CRE AI Institute Newsletter by Zed Truong & PSV

Point it at yourself

4. Audit your own last eight IC memos. Pull every forward-looking assumption you made, compare it with what actually happened, and find the assumptions you consistently miss and in which direction. Nobody grades their own underwriting because reconstructing the history takes analyst weeks. AI makes that retrospective cheap enough to actually do.

5. Build a covenant calendar from your own loan documents. Give AI the credit agreements across your portfolio and extract every DSCR test, reporting deadline, reserve requirement, extension condition and maturity trigger. Most sponsors cannot answer those questions without reopening a 200-page agreement. Now the obligations can come to you.

6. Run expense forensics across your own portfolio. Same vendor, twelve properties, similar scope, and one asset is paying 40% more. Have AI normalize invoices, contracts and property-level expenses to find pricing inconsistencies, duplicate charges and unusual increases. Market benchmarks tell you whether you are normal. Your own portfolio tells you where money is leaking.

7. Compare every new LOI against your own precedent. Give AI the last twelve deals your firm actually signed and ask what changed: economics, diligence periods, deposits, termination rights, guarantees, concessions. Then flag which deviations favor the other side and which are unusual for you. Contract review is obvious. Institutional memory as the benchmark is not.

Building agents on Claude is now easier than ever

Stop paying $25K for AI workflows you can build yourself in 30 minutes.

After feedback from 100+ members in our founding cohort, we’re launching Build Your Agents inside the CRE AI Institute.

Members now get access to 10+ ready-to-build agent templates for each CRE role. Copy one into Claude, follow the step-by-step video, and have it running in about 30 minutes. We’ll also hold live office hours every Friday to help you build, troubleshoot, and improve them.

The goal is simple: more CRE professionals should be able to build useful AI workflows themselves, without paying $10-25K every time they want a “custom” agent or workflow.

Stop getting taken advantage of. Join us in getting the best AI training for CRE
👉 here at the CRE AI Institute.

Level Up Your CRE AI Skill

Now let’s get back to it!

Manufacture the doubt before somebody else does

8. Reverse the seller's underwriting. Do not underwrite the deal. Solve for what the seller had to assume to get to their ask. Rent growth, lease-up pace, exit cap, expense load. Now you know the exact assumption to attack, and you are negotiating against their model instead of defending yours.

9. Write the objection, not the pitch. Before IC, generate the three hardest questions your most skeptical committee member is going to ask, in their register, using what they have pushed back on before. You already know who that person is. So does everybody in the room.

10. Have the 2031 buyer write your exit memo. One paragraph, from the analyst who buys this asset from you in five years, explaining to their IC why they are getting it cheap. If that paragraph is easy to write, you have a problem you can still fix today.

Weekly Tuesday prompt

Number four is the most uncomfortable one on the list, and probably the most valuable. It compounds because you are not just improving the deal on your desk. You are improving the judgment behind every deal that comes after it.

you're a post-mortem analyst. i'm attaching [insert number] investment committee memos we wrote between [insert dates], plus what actually happened on each of those deals: [insert what you have, e.g. current occupancy, actual T-12 NOI, actual capex spent, hold status].

for each memo, pull every forward-looking claim we made that can now be checked. rent growth, lease-up timing, expense assumptions, capex budget, exit timing, exit cap.

build one table: deal, the claim we made, the number we used, what actually happened, the miss in percent, and whether we missed high or low. quote the exact sentence from the memo in every row.

then tell me which assumption we get wrong most often and in which direction.

if a memo doesn't state something, write [NOT IN THE MEMOS] rather than inferring it from the deal type. don't soften a miss, don't average a bad one away, and don't make up data.

Good output should sting a little. If the table comes back flattering, your inputs were probably too thin or the quotes were invented. Check two of the cited sentences against the actual memo before you trust the rest.

The real value comes after the table. You discover that you keep getting the same assumption wrong, in the same direction, deal after deal. Most shops have a bias like this, but I guess it is part of the game. Or is it?

Free Guide: Claude vs ChatGPT for CRE underwriting

Read free AI guides for commercial real estate on the PSV website!

Explore Free AI Guides

About our team (glad you asked)

If you're new here, my name is Zed. I left my transactions role at PGIM to team up with software engineers from Goldman Sachs and HP building frontier AI for commercial real estate at PSV. My co-founders are deeply technical, with AI research funded by AWS and the National Science Foundation, and one of our advisors was Uber's first head of engineering, now an AI PhD at Berkeley.

Combined, our team has closed over $1 billion in CRE transactions. We're based in San Francisco, in the middle of the AI world, and most people gatekeep what we learn here. I'd rather translate it into quick 10-minute reads for fellow real estate people, for free, forever. (Fun fact: we've got robot dogs in our office!)

Everyone is touring units. Go check the crawlspace.

Follow PSV on LinkedIn

Sources

1. PSV & CRE AI Institute Research Desk.

2. Reality check: The true pace and payoffs of AI adoption in corporate real estate, Global Real Estate Technology Survey, October 27, 2025 (JLL)

3. 2026 commercial real estate outlook, September 29, 2025 (Deloitte Insights)

Tenant questions. Repairs. Follow-ups. Rent issues. Vendor coordination. Don’t get me started on accounting and the mid-day call about water pressure that ends up being the city’s fault.

None of it is especially hard. There is just always something to deal with.

That is why AI matters so much here. A growing share of the repetitive work can now be handled or prepared automatically, while the manager stays focused on the parts that actually require judgment and a human relationship.

A UK lettings company now says one manager can carry roughly 300 units instead of 100, and investors just put $170 million behind the operating model.

That does not mean every manager should run 300 units. It means units per head is becoming a technology question, not just a staffing assumption.

By the end of this issue, you’ll have a stress test for figuring out how much more work your team can handle before you need to hire, without copying a startup’s claim or cutting a job on faith.

Welcome. This is Volume 17 of Keeping Up With AI in CRE.

It all comes down to the per-unit

Most property-management staffing models start with simple math:

Units ÷ property managers = capacity.

But anyone who has managed real estate knows that number can be misleading.

One manager might oversee a newer building with clean files and stable tenants. Another might manage the same number of units across four older properties, multiple ownership groups, and an active renovation.

Same unit count. Completely different workload.

Dwelly is trying to change the work underneath that equation. It acquires independent lettings agencies, moves them onto one operating platform, and uses AI to handle as much repetitive operational work as possible.

Then you get the number everyone will remember: 300 units per Dwelly property manager, versus roughly 100 traditionally.

One caveat before anyone screenshots that 300-unit number for an IC deck.

This is UK residential, not US commercial real estate. And it is Dwelly’s own comparison, not an independently audited industry benchmark. Building age, tenant profile, geography, service model, and accounting responsibilities can all change the workload.

But the bigger point still holds.

An institutional investor just backed a company whose economics depend partly on increasing the number of units one employee can manage.

Whether Dwelly actually gets to 300 is their problem.

The fact that units per head is now a technology-driven input is now yours.

What actually changes the ratio

A property manager does not go from 100 units to 300 because they answer emails a little faster.

The ratio moves when entire categories of work disappear from the week.

And that idea applies across CRE.

An acquisitions team can review more deals. A broker can cover more owners. An asset manager can oversee more properties. A property manager can carry more units.

JLL has said publicly that it expects to grow revenue with AI without adding headcount over a two-year period. Different roles, same underlying idea: increase how much work each person can carry without lowering the quality of the output.

So the question is not just whether Dwelly can get to 300 units per manager.

It is:

What is actually consuming your team’s time today, and which parts of that workload can AI remove?

For a property manager, that might be tenant communication, work orders, delinquency follow-up, and owner reporting. For a broker, it might be research, prospecting, CRM updates, and follow-up. For an acquisitions team, it might be document review, underwriting prep, and IC materials.

That is where the operating leverage comes from.

Find the work before you automate it

This is where most AI implementation goes backwards. Most teams start with a tool, then go looking for a problem to give it. Even worse, many teams start with multiple tools.

Notebook LM this, Manus that. Perplexity this. Kimi that. I mean, have you ever met a successful broker who covers all major and alternative asset classes? Me neither!

Simplicity is your friend. For most CRE teams, one Claude or ChatGPT subscription can already do far more than they realize.

Inside the CRE AI Institute, members have built dashboards, relationship-management agents, market-news agents, and CRE signal trackers without spending thousands of dollars on consultants or complicated software.

In many cases, the difference was a ten-minute lesson and a little more time learning how to work with Claude.

What’s your excuse not to start?

Explore CRE AI Case Studies

Anyways, let’s get back to it.

A better approach is to start with the work.

For one week, track the recurring tasks your team actually spends time on. Not the job description. The real work.

For a property manager, that might be work orders, tenant emails, delinquency follow-up, owner reporting, and vendor coordination.

For a broker, it might be owner research, CRM updates, follow-up, BOV preparation, and pulling comps.

For an acquisitions team, it might be screening deals, reviewing OMs, normalizing rent rolls, updating models, and preparing IC materials.

Then sort the work into three buckets:

Repeatable: AI can draft, extract, route, summarize, or prepare it.

Judgment: AI can prepare the information, but a person still needs to decide.

Human: the work depends on a relationship, negotiation, physical presence, or accountability.

Then rank everything by total hours.

That last part matters.

Automating a ten-minute annoyance feels productive. Removing a task that quietly consumes six hours every week changes how much work the team can actually carry.

That is the number I would optimize for first.

Sharing is caring

Finding this useful? Share it with the owner or operator who still calculates staffing capacity from the org chart instead of the operating week.

Share

Keep it simple

You do not need a new AI platform to run this test. Paste the five-day logs into Claude.

The entries can be messy. The evidence rules cannot.

You are an operations analyst for a [PROPERTY TYPE] portfolio.

Attached is a five-day task log kept by [NUMBER] property managers covering [NUMBER] units.

1. Group every logged task into recurring task types and total the minutes for each.

2. Tag each task type SCRIPTED, JUDGMENT, or ON SITE. Give one sentence explaining each tag.

3. Rank the task types by total minutes and show the top four as a percentage of logged time.

4. Calculate scripted minutes per unit per week.

5. Model how many units one manager could carry if scripted work arrived already drafted, summarized, or routed, while holding judgment and on-site work constant.

6. Show every step of the arithmetic.

7. List everything the log does not reveal that could reduce the capacity estimate.

Use only the attached logs. Cite the entries behind every number. Write [NOT IN LOG] instead of estimating. Do not use outside staffing benchmarks and do not recommend a headcount change.

Return:
A. Task-type table
B. Top-four time sinks
C. Current units-per-head calculation
D. Scenario calculation
E. Missing evidence and manual-verification checklist

No preamble. Use tables where possible.

Read section E first.

The biggest risk in a staffing model is usually the work nobody captured: the emergency call, the surprise inspection, the owner request that took “five minutes” six different times.

Claude can make the math look clean. It cannot make an incomplete week complete.

That is why this is a capacity test, not a headcount recommendation.

Bring one real bottleneck into the CRE AI Institute

The prompt can read one week. It cannot redesign the week around it.

That is where most internal AI projects stall. Someone finds a repetitive task, writes a decent prompt, and saves it in a personal folder. Three weeks later the prompt is gone, the manager is back in Excel, and nobody knows whether it ever saved an hour.

Inside the CRE AI Institute, bring one real task from the scripted pile. The 20-plus modules and prompt library help you build the first version. Weekly office hours are where we deal with the part the clean example missed.

• Use the task log to prove where the time goes.

• Define what Claude may draft, what needs approval, and what stays human.

• Build the first version with the prompt library and course modules.

• Bring the awkward edge cases to weekly office hours.

• Test the workflow against a real week before calling the time savings real.

If you join, do not show up with “we need an AI strategy.” Show up with a manual task. That is enough to start.

Build Your First CRE AI Workflow

About our team (glad you asked)

If you are new here, my name is Zed. I left my transactions role at PGIM to team up with software engineers from Goldman Sachs and HP building frontier AI for commercial real estate at PSV.

Our team has closed more than $1 billion in CRE transactions. We are based in San Francisco, in the middle of the AI world, and I would rather translate what we learn into practical ten-minute reads for real estate people than gatekeep it.

Quick summary

• Dwelly raised $170 million while managing more than 15,000 properties on a £350 million rent roll [1].

• The company reports roughly 300 units per manager versus roughly 100 traditionally. That is a company-reported UK residential comparison, not an industry benchmark [1].

• The important shift is that institutional capital is pricing units per head as a technology-sensitive operating input.

• Stress test your own ratio: show the real week, separate the work by who must own it, find the time sinks, model the higher unit count, and verify it before changing payroll.

• Keep the evidence fence. [NOT IN LOG] beats a precise staffing number built on invisible assumptions.

Nobody raises class size by telling the teacher to talk faster. Change what the teacher has to carry.

Follow PSV on LinkedIn

Also worth knowing this week

Texas put roughly 474 gigawatts of interconnection requests behind an audit. On August 3, Governor Abbott directed the PUCT and ERCOT to verify every data center moving through the interconnection process. The queue is more than five times the state’s record peak demand, and the directive says noncompliant projects will be denied connection [3]. The permissive market did not close the door. It started checking the guest list.

The largest US grid plans to curtail certain large data-center loads beginning in June 2027. PJM, which serves 67 million people from Virginia to Illinois, said loads of 50 MW and above may be temporarily curtailed during unmanageable demand events [4]. “Powered” is no longer a yes-or-no diligence answer.

Anthropic reviewed 141,006 cybersecurity-evaluation runs and identified three incidents involving real systems. The company attributed them to evaluation-environment misconfiguration and said it found no evidence of models pursuing goals of their own [5]. If an AI agent has a login, environment controls matter as much as the prompt.

Anthropic canceled Claude Sonnet 5’s planned price increase. Its current documentation says the $2/$10 per million input/output token rate is now the standard price, and the scheduled move to $3/$15 on September 1 will not occur [6]. If you already increased the budget, move that line back down.

Blackstone and Brookfield are building proprietary real-estate AI. Commercial Observer reported that some of the largest owners are developing tools internally rather than relying entirely on vendors [7]. The strategic question is becoming which layer to own, not whether to use AI.

Sources

1. CRE AI Institute Research Desk - PSV.

2. Dwelly raises $170M to turn UK lettings agencies into software - TNW.

3. How commercial property management firms are using AI - Facilities Dive.

4. Governor Abbott Directs Comprehensive Data Center Audit - Office of the Texas Governor.

5. Data centers may face temporary power cuts on largest US grid - TechCrunch.

6. Investigating incidents in cybersecurity evaluations - Anthropic.

7. Pricing - Anthropic developer documentation.

8. Blackstone and Brookfield Are Using AI to Build Real Estate Tech In-House - Commercial Observer.

But the lobby is not the building.

It is the part investors photograph. The elevators, electrical rooms, access control, plumbing, and operating staff are the factors that determine whether the place actually works.

CRE dashboards have always had the same problem.

The charts are the lobby. The expensive part is everything behind them: connecting the data, cleaning it, defining the metrics, controlling permissions, and making sure Monday’s number still means the same thing on Friday.

Claude is making the lobby extraordinarily cheap.

You can now describe the portfolio view you want in normal language, attach a spreadsheet, and get a working dashboard or prototype without first writing a software specification. See below, example built by one of our members at the CRE AI Institute - this took only 30 minutes of his time.

Anthropic is explicitly pushing Claude toward file creation, interactive artifacts, and conversational design.

That does not mean every $50,000 dashboard is now a $50 prompt.

It means the visible interface is losing scarcity. Data plumbing, governance, workflow design, and production reliability are becoming the actual product.

By the end of this issue, you will have a five-layer test for deciding whether your next CRE dashboard should be prompted, purchased, or professionally built.

Welcome. This is Volume 16 of Keeping up with AI in CRE.

Free AI Guide This Tuesday - AI, explained with no buzzwords for CRE professionals.

Check out the free guide below on how to “speak AI” and impress your committee, without being confused by tech jargons. Don’t be that buzzword guy!

Read the Guide

Now let’s get back to it.

The interface was never the whole cost

A CRE dashboard usually starts as a simple request. The acquisitions head wants pipeline visibility. Asset management wants budget-to-actuals. The managing partner wants one place to see what needs attention.

Then the simple request hits the building systems.

The pipeline lives in a CRM. Property facts live in Excel. Documents live in Drive. Actuals come from accounting. Market data sits behind a license. The same asset appears under three names, two entity structures, and four different IDs.

By the time the data is connected and normalized, the chart itself is the easy part.

Conversational design flips the sequence. You can build the useful screen first, put it in front of the team, and learn which decisions it actually changes before wiring every system.

That is a major shift. A prototype used to be expensive enough that firms treated the original requirements as truth. Now the prototype can become the requirements process.

The risk is confusing a convincing interface with a dependable system.

The five layers of a real dashboard

Use these layers to decide what Claude can do now and where professional implementation still earns its keep.

1. The conversation layer

This is the request in plain English: what the user wants to know, what action follows, and what exceptions should be visible. Claude is excellent here because users can refine the screen by reacting to it instead of translating their job into technical requirements.

Example prompt: “Help me define this dashboard in plain English. Ask me what decisions it should support, what questions users need answered, what actions follow, and which exceptions or edge cases must be visible.”

2. The interface layer

This is the lobby: charts, tables, maps, filters, drill-downs, and buttons. AI can generate this quickly. For a single user or a temporary decision, that may be enough.

“Design the dashboard interface for this use case. Recommend the charts, tables, KPIs, filters, drill-downs, maps, and actions needed. Keep the screen focused on the decisions the user actually needs to make.”

3. The data layer

This is where most dashboard projects become real work. Every source needs an owner, refresh rule, field definition, stable identifier, and error state. MCP can standardize how models connect to systems, but a connector does not decide which rent roll is authoritative or whether “committed” means signed, funded, or merely approved.

4. The control layer

Permissions, audit logs, review gates, security, backups, uptime, and change management live here. If a dashboard can expose tenant data, change a deal stage, send an email, or influence a capital call, the control layer is not optional.

5. The workflow layer

The dashboard must sit inside a recurring decision. Who opens it, when, what do they do next, and what happens when a number is wrong? A beautiful screen with no operating cadence becomes another tab nobody checks.

Prototype first, harden second

Start with one decision, one owner, and one data file. Build the conversational prototype in a day. Use it for two weeks. Track which views get opened, which numbers get challenged, which actions follow, and where users still return to Excel. Only then should you connect source systems and engineer the controls.

The Dashboard Specification Prompt

Before you buy or build anything, give Claude this prompt with a sample export and one real weekly meeting agenda.

You are designing a commercial real estate operating dashboard. Use the attached sample data and meeting agenda.

Start with the decision, not the chart.

Output only these sections:

Users: who will use the dashboard and what authority each person has.

Recurring decisions: the weekly or monthly decisions the dashboard must support.

Minimum views: the smallest set of tables, charts, filters, and drill-downs required.

Source map: for every field, name the source file or system, owner, refresh frequency, and authoritative identifier.

Definitions: write an exact definition for every KPI. Mark disagreements or missing definitions [NOT DEFINED].

Controls: permissions, approval gates, audit history, error states, and actions that must remain read-only.

Prototype plan: what can be built from the sample files now.

Production gaps: what must be integrated, secured, tested, and monitored before firmwide deployment.

End with a recommendation: prompt a one-user prototype, buy existing software, or build a production system. Explain the gating reason in one sentence.

If the source map is longer than the feature list, that is normal. You have finally found the real project.

The $50,000 dashboard is not disappearing. It is being unbundled. The screen may cost almost nothing to prototype. The dependable data and controls behind it are where serious firms will still spend.

The question is, when should you hire a vendor, and when should you start building internally with Claude?

Outsourcing vs. Do-It-Yourself

The line is not “Can Claude build this?” It usually can. The better question is: What happens if the dashboard is wrong? Who takes the responsibility? A good rule of thumb is below.

Build the first version yourself. Outsource the parts where reliability, governance, and integration start to matter.

Do it yourself when:

• You are the primary user.

• The dashboard is exploratory, temporary, or used for internal analysis.

• The underlying data is already clean and easy to access.

• A mistake is easy to notice and cheap to fix.

• You are comfortable maintaining prompts, connectors, and logic as things change.

This is where Claude is extremely powerful. You can go from idea to working interface quickly without turning every dashboard request into a software project.

Outsource when:

• Multiple people or departments depend on the dashboard.

• Data comes from several systems with conflicting definitions.

• Permissions differ by user, property, fund, or role.

• Bad data could affect investors, tenants, reporting, compliance, or capital decisions.

• It needs to work reliably every day without the person who built it babysitting it.

• You need proper testing, monitoring, auditability, and long-term maintenance.

Need help with building custom AI dashboards?

People think dashboards are just for visualization and staying organized. In the age of AI, they’re worth more than that. The real value is in intelligence & judgement.

Your Claude, ChatGPT, or Copilot might get you 60% of the way out of the box.

The missing 40% is your firm’s context.

It doesn’t know that you underwrite San Diego differently than Phoenix. It doesn’t know why a $32/SF comp is misleading, which landlord quietly gives six months of free rent, which broker’s guidance you actually trust, or that a seller has retraded the last three deals.

It sees the data. It doesn’t know the story behind the data.

Your AI needs a brain that captures those assumptions, relationships, decisions, and exceptions, and a dashboard that makes that intelligence usable.

That’s what we build at PSV. Really cool AI brains and really cool dashboards for CRE.
You should talk to our team HERE if this sounds interesting.

Work With Me

Want to take a stab at building with Claude for CRE tasks?

Join the CRE AI Institute

About our team (glad you asked)

If you’re new here, my name is Zed. I left my transactions role at PGIM to team up with software engineers from Goldman Sachs and HP to build the #1 AI brain for commercial real estate at PSV. My co-founders are deeply technical, with AI research funded by AWS and the National Science Foundation, and one of our advisors was Uber’s first head of engineering, now an AI PhD at Berkeley.

We have over $1 billion in CRE transaction experience. Building at the intersection of the CRE space and AI world, our goal is to shed some light into this new technology revolution, where misinformation and hype is rampant.

Fun fact: Gedeon speaks fluent french!

About The Team

Quick summary

• Claude is collapsing the cost and time required to create the first useful CRE dashboard interface.

• The charts were never the whole product. Data definitions, integrations, permissions, auditability, and workflow remain the hard part.

• A prototype proves the decision and user behavior. A production system proves trust, security, and reliability.

• Begin with one decision, one owner, one sample file, and one recurring meeting.

• Use the five-layer test before deciding to prompt, buy, or build.

• Spend engineering dollars only after the conversational prototype shows what deserves to be hardened.

The lobby got cheap. The building systems did not.

Follow PSV on LinkedIn

Also worth knowing this week

The $50 billion Texas data center lease is contractually a $19.6 billion lease. Hut 8 disclosed two 352 MW leases at Beacon Point on July 20: 15-year base terms, triple net, 3% annual escalators, three five-year renewal options each, and $50.2 billion only if every option is exercised [3]. The FT then reported the unnamed “high-investment-grade” tenant is Nvidia, which Nvidia declined to confirm and Reuters could not verify [4]. Base term is what somebody owes. The rest is a ceiling.

Nvidia is reportedly in talks to guarantee roughly $250 billion of OpenAI’s lease and construction debt on a 10-gigawatt Ohio campus. The chips are explicitly excluded, phase one is around 800 MW targeted for 2028, and nobody involved would confirm any of it [5]. If a supplier has to guarantee the rent, the guarantor is the credit.

A UK lettings platform raised $170 million on the claim one manager can run 300 units instead of 100. Dwelly’s round split $95 million equity and $75 million debt, led by EQT Growth, across 15,000-plus properties [6]. UK residential, and the ratio is the company’s own number, so don’t quote it as a benchmark. Worth knowing your own units-per-head figure though, because capital is now underwriting it.

JLL and MIT say AI leasing has decoupled from the tech downturn. US tech employment fell 1.5% in early 2026 while tech office leasing rebounded, and AI companies have made up nearly 30% of San Francisco leasing since 2025 [7].

AMD will deploy up to 2 gigawatts of Instinct MI450 GPUs for Anthropic, with the first gigawatt starting in the first half of 2027 and an AMD equity investment of up to $5 billion [8]. Every one of these announcements is a power bill somebody’s grid has to absorb, and increasingly a permit somebody has to win.

Sources

1. CRE AI Institute Research Lab (PSV)

2. Create and edit files with Claude (Anthropic)

3. Model Context Protocol documentation (Anthropic)

4. Hut 8 Fully Commercializes 1 GW Beacon Point Campus with Second 352 MW Lease (Hut 8 / Yahoo Finance)

5. Nvidia behind $50 billion lease on Texas data center, FT reports (Reuters)

6. Nvidia in talks to guarantee $250 billion financing for OpenAI Ohio campus (Reuters on WSJ)

7. Dwelly raises $170M to turn UK lettings agencies into software (TNW)

8. AI driving deep divergence across real estate markets (JLL with MIT)

9. AMD and Anthropic announce strategic partnership to deploy up to 2 gigawatts (AMD)

The broker sends the book to forty buyers. Thirty never open it. Eight say the basis is too high. Two tour.

Then one bidder hears the other is still engaged and suddenly finds another $750,000.

Nothing about the property changed.

The rent roll was the same. The debt market was the same. The underwriting was the same. Even the PDF was the same.

So where did the extra $750,000 come from?

That question matters more as Claude starts absorbing the visible work of investment sales.

It can read the OM, normalize the rent roll, build the buyer list, draft the outreach, create the presentation, compare bids, and answer a surprising amount of diligence.

Soon, much of what clients associate with brokerage work will take minutes instead of weeks. Which raises a much harder question:

Once the analysis, materials, outreach, and process management are largely automated with AI, what exactly is the broker still getting paid for?

We’ll answer this question in a quick 10-minute read.

Welcome. This is Volume 15 of Keeping up with AI in CRE.

Free AI Guide This Thursday - Claude for Deal Prospecting

Check out the free guide below on how to replace an expensive AI consultant by building yourself an outbound prospecting system.

Check Out The Guide

Now let’s get back to it.

Has it ever really been about the OM?

The OM feels like the product because it is the object everyone sees. It is also the easiest part of the process to automate.

Give Claude the rent roll, T-12, leases, photos, and market notes, and it can turn the same source material into an executive summary, buyer FAQ, diligence list, email campaign, call brief, and multiple versions of a pitch.

That is real value. It eliminates handoffs, reduces transcription, and allows a smaller team to prepare a process that once required several analysts and associates.

But the sale price is not hiding inside the PDF waiting to be extracted.

The clearing price emerges from a specific group of buyers, with specific capital, during a specific week, reacting to each other inside a controlled process.

A good broker knows which “no” is final, which “no” means call me at a different basis, and which silent buyer is still underwriting the deal.

Claude has broad knowledge. The broker has private, local, and time-sensitive knowledge.

That private layer is where much of the fee survives. But not all brokers will.

The closer a task gets to documents and repeatable rules, the more Claude can own. The closer it gets to buyer psychology, disclosure judgment, negotiation, or incomplete human signals, the more the broker matters.

So if your value right now is simply “doing the work,” good luck. The broker who uses Claude daily is going to out-pace you, even if he’s not as qualified.

So what should you do?

The AI-First Broker

The better operating model is not a traditional brokerage team with a chatbot bolted onto it.

It is a smaller senior team supported by an AI production layer.

A good broker remains in control of buyer selection, disclosure, pricing strategy, negotiation, and escalation.

A great broker lets Claude handles the work required to make those decisions faster and with better information.

Here is what that looks like in practice.

1. Build a smarter buyer universe

Most buyer lists are static exports from a CRM.

Claude can help turn them into a live market map.

Feed it prior transactions, third-party intellifence, geography, check size, asset preferences, recent acquisitions, call notes, and known objections. Then ask it to rank buyers by fit and explain why each one belongs on the list.

Brokers can decide how to add the private layer: who has capital today, who can actually close, who is wasting time, and who needs a phone call rather than another email.

We teach you how to do this securely as part of the CRE AI Institute.

2. Personalize the process without slowing it down

A generic email blast signals that the broker has not thought seriously about the buyer.

Claude can produce buyer-specific positioning based on strategy, portfolio gaps, geography, and prior activity. It can also create follow-ups, meeting briefs, and call questions for every serious prospect.

That lets the broker enter each conversation with a sharper thesis:

“This fills the gap between your Phoenix and Inland Empire portfolios.”

“This matches the smaller-bay strategy you described on our last call.”

“This lease rollover is the risk, but it is also where your operating platform creates upside.”

The machine prepares the argument. The broker earns the response.

3. Turn every conversation into process intelligence

The most valuable information in a sale often never enters the CRM.

A buyer says the basis is too high. Another is worried about rollover. A third likes the deal but cannot get its lender comfortable.

Claude can summarize every call, classify objections, update buyer status, identify repeated concerns, and show where the process is weakening.

Instead of relying on memory, the broker can see:

• which objection is becoming a pattern

• which buyers are still doing work

• where the materials are creating confusion

• which issue is likely to produce a retrade

• where additional information could restore conviction

This is where AI becomes more than a writing tool. It becomes a live intelligence layer around the process.

4. Prepare for negotiations before they happen

Claude should not run the negotiation. It should make the broker better prepared for it.

Before a call, it can summarize the buyer’s position, identify open issues, compare the bid against other offers, model likely concessions, and produce a negotiation range.

It can also distinguish between price and certainty.

A higher bid with weak financing, broad contingencies, and an aggressive diligence period may be worse than a lower bid with real capital and limited execution risk.

Claude can organize the evidence. The broker still decides how hard to push.

5. Make diligence feel controlled

Diligence is where a polished sale process often becomes an email scavenger hunt.

Claude can maintain the request list, identify missing documents, compare versions, summarize buyer comments, flag inconsistencies, and prepare daily status reports.

The broker then spends less time chasing files and more time solving the issues that could actually threaten the closing.

The goal is not to remove the broker from diligence. It is to stop using the broker as a human project-management system.

CRE AI Institute Workflow Tip

Take your last closed deal and give Claude the deal files, process documents, emails, call notes, buyer list, bid comparison, and closing checklist.

Then use this prompt:

You are redesigning a commercial real estate investment-sales process to make the brokerage process materially more effective.

Review the attached deal files, process documents, emails, call notes, buyer records, bid comparison, diligence tracker, and closing checklist.

Output five sections:

1. Production work Claude could own, including document review, research, drafting, tracking, and reporting.

2. Decisions Claude could prepare but the broker should make, including buyer selection, process timing, pricing strategy, disclosure, and escalation.

3. Relationship and negotiation work that should remain broker-controlled.

4. Information that was lost, delayed, or trapped in emails, calls, spreadsheets, or individual memory during the process.

5. A redesigned workflow showing how one senior broker could run this deal faster, with fewer handoffs and better buyer intelligence.

For every recommendation, include:

• the required input

• Claude’s role

• the broker’s role

• the approval gate

• the likely time saved

• the risk if automated without human review

Do not assign Claude relationships, authority, intent, or knowledge that is not evidenced in the attached materials.

End with:

• the five tasks consuming the most human time that can move to Claude now

• the five decisions that should remain human-controlled

• the three changes most likely to improve price, certainty, or speed on the next sale

If the output says everything is automatable, the process description is incomplete. Add the calls, judgment, exceptions, and private information that actually moved the deal.

The result should not be a plan to remove the broker.

It should be a plan to remove the work preventing the broker from brokering.

Want to learn how to implement AI practically without buzzwords via step-by-step video guides?

Join the CRE AI Institute

Show your support for free AI in CRE content by sharing

Know a colleague still splitting a 140-page OM into four chats at 9pm? Forward them this. That is honestly how most people here found us. And if you refer 5 people, you get free access to our copy-and-paste Claude Skill Vault, the exact prompts and workflows we deploy for CRE clients, including the deal brief above.

Thanks so much for sharing! It helps our team continue to grind out really good content for you alll

Share

About our team (glad you asked)

If you’re new here, my name is Zed. I left my transactions role at PGIM to team up with software engineers from Goldman Sachs and HP building frontier AI for commercial real estate at PSV. My co-founders are deeply technical, with AI research funded by AWS and the National Science Foundation, and one of our advisors was Uber’s first head of engineering, now an AI PhD at Berkeley.

Combined, our team closed over $1 billion in CRE transactions. We’re based in San Francisco, in the middle of the AI world, and most people here gatekeep what they learn. I’d rather translate it into quick 10-minute reads for fellow real estate people, for free, forever. (Fun fact: we got robot dogs in our office!)

Quick summary

• Claude can own much of investment-sales production: extraction, first-pass analysis, drafting, buyer research, and process tracking.

• The visible deliverable was never the full moat. Buyer access, private information, price discovery, negotiation, and execution remain human-heavy.

• Use task-level replacement decisions. “Can AI replace the broker?” is too broad to operate.

• Require human approval before facts, disclosures, outreach, pricing signals, or negotiation positions leave the firm.

• The broker who sells production is exposed. The broker who creates competition becomes more leveraged.

• Run the replacement matrix on one closed deal before redesigning the team.

AI can make the book. The broker still has to make the market.

Follow PSV on LinkedIn

Also worth knowing this week

The $50 billion Texas data center lease is contractually a $19.6 billion lease. Hut 8 disclosed two 352 MW leases at Beacon Point on July 20: 15-year base terms, triple net, 3% annual escalators, three five-year renewal options each, and $50.2 billion only if every option is exercised [3]. The FT then reported the unnamed “high-investment-grade” tenant is Nvidia, which Nvidia declined to confirm and Reuters could not verify [4]. Base term is what somebody owes. The rest is a ceiling.

Nvidia is reportedly in talks to guarantee roughly $250 billion of OpenAI’s lease and construction debt on a 10-gigawatt Ohio campus. The chips are explicitly excluded, phase one is around 800 MW targeted for 2028, and nobody involved would confirm any of it [5]. If a supplier has to guarantee the rent, the guarantor is the credit.

A UK lettings platform raised $170 million on the claim one manager can run 300 units instead of 100. Dwelly’s round split $95 million equity and $75 million debt, led by EQT Growth, across 15,000-plus properties [6]. UK residential, and the ratio is the company’s own number, so don’t quote it as a benchmark. Worth knowing your own units-per-head figure though, because capital is now underwriting it.

JLL and MIT say AI leasing has decoupled from the tech downturn. US tech employment fell 1.5% in early 2026 while tech office leasing rebounded, and AI companies have made up nearly 30% of San Francisco leasing since 2025 [7].

AMD will deploy up to 2 gigawatts of Instinct MI450 GPUs for Anthropic, with the first gigawatt starting in the first half of 2027 and an AMD equity investment of up to $5 billion [8]. Every one of these announcements is a power bill somebody’s grid has to absorb, and increasingly a permit somebody has to win.

Sources

1. PSV Research Desk

2. Claude for Financial Services (Anthropic)

3. Investment Sales (JLL)

4. Hut 8 Fully Commercializes 1 GW Beacon Point Campus with Second 352 MW Lease (Hut 8 / Yahoo Finance)

5. Nvidia behind $50 billion lease on Texas data center, FT reports (Reuters)

6. Nvidia in talks to guarantee $250 billion financing for OpenAI Ohio campus (Reuters on WSJ)

7. Dwelly raises $170M to turn UK lettings agencies into software (TNW)

8. AI driving deep divergence across real estate markets (JLL with MIT)

9. AMD and Anthropic announce strategic partnership to deploy up to 2 gigawatts (AMD)

Every time the wheel nudged me back toward the center, I pulled the other way. The car corrected. I corrected back. The two of us wobbled down the 101 arguing about where the lane was.

The car was not the problem. My hands were.

At the time, I wanted to become a pro boxer and had overworked them badly enough that even small movements felt unnatural. So I kept resisting a system that was trying to help. But why should you care about my short-lived boxing career?

Because something similar happened to a lot of Claude prompts last Friday.

Claude got better at handling certain tasks on its own. But many saved prompts still contain old instructions telling it exactly how to do those same tasks step by step.

Now the model and the prompt are both trying to steer.

You can skip this if you do not reuse saved prompts on deals. But if you do, this is a ten-minute cleanup that will likely pay for itself.

Welcome. This is Volume 13 of Keeping up with AI in CRE.

The 30-second version

On July 24, Anthropic released Claude Opus 5. Three changes matter for CRE.

Pricing stayed the same as Opus 4.8, so reviewing OMs, leases, and deal files should not cost more.

The 1 million-token context window is now the default. Think of it as a larger deal room: Claude can hold the OM, rent roll, T-12, leases, and market reports at once.

Extended thinking is also on by default. Use lower effort for emails and summaries, and higher effort for underwriting, lease analysis, and diligence.

Anthropic calls it “near Fable 5 intelligence at half the cost.” However, one line in Anthropic’s own documentation should send you straight to your saved prompts.

Claude now checks its own work automatically. Anthropic says to remove older instructions like “include a final verification step” or “use a subagent to verify,” because they can make the model over-check simple work.

In CRE terms, you may be asking Claude to review the OM, then review its review, then assign another agent to review both.

Most people have their Claude set up to be an intelligent analyst, which means checking its work, without realizing they have that feature.

Your careful prompt hygiene from six months ago may now be the hands fighting the wheel. And it will cost you a heck lot of tokens. But don’t worry, we’ll fix it below.

Free AI Case Study This Thursday - Build Your Own Deal Management Dashboard

Check out the free video guide below on how to replace an expensive dashboard subscription with just a few hours of focused work.

Learn How It Works

Now let’s get back to it.

Why token efficiency matters

Most people think of tokens as a technical detail. They are really the meter running in the background.

When Claude is connected through an API, you are generally billed for what the model reads and what it writes. Every oversized prompt, repeated instruction, unnecessary file, and redundant verification step adds to the cost. At one deal, that may barely register. Across hundreds of OMs, leases, underwriting runs, and follow-ups, it becomes an operating expense.

Think of it like change orders on a development project. One extra request does not break the budget. But enough small additions, unclear scopes, and duplicated work can push the final cost far beyond the original estimate.

That is also how overages happen. A workflow may be priced around normal usage, but if users repeatedly rerun large files, ask for excessive outputs, or force Claude to redo work it already completed, the API bill can exceed the allowance built into the contract.

Token efficiency does not mean giving Claude less context. It means giving it the right context once, removing instructions the model no longer needs, and avoiding work that does not improve the final answer.

The goal is simple: better output per token, more predictable costs per deal, and fewer surprise overages at the end of the month.

Now, let’s clean up that prompt.

The Prompt Retrofit

Pull up the prompt you actually reuse: the deal brief, rent roll normalizer, lease abstractor, or whatever your one good prompt is. Then make these five edits.

1. Delete the verification boilerplate

Remove lines like “double-check every number,” “include a final verification step,” or “use a second pass to verify.” Claude now checks its own work automatically, and Anthropic says these older instructions can cause over-verification. You do not necessarily get a more accurate answer. You often get a longer one that spends more tokens reviewing itself.

2. Keep the sourcing rules

This is where people make the wrong edit. “Verify your work” tells Claude how much effort to use. “Write NOT IN DOCS instead of estimating, and cite the page for every figure” tells Claude where the answer is allowed to come from. The first may now be redundant. The second is still essential. Nothing about this release guarantees Claude will admit when a figure is missing, so keep provenance rules in every deal prompt.

3. Stop splitting one deal across four chats

In Volume 7, I recommended stripping down the OM before uploading it. That made sense when large-context analysis came at a premium. It makes less sense now. The 140-page OM, T-12, rent roll, lease stack, and market notes can go in together. Every time you split a deal across separate chats, you create another place for Claude to lose an assumption, miss a relationship, or forget why a number mattered. The larger context window is not just about fitting more pages. It is about preserving the full deal narrative.

4. Define the deliverable

Opus 5 tends to produce longer answers. That is useful when you want a detailed investment memo and annoying when you asked for twelve rent-roll exceptions and received a five-page essay. Put the shape directly in the prompt: “one page,” “twelve rows maximum,” “table only, no preamble,” or “three bullets per section.” A few words in the prompt can save hundreds in the output.

5. Use the effort setting intentionally

For teams using Claude Code or the API, effort is both a quality lever and a cost lever. Use low or medium for extraction, formatting, and repetitive document work. Use high for underwriting, lease-risk analysis, and investment memos. Reserve xhigh or max for the reasoning tasks that genuinely justify the extra time and tokens. In the Claude app, the default handles this for you, which makes the other four edits even more important.

The test

Take one live deal and run it twice: first with the old prompt unchanged, then with the retrofitted version. Compare the outputs side by side. You want the same key numbers and conclusions with less padding. If the new version loses something important, add back the specific instruction that produced it rather than restoring every piece of old boilerplate. Ten minutes, one deal, and you know which instructions are still earning their place.

The Prompt, Retrofitted

Here is my deal brief prompt with the surgery already done. Attach the full deal package at once.

You are a CRE acquisitions analyst. Read every attached document in one pass and produce a compact deal brief I can reuse without reopening or re-pasting the files. Be dense and numeric.

Output only:

Snapshot: asset, property type, address, year built, units or square feet, and submarket.

Price and debt: asking price, price per unit or square foot, in-place cap rate, and existing debt terms.

T-12 operations: normalized revenue, major expense lines, NOI, and one-time items shown separately.

Rent roll: unit mix, in-place versus market rent, occupancy, and concessions.

Leases: expiration schedule and unusual clauses.

The five numbers that drive the deal.

What the documents do not contain.

Every figure must come from the attached documents. Cite the page or section for each one. Write [NOT IN DOCS] instead of estimating, and do not fill gaps using general knowledge about [insert market].

Keep the brief under one page. No preamble and no summary of what you are about to do.

What should come back is a one-page brief with citations you can spot-check in about ninety seconds. Section 7 will probably feel uncomfortably long the first time. That is the point. It is not a weakness in the output. It is your diligence list.

At published API prices, a package totaling roughly 400,000 input tokens and 25,000 output tokens would cost less than $3 in model usage. Of course that is illustrative math, not a direct quote. Please price your own package before repeating the number to your boss!!

Not sure where your AI knowledge stands?

Join the CRE AI Institute

Show your support for free AI in CRE content by sharing

Know a colleague still splitting a 140-page OM into four chats at 9pm? Forward them this. That is honestly how most people here found us. And if you refer 5 people, you get free access to our copy-and-paste Claude Skill Vault, the exact prompts and workflows we deploy for CRE clients, including the deal brief above.

About our team (glad you asked)

If you're new here, my name is Zed. I left my transactions role at PGIM to team up with software engineers from Goldman Sachs and HP building frontier AI for commercial real estate at PSV. My co-founders are deeply technical, with AI research funded by AWS and the National Science Foundation, and one of our advisors was Uber's first head of engineering, now an AI PhD at Berkeley.

Combined, our team closed over $1 billion in CRE transactions. We're based in San Francisco, in the middle of the AI world, and most people here gatekeep what they learn. I'd rather translate it into quick 10-minute reads for fellow real estate people, for free, forever. (Fun fact: we got robot dogs in our office!)

Quick summary

• Claude Opus 5 shipped July 24 at $5 and $25 per million tokens, unchanged from Opus 4.8 [1][2].

• Anthropic's docs say to remove verification instructions carried over from older models, because they now cause over-verification [2].

• Keep your provenance rule. "Cite the page, write [NOT IN DOCS]" is about where numbers came from, and nothing in this release replaces it.

• 1M tokens is now both the default and the maximum context, so the whole deal package goes in one pass instead of four chats [2].

• Responses run longer by default. Say how long you want it.

• Run old prompt versus retrofitted prompt on one live deal before you roll it out to the team.

Stop fighting the wheel. The car got better at staying in the lane, and your job moved to deciding where you're going.

Follow PSV on LinkedIn

Also worth knowing this week

The $50 billion Texas data center lease is contractually a $19.6 billion lease. Hut 8 disclosed two 352 MW leases at Beacon Point on July 20: 15-year base terms, triple net, 3% annual escalators, three five-year renewal options each, and $50.2 billion only if every option is exercised [3]. The FT then reported the unnamed "high-investment-grade" tenant is Nvidia, which Nvidia declined to confirm and Reuters could not verify [4]. Base term is what somebody owes. The rest is a ceiling.

Nvidia is reportedly in talks to guarantee roughly $250 billion of OpenAI's lease and construction debt on a 10-gigawatt Ohio campus. The chips are explicitly excluded, phase one is around 800 MW targeted for 2028, and nobody involved would confirm any of it [5]. If a supplier has to guarantee the rent, the guarantor is the credit.

A UK lettings platform raised $170 million on the claim one manager can run 300 units instead of 100. Dwelly's round split $95 million equity and $75 million debt, led by EQT Growth, across 15,000-plus properties [6]. UK residential, and the ratio is the company's own number, so don't quote it as a benchmark. Worth knowing your own units-per-head figure though, because capital is now underwriting it.

JLL and MIT say AI leasing has decoupled from the tech downturn. US tech employment fell 1.5% in early 2026 while tech office leasing rebounded, and AI companies have made up nearly 30% of San Francisco leasing since 2025 [7].

AMD will deploy up to 2 gigawatts of Instinct MI450 GPUs for Anthropic, with the first gigawatt starting in the first half of 2027 and an AMD equity investment of up to $5 billion [8]. Every one of these announcements is a power bill somebody's grid has to absorb, and increasingly a permit somebody has to win.

Sources

1. The CRE AI Institute Publications

2. Introducing Claude Opus 5 (Anthropic): https://www.anthropic.com/news/claude-opus-5

3. What's new in Claude Opus 5 (Anthropic developer docs): https://platform.claude.com/docs/en/about-claude/models/whats-new-opus-5

4. Hut 8 Fully Commercializes 1 GW Beacon Point Campus with Second 352 MW Lease (Hut 8 / Yahoo Finance): https://finance.yahoo.com/technology/ai/articles/hut-8-fully-commercializes-1-133338063.html

5. Nvidia behind $50 billion lease on Texas data center, FT reports (Reuters): https://kfgo.com/2026/07/27/nvidia-behind-50-billion-lease-on-texas-data-center-ft-reports/

6. Nvidia in talks to guarantee $250 billion financing for OpenAI Ohio campus (Reuters on WSJ): https://virginiabusiness.com/nvidia-talks-guarantee-250-billion-financing-openai-ohio/

7. Dwelly raises $170M to turn UK lettings agencies into software (TNW): https://thenextweb.com/news/dwelly-raises-170m-to-turn-uk-lettings-agencies-into-software

8. AI driving deep divergence across real estate markets (JLL with MIT): https://www.jll.com/en-in/newsroom/ai-driving-deep-divergence-across-real-estate-markets

9. AMD and Anthropic announce strategic partnership to deploy up to 2 gigawatts (AMD): https://newsroom.amd.com/news/amd-anthropic-strategic-partnership/

Now imagine that analyst forgets all of it overnight and shows up brand new every single morning. That's Claude with no company brain. Brilliant, eager, and guessing.

Because of this, true AI adoption is still in the single digits. If there is even a chance this sounds like you, give us the next 10 minutes. We’ll show you how to fix it.

Welcome. This is Volume 12 of Keeping up with AI in CRE.

The 30-second version

MIT says 95% of organizations are getting zero return on generative AI, after $30 to $40 billion spent. A JLL survey says only 5% of CRE firms hit their AI goals, while 88% are piloting something. On July 7, Commercial Observer read those two numbers together and named the culprit: the tools have no context.

The firms pulling real value didn't buy a smarter model than you. They fed theirs the stuff yours is missing: the comps you trust, the expense ratios you assume before a T-12, the debt your shop actually gets quoted.

That grounded source of truth has a name, and you can build Claude one this week without an engineer. It's the House File. The catch is the one rule almost nobody sets up, and it's the difference between a tool that flags a hole and one that paves right over it. There’s more..

Featured AI Win This Week From the CRE AI Institute - Matthew Cook

Matthew replaced what would’ve been a 5-figure dashboard subscription and built a mission control as a source of truth for his business. He can now manage all of his relationships, to-do items, market news, and deal pipeline all in one place. Congrats Matthew, welcome to the Institute!

If you are interested in becoming an AI-first CRE professional check out the CRE AI Institute below:

Join The Institute

Now, let’s get back to it.

MIT found an implementation problem

The MIT finding was not that employees cannot get value from AI.

Workers are already using consumer tools to draft emails, summarize documents, research markets, and finish individual tasks faster. The failure happens when companies try to turn those isolated wins into a repeatable operating system.

MIT’s researchers examined more than 300 public AI initiatives, interviewed roughly 150 executives, and surveyed about 350 employees. Their conclusion was that the divide between the successful 5% and everyone else was not primarily model quality, regulation, or technical talent.

It was learning.

Most enterprise AI systems did not retain corrections, adapt to how the organization worked, or improve from repeated use. An employee could tell the tool that a certain answer was wrong today, then receive essentially the same wrong answer next week.

Scary.

The CRE version reached the same conclusion

JLL found the real estate version of the same problem. Eighty-eight percent of investors, owners, and landlords had begun piloting AI, often across five use cases at once. Yet only 5% said they had achieved all their program goals.

That does not mean the other 95% received nothing. Nearly half had reportedly achieved two or three goals. But it shows how quickly access to AI has outpaced the systems required to scale it. Firms are launching more pilots than their data, workflows, governance, and employees are prepared to support.

A general-purpose model may understand what a cap rate is, but it does not automatically know which comps your acquisitions team trusts, how your asset managers normalize expenses, or which assumptions your investment committee rejects.

Without that institutional context, AI remains useful to individuals but unreliable to the firm.

So how do you fix it? By building what can be called a second brain. But before that, take the AI readiness quiz to see where you really stand.

How AI Ready Are You?

Okay, let’s get to fixing.

How to build your House File in 20 minutes

Step 1. Create one markdown file

Open a blank document and save it as:

HOUSE_FILE.md

Markdown is just plain text with headings. Claude reads it cleanly, it is easy to update, and you do not need special software.

Paste this structure inside:

# Firm Overview

## Markets We Cover
- 

## Asset Classes
- 

# Underwriting Standards

## Market Rent
- 

## Rent Growth
- 

## Vacancy and Downtime
- 

## Operating Expenses
- 

## Exit Assumptions
- 

## Debt Assumptions
- 

# Source Hierarchy

Use sources in this order:

1. Executed property documents
2. Internal firm data
3. Verified broker comps
4. Third-party market reports
5. Offering memorandums

# Approval Rules

Flag these for human review:
- Assumptions outside firm standards
- Conflicting sources
- Missing deal documents
- Unsupported broker projections

# Lessons and Exceptions

- 

Do not fill it with definitions Claude already knows. Add the assumptions your team normally explains verbally.

Step 2. Add the instruction at the top

Paste this directly beneath the title:

Use this file as the firm's operating standard.

For every material conclusion:
1. Cite the section supporting it.
2. Separate documented facts from house assumptions.
3. Show conflicting information instead of resolving it silently.
4. If the answer is unsupported, write [NOT IN HOUSE FILE].
5. Never invent a number to complete an analysis.

This is the part most firms miss. Uploading documents gives Claude information. These rules tell it how to behave when the information is incomplete.

Step 3. Upload it to a Claude Project

Create a Project for your firm, strategy, market, or asset class.

Upload HOUSE_FILE.md into the Project knowledge alongside the files you repeatedly use, such as:

• Approved comp sets

• Previous investment memos

• Underwriting templates

• Expense benchmarks

• Lender quotes

• Market reports

Keep the House File small. It should contain the rules. The supporting files contain the evidence.

Step 4. Start every analysis with this prompt

Analyze this deal using the House File and Project documents.

For each important assumption, show:

1. The recommended input
2. Whether it comes from deal evidence or the House File
3. The exact supporting source
4. Any conflicting information
5. Anything marked [NOT IN HOUSE FILE]

Do not fill missing information with general market assumptions unless I explicitly ask you to.

Then attach the rent roll, T-12, OM, lease, or other deal documents.

Step 5. Test it before trusting it

Ask Claude a question your House File does not answer:

What renewal probability should we use for this tenant?

A grounded system should either cite your standard or return:

[NOT IN HOUSE FILE]

Then give it conflicting information:

The OM says market rent is $24, our broker comp says $21, and the latest executed lease is $20.50. Which figure should control?

Claude should apply your source hierarchy, explain the conflict, and cite the winning source.

If it simply chooses a number without showing its work, tighten the rules before using it on a live deal.

Step 6. Update one section after every deal

Do not rewrite the whole file.

Add only what the deal taught you:

## Lessons and Exceptions

### July 2026 — Phoenix Industrial
- Broker asking rents overstated executed rents by 9%.
- Use signed leases over asking comps when available.
- This is now a firmwide standard.

Five minutes after every deal is enough.

The goal is not to give Claude more documents.

It is to give Claude a clear set of rules for deciding what to trust, what to flag, and when to admit it does not know.

A quick note

Know a colleague whose analyst is still getting made-up cap rates back from a chatbot? Forward them this. That's how a lot of folks here found us. And refer 5 people and you get free access to the copy-and-paste Claude Skill Vault, the same grounded prompts we deploy for CRE clients.

Share

Want to work with us to get your firm up-to-date with AI?

Book a call with our team

About our team (glad you asked)

If you're new here, my name is Zed. I left my transactions role at PGIM to team up with software engineers from Goldman Sachs and HP building frontier AI for commercial real estate at PSV. My co-founders are deeply technical, with AI research funded by AWS and the National Science Foundation, and one of our advisors was Uber's first head of engineering, now an AI PhD at Berkeley.

Combined, our team has closed over $1 billion in CRE transactions. We're based in San Francisco, in the middle of the AI world, and most people gatekeep what we learn here. I'd rather translate it into quick 10-minute reads for fellow real estate people, for free, forever. (Fun fact: we've got robot dogs in our office!)

If this was useful, subscribe so you keep getting it, and follow us so we can keep making it.

Follow PSV on LinkedIn

Follow me on X

Quick summary

• 95% of orgs get zero return on generative AI, and only 5% of CRE firms hit their AI goals. The gap is context, not model size.

• Build Claude a House File: your trusted comps, expense standards, debt terms, submarkets, and underwriting rules, in one grounded Project.

• Add the rule that makes it honest: answer from the file, or write [NOT IN HOUSE FILE]. Cite or flag, never guess.

• Test it adversarially before a live deal. If it invents an answer, tighten the file.

• Feed it after every deal so the brain compounds. That feedback loop is the whole edge.

Also worth knowing this week

Construction budgets built on 2024 prices are quietly underwater. JLL's mid-year update (July 7) puts costs up roughly 5% year over year and accelerating, with steel, aluminum, and copper still at 50% tariffs. If you're pricing a capex plan or a ground-up deal, your GC's number has drifted. [2]

Anthropic extended free Fable 5 access again. The second bump in a week, landing right after OpenAI shipped GPT-5.6 and its Sol model on July 10. The price and access war keeps making the top-tier models cheaper to run on your deals. Stay model-agnostic. [3]

AI is now reshaping how North American CRE firms actually invest. A July 14 roundup profiles BGO timing exits across 2,500 submarkets, Hazelview baking internal data into underwriting, and Northmarq scoring tenant-renewal odds. 88% of firms are piloting AI; only 27% say rollout is going smoothly. [4]

Anthropic launched Claude for Teachers. Free premium access for verified US K-12 educators, with student data walled off from model training. Not a CRE story, but a tell for where "governed, data-protected" AI is heading, which is exactly the standard your vendor agreements should be catching up to. [5]

Anthropic put $10 million into Canadian AI research. A small line item that says the labs are still spending to widen the moat. The models under your workflows keep getting better whether or not you're paying attention, so the edge is in how you deploy them, not which one you pick. [6]

Sources

1. CRE AI Institute Publications (PSV)

2. Commercial Real Estate Needs a Company Brain, Not a Bigger AI Model (Commercial Observer)

3. JLL Report Details Rising Construction Costs in 2026 (JLL)

4. AI Model Wars: Anthropic Extends Fable Access Again After OpenAI's Sol Release (Forbes)

5. AI Adoption Reshapes CRE Strategies for North American Firms (CRE Daily)

6. Introducing Claude for Teachers (Anthropic)

7. Anthropic Commits $10 Million to Canadian AI Research (Anthropic)

Billionaire Travis Kalanick (Uber’s Founder) describes his new company as a computer:

Manufacturing is the processor.
Transportation is the network.
Real estate is the storage.

Not just an asset. But a crucial component inside a much larger machine. And one of the smartest real estate technology investors just wrote the check.

If you understand what they are actually buying, you are seeing the next asset class before the market has even named it. Most people are still debating whether AI-powered warehouses are even real…

Meanwhile, the capital has already moved.

Welcome to PropTech 2.0, this is volume 12 of Keeping Up With AI in CRE by the CRE AI Institute.

The 30-second version

On July 22, Fifth Wall, led by former Blackstone real estate investor Brendan Wallace, invested $135 million in robotic mining and autonomous transport. That is a meaningful shift away from the software that defined PropTech 1.0.

The signal for CRE investors is clear: in the physical-AI economy, the durable value may sit in the land beneath an operating business and the automation that makes that business safer, more productive, and ultimately more leasable.

This might be the next “roll-up” play.

1. Real estate and the physical AI economy.

On July 22, Atoms, Travis Kalanick’s holding company built on the bones of CloudKitchens, raised $1.7 billion in a round led by Andreessen Horowitz, with Ben Horowitz taking a board seat. It also acquired industrial automation company Pronto and is now targeting mining and heavy industry with autonomous machines.

The names in the round are the tell: Bain Capital, Uber, and Fifth Wall, the largest venture firm built specifically around real estate and the built world, backed by some of the country’s biggest owners and operators.

Fifth Wall did not invest in another property platform. It invested in the technology changing how physical assets are operated.

That does not mean commercial real estate investors should start buying mining companies. It means the next opportunity may sit at the intersection of real estate, infrastructure, and the automated businesses that occupy them.

Fun fact: one of our advisors is Uber’s first head of engineering and Travis Kalanick’s close friend. You can learn more about our team here.

2. The same dollar has two price tags.

A dollar of profit inside an operating business might trade at six to ten times earnings. Turn that same dollar into contractual rent under a long-term lease, and suddenly the market may value it at 15 to 18 times (high-level math).

That spread is the basic logic behind net lease investing. It is also why the debate over whether a data center is “real estate” or an “operating business” misses the point. The real question is which portion of the cash flow can be separated, contracted, and valued like rent.

The better question is: how much of the operating cash flow can be separated, contracted, and underwritten as rent?

But the separation itself does not create value. It only allows the market to recognize value that already exists. The business must first generate reliable cash flow. Then a credible tenant must agree to pay the landlord for years.

That second piece is where the physical-AI economy breaks the traditional net lease model.

3. The “credit” AI-first tenants don’t exist yet.

Robotaxi depots, autonomous truckports, and robotic mining sites may eventually become institutional net lease assets. But not yet.

The problem is not the real estate. It is the tenant credit.

A 15-year lease only deserves a low cap rate if investors believe the operator will still be paying in year 15. Most physical-AI companies are venture-funded, pre-profit, and dependent on future capital.

So the first deals will likely come with shorter leases, higher yields, corporate guarantees, throughput-based rent, equity kickers, or structures that combine real estate and operating income.

Senior housing solved a similar problem through RIDEA structures. The missing tenant credit looks like the risk.

Or maybe it is the entry.

4. The money isn't in owning it once it's safe.

By the time robotic-logistics real estate is stabilized, master-leased to an investment-grade name, and trading at a 5.5% cap in 2034, you're the buyer at the end of the line, not the winner.

The winner took operating-shaped risk at 9 to 11 percent going-in yields for the five to eight years while a dozen shaky operators consolidated into two or three names a bond desk will underwrite, and got re-rated the day the first public vehicle priced.

That's not some exotic new trade. It's the exact sequence data-center developers ran from roughly 2001 to 2015. Same structure, new tenants.

The trap on the other side is over-renting. Write the rent too high to juice the paper value, and you hand the operating business a cost structure it can't survive a bad year with. Sears carved its real estate out as Seritage in 2015 and was bankrupt by 2018. Red Lobster's sale-leaseback gets much of the blame for the bankruptcy that followed (or maybe it was the all-you-can-eat Shrimp deal).

The rent pulled the value out and the business died holding the bill. Bigger number, worse company.

Note: if you want to stay ahead of the technology curve in CRE with up-to-date AI training, an an AI-first community of vetted CRE professionals, apply below:

Apply for CREAII

5. The version you can actually run.

You're not writing a $1.7 billion check into Kalanick's robots. Neither am I. But the structure underneath that check has a small-dollar cousin, and it happens to sit right where a few of us already work.

Own the dirt under an operating business, lease it back to the operator, and separately hand that operator the AI that takes its labor cost down and its reliability up. Every point of margin you automate into that business is a point of rent coverage on your own lease. You're underwriting your tenant, then quietly making your tenant more bankable. That's an unusually clean alignment, and almost nobody is set up to do both halves of it.

The screen is one number: real estate value as a percent of what you paid. Pay 7 times earnings for an operator whose dirt alone is worth 80% of the price, and you got the business for almost nothing. Automate the boring, rules-based work until the coverage ratio is real, standardize it across sites, and you've turned a messy operating company into a lease a net-lease buyer will actually price. The automation isn't the whole trade. It's the permission slip that makes the cap rate real.'

Leave a comment

What I'd actually do this week

Before you write off some grubby operating business as "not real estate," run it through the property-company lens. Paste this into Claude, and turn on extended thinking so it shows the math:

you're a real estate and private equity analyst. i'm looking at [insert operating business or asset class, e.g. autonomous-truck maintenance depots, a regional laundry operator, robotic warehouse sites]. decompose a typical operator's EBITDA into (a) the implicit market rent the real estate would command leased to a third party and (b) the true operating profit left over. estimate rent coverage (EBITDAR / rent) at a conservative rent, then show how coverage changes if i automate [insert task] and cut opex by [insert %]. tell me real-estate-value-as-a-percent-of-purchase-price at [insert multiple]. flag the rent level where coverage falls under 1.5x. cite a source for any market figure, and write [NOT VERIFIED] where you're estimating. don't invent numbers.

Good output is a coverage table you can sort, a plain verdict on whether this is mostly real estate or mostly operations, and a set of [NOT VERIFIED] cells telling you exactly what to confirm before you trust any of it. Verify those first.

Run the toy version and the lens does the talking. A $10M-earnings operator bought at 7x is a $70M business on paper. Split the dollar into rent and operating profit, lift the coverage with automation, and the rent stream alone re-rates toward numbers the blended multiple was hiding. Illustrative, not a promise. The point is the lens, not the figure.

That prompt is the back-of-the-napkin version. The full property-company screen, the coverage math, the automation-to-credit bridge, and the structure library (sale-leaseback versus born-split versus the run-it-yourself manager answer) is one of the systems we build inside the CRE AI Institute. Same deal you were about to call "not real estate," run through the lens that actually prices it.

Join the CRE AI Institute

A quick note

Know a colleague who still argues "that's an operating business, not real estate" like the two are enemies? Forward them this issue. That's honestly how most people here found us. And if you refer 5 people, you get free access to our copy-and-paste Claude Skill Vault, the exact prompts we run for CRE clients.

Share

Also, you can keep reading about where AI and real estate collide every week, or you can join the room actually running the plays: the training to do it on your own deals (20+ modules plus the full prompt, skills, and plugins library), a weekly office hour to get unstuck, and a vetted room of brokers, operators, and investors building the same way. For anyone eyeing a move, the firms hiring AI-fluent CRE talent come to us first. The 7-day free trial is open, so if you're serious about AI, apply below.

Join the CRE AI Institute

About our team (glad you asked)

If you're new here, my name is Zed. I left my transactions role at PGIM to team up with software engineers from Goldman Sachs and HP building frontier AI for commercial real estate at PSV. My co-founders are deeply technical, with AI research funded by AWS and the National Science Foundation, and one of our advisors was Uber's first head of engineering, now an AI PhD at Berkeley.

Combined, our team has closed over $1 billion in CRE transactions. We're based in San Francisco, in the middle of the AI world, and most people gatekeep what we learn here. I'd rather translate it into quick 10-minute reads for fellow real estate people, for free, forever. (Fun fact: we've got robot dogs in our office!)

If you found this helpful, subscribe so you keep getting it, and follow us so we can keep making it.

Follow PSV on LinkedIn

Follow me on X

Sources

1. CRE AI Institute Publications & Research.

2. Travis Kalanick's robotics company raises $1.7B, led by a16z (TechCrunch)

3. Kalanick's Atoms raises $1.7B in a16z-led round for industrial AI (Yahoo Finance)

Last Tuesday, New York ran that play on the hottest asset class in real estate. The plant is the power grid. The hookup is a 50-megawatt interconnection. The study covers the entire state.

Most coverage filed it under tech news. It's a land-use story, and if you read it that way this week, congratulations, you’re one of the first.

Welcome. This is Volume 11 of Keeping up with AI in CRE, by the CRE AI Institute.

The 30-second version

On July 14, 2026, Governor Kathy Hochul signed Executive Order No. 62, the first statewide moratorium on new hyperscale data centers in the country. An executive order, not legislation, effective immediately. New York's environmental agency will now sit on state permits for any new data center that consumes, or could consume, 50 megawatts or more (roughly the draw of 40,000 homes) while the state writes new standards for energy, water, and community benefits. Factories, universities, hospitals, and research facilities are exempt, and anything already approved keeps building.

The hottest asset class in real estate just met the oldest force in real estate: the permit office.

The order's own text names the real fight: nearly 12 gigawatts of data center requests sit in New York's grid queue, which is the polite way of saying a small country asked to plug in. Albany says the pause runs "up to one year." The order itself names no end date, and the gap between those two sentences is where your underwriting lives.

1. The whole order turns on a date stamp

The order does one narrow thing. It tells the Department of Environmental Conservation to hold discretionary permits for covered projects, but only where the application was not deemed complete before July 14. Deemed complete is the date stamp at the permit counter. Got the stamp before the window closed? You build. Still assembling exhibits? You wait. Local zoning is untouched, and facilities primarily for manufacturing, research, education, or medical care are carved out entirely.

The soft spot is duration. The press release says "up to one year." The order ties the end to a statewide environmental study with no date attached, the kind of open-ended "when the study is done" that anyone with a sewer moratorium story has learned to distrust; Jones Day flagged the gap within days. In lease terms, a holdover with no outside date. The reason Albany reached for that open end showed up six days later, with a dollar sign on it.

2. Who pays for the plant

That dollar figure sits next door. On July 15, PJM, the grid operator covering 13 states plus DC, cleared a $16.4 billion capacity auction for 2028-29. Five days later its independent market monitor pinned about $6.3 billion of that (38 percent) on data center demand, and $29.4 billion across the last four auctions. Wholesale power prices in PJM jumped 75.5 percent between Q1 2025 and Q1 2026, with data center load growth named the primary cause; the monitor calls it "really a paradigm shift."

The order states its reasoning plainly: those 12 gigawatts of queued demand should not get wired up on everyday New Yorkers' bills. A capacity auction is how that plant expansion lands on the whole town's tab anyway. Maybe what a state really sells a data center was never land at all. It's headroom on the grid, priced before it gets promised.

By the way post is public so feel free to share - it would mean a lot to our team. We all put in tons of work into writing free content for you!

Share

3. The moratorium has a copy machine

New York signed first, but the paperwork was already circulating. Maine's legislature passed a 20-megawatt version; Governor Mills vetoed it on April 24, 2026. Denver approved a one-year citywide pause in mid-May. By June the Rockefeller Institute counted 14 states with moratorium bills, one law firm had tallied 300-plus data-center bills in 30 statehouses in the first six weeks of 2026, and a tracker cited by Columbia's Sabin Center counts over 100 local moratoria.

The quieter version is already national: 24 states had approved special large-load utility tariffs by July 2026, several letting a utility lock big customers into 12- to 15-year terms and bill them for capacity whether they use it or not (Ohio cleared AEP's version in July 2025; Georgia authorized terms up to 15 years for its 100-megawatt class back in January 2025).Hookup fees, by another name.

Washington noticed. The President called the freeze "a terrible decision" and data centers "LIQUID GOLD" on Truth Social; Hochul replied that New Yorkers "deserve more than scraps."

All this demand is other people's AI. If you're less sure where your own AI knowledge stands, join over 100 institutional-grade CRE professionals at the CRE AI Institute.

Level Up Your AI Skills for CRE

4. You will never own a data center. Your deals noticed anyway.

The same boom Albany just paused was already inside your numbers. Contractors with data center work carry 12.2-month backlogs against 8.3 months for everyone else (JLL, mid-2026), while everything that isn't a data center (office, industrial, mixed-use) grows below 1 percent in real terms. Land near power that went for $10,000 to $30,000 an acre before the boom now trades at $200,000 to $1 million (Bisnow, January 2026). Your GC's pricing and your power line item: same boom, different invoice.

So a statewide pause helps you? Partly. It also kneecaps the one construction segment that was hiring. The national plumbers and pipefitters union says the order "kills good-paying union jobs," and Construction Dive notes data centers were construction's golden goose in 2026. The industry's counterpunch lands too: blocked compute gets built anyway, somewhere else, often on a dirtier grid. Cheaper bills and fewer cranes.

5. The new line on the diligence checklist

Which brings us to what you write down. Foley & Lardner called the order a potential "first domino" and now lists regulatory risk as a key development-planning factor. PitchBook's analyst put it sharpest: New York converted the constraint "from a physics problem into a political one."

Steal the question for every deal, not just the 50-megawatt kind. Before you underwrite anything power-hungry, or anything next to something power-hungry: what is this utility's large-load tariff, is there a moratorium bill in this statehouse, and who eats the upgrade if the rules change mid-construction?

Your grandfather's developer asked the same thing at the town counter. Who pays for the plant, and did my application get the stamp before the window closed?

For CRE pros serious about AI

Know a colleague who learns about their market’s utility tariff from the invoice? Forward them this issue. That’s how most people here found us. And if you refer 5 people, you get free access to the copy-and-paste Claude Skill Vault.

We just opened a 7-day free trial of the CRE AI Institute, so you can walk the halls before committing to anything. Inside: 20+ modules on running AI on your own deals, the full prompt, skills, and plugins library, weekly office hours, and a vetted room of AI-forward CRE operators, quarterly in person.

Serious about AI? Apply below.

Join the CRE AI Institute

This week's Claude move

This whole issue asked one question: who pays for the plant. Your share of the answer is already sitting in your own T-12s. Everyone knows Claude can read a PDF. Fewer know it can take a whole stack of spreadsheets in one chat and run real arithmetic across them: not summarize, compute.

Export the trailing-12 for every property you run, drop them into Claude, and paste this:

you're an asset management analyst for a CRE owner. i've attached T-12s for [insert number] properties. for each property, pull every utility line item (electric, gas, water) and build one table: property, trailing-12 utility spend, utilities as a % of total opex, and $ per unit or per SF. rank the portfolio by electric exposure. then model electric costs at +10% and +25% and show the NOI impact per property in dollars. cite the file, tab, and row for every number you pull. if utilities are buried in CAM or a number isn't in the sheet, write [NOT IN SHEET], don't guess.

Good output is a ranked table where every figure traces back to a file and row, with honest [NOT IN SHEET] flags wherever a ledger hides utilities inside recoveries. Verify it by tying the top-ranked property back to its source workbook by hand before you show anyone. The payoff is the afternoon of rekeying a dozen workbooks you didn't spend, and this issue stops being news and becomes a number with your name on it.

That prompt is the pocket version of the portfolio power screen. The full version (the tariff and moratorium pulls for the markets you actually buy in, the scenario table wired to your actual recovery clauses, and the verification pass that makes Claude flag instead of guess) is one of the systems we build inside the CRE AI Institute.

About our team (glad you asked)

If you're new here, my name is Zed. I left my transactions role at PGIM to team up with software engineers from Goldman Sachs and HP building frontier AI for commercial real estate at PSV. My co-founders are deeply technical, with AI research funded by AWS and the National Science Foundation, and one of our advisors was Uber's first head of engineering, now an AI PhD at Berkeley.

Combined, our team has closed over $1 billion in CRE transactions. We're based in San Francisco, in the middle of the AI world, and most people gatekeep what we learn here. I'd rather translate it into quick 10-minute reads for fellow real estate people, for free, forever. (Fun fact: we've got robot dogs in our office!)

If you found this helpful, subscribe so you keep getting it, and follow us so we can keep making it.

Follow PSV on LinkedIn

Sources

1. CRE AI Institute - AI Research and Publications.

2. Executive Order No. 62: Establishing a Temporary Moratorium on Data Centers in New York (Governor Kathy Hochul)

3. First Statewide Moratorium on New Hyperscale Data Centers Launched by Governor Kathy Hochul (Governor Kathy Hochul)

4. New York Becomes First State to Impose Data Center Moratorium (Broadband Breakfast)

5. New York Imposes Moratorium on Hyperscale Data Center Development Under Executive Order 62 (Harter Secrest & Emery)

6. New York Pauses Permits for Largest Data Centers While Hochul Weighs Broader Bill (amNY)

7. New York Enacts First Statewide Data Center Moratorium (Jones Day)

8. Data Centers Drove $6.3B in PJM Capacity Auction Costs: Market Monitor (Utility Dive)

9. Datacenters Slurping Up So Much Juice They Boosted Prices 75% in Largest US Energy Market (The Register)

10. Governor Mills Announces Decision on LD 307 (Office of Governor Janet Mills)

11. Denver City Council Approves Data Center Moratorium (Government Technology)

12. Updates on the Cloud: More Moratoriums on Data Centers (Rockefeller Institute of Government)

13. State Regulation of Data Centers in 2026: A Shifting Landscape (ArentFox Schiff)

14. Local Moratoria Against Data Center Construction: Considerations for Municipal Governments (Columbia Law School Sabin Center)

15. Large Load Projects and Tariffs, July 2026 (Edison Electric Institute)

16. Regulator Approves AEP Ohio's Landmark Data Center Tariff (Power Magazine)

17. PSC Approves Rule to Allow New Power Usage Terms for Data Centers (Georgia Public Service Commission)

18. Trump Calls New York's Data Centre Pause a 'Terrible Decision'. Hochul Is Not Moving. (The Next Web)

19. Construction Costs Rise in 2026: JLL Report Identifies Strategic Procurement Opportunities (JLL)

20. If You Entitle It, They Will Come: How Powered Land Plugs CRE Into The Data Center Boom (Bisnow)

21. New York Enacts Nation's First Statewide Temporary Ban on New Data Centers. Here's Why. (CBS New York)

22. New York Contractors Decry State's Data Center Moratorium, the Nation's First (Construction Dive)

23. New York Becomes First State to Impose Moratorium on Data Center Development (ESG Dive)

24. New York's AI Data Center Moratorium: A Potential Turning Point for U.S. Data Center Development (Foley & Lardner)

25. NY's Data Center Freeze Adds a New Risk for AI Investors: State Regulation (PitchBook via Yahoo Finance)

But CRE firms routinely give software vendors access to rent rolls, leases, T-12s, and tenant data without asking the same questions.

A big AI deal that closed this week showed why these terms are becoming a real diligence issue - and why a digital key may open more than you think.

Welcome. This is Volume 10 of Keeping up with AI in CRE.

The 30-second version

On July 14, RealPage (Richardson, Texas, multifamily software and data analytics, sitting on 25 petabytes of data) acquired Cherre, the Manhattan company founded in 2016 whose platform unifies and analyzes more than 4 billion entities and $4 trillion in real assets globally. Cherre keeps its brand and its Manhattan HQ, and co-founder L.D. Salmanson becomes Senior Vice President. What’s the price? Well when you ask a broker for comps, sometimes you get it, and sometimes you don’t.

RealPage President and CEO Dirk Wakeham declined to provide specific figures.

A software company buying a data company isn't news. What Cherre sells is: it connects, resolves, and governs fragmented data to create "a trusted foundation for AI, analytics, and decision-making."

Governs. That word is about to do a lot of work, thanks to something filed in a North Carolina courtroom last November that most of the coverage walked right past...

1. RealPage bought the layer that sits above its own data

25 petabytes is a warehouse. Cherre is the thing that tells you what's in the warehouse, where every box came from, and who's allowed to open which ones.

Wakeham said it plainly. "The opportunity for the 1 plus 1 equals 3 here is pointing the Cherre technology...against our 25 petabytes of data."

The stated logic is a bridge. Property-level data on one side, portfolio and fund-level context on the other, AI running across property management, portfolio oversight, and capital allocation. On paper, a sensible trade. And if you're a RealPage customer, your next renewal call just got more interesting than another argument about the renewal number. The math checks out fine. The open question is: whose data?

2. The pipes were always the #1 asset

Everyone wants to talk about the best AI model. In real estate terminology, the model is the tenant. The data layer is the building.

Cherre's pitch is that real estate data stays useless until somebody connects it, resolves it (this entity and that entity are the same LLC), and governs it (this field came from here, and it may be used for that). That's provenance: the boring plumbing under every AI answer a CRE product will ever show you, and the difference between a platform that can tell you where a number came from and one that just hands you the number.

A governed data layer is what makes an AI output defensible in front of an investment committee. It's also what makes your data legible, portable, and useful to somebody who is not you.

So the next time a vendor demos an AI feature, ask one unglamorous question: where did this number come from. If the answer is a shrug and a roadmap, you're renting a room in someone else's building. Ask for the floor plan.

Not sure where you stand with AI and data?

Take AI Readiness Quiz

3. The part sitting in the court file

In November 2025, the DOJ filed a proposed settlement resolving antitrust claims involving certain RealPage rental-pricing products.

RealPage did not admit liability, and the settlement was not a finding of guilt.

But the proposed restrictions were specific. They limited the use of nonpublic, competitively sensitive landlord data and restricted what information could be used to train certain models. Compliance would be reviewed by a court-appointed monitor.

Then RealPage acquired a company whose core product is tracking where data came from and governing how it may be used.

That does not automatically make the acquisition suspicious. It may be the opposite.

When you must prove that certain data was not used for training, serious data-governance infrastructure becomes essential.

But the larger lesson applies to every CRE firm:

“What is this vendor allowed to do with our data?” is no longer an abstract technology question. It can become a contractual, regulatory, and diligence issue.

Your software vendor already has the key. Do you know which rooms it opens?

4. What your agreement actually says about your rent roll

Every operator hands a vendor the same pile. Rent roll, leases, T-12, unit-level pricing, tenant PII, maintenance history. Then we spend three weeks negotiating the fee and sign the data terms unread, because they live in the exhibit at the back that nobody opens.

The clauses that matter are almost never the scary-sounding ones. They're the polite ones. "Aggregated and anonymized." "Service improvement." "Derived data." Each is doing enormous work.

Walk just the first one. Your rent roll goes in. It gets stripped of the unit numbers and your entity name, which makes it "anonymized." It gets pooled with data from a few thousand other properties, which makes it "aggregated." What comes out the other side is a market benchmark, and that benchmark is a product. It can be sold. Even to the group bidding against you on Tuesday. Usually in the contract you signed. It survives termination, which means you can fire the vendor and your data keeps working there.

Of course, none of that is an accusation against RealPage or anyone else. Just standard boilerplate, across the whole category.

A quick note before the takeaway

I run a community of AI-first CRE professionals where we go over how to build production-ready AI agents for manual CRE work that consultants charge $25,000 for each.

Last week, we helped members replace their five-figure deal pipeline dashboard in under 2 hours with Claude Code.

Plus there’s over 95 AI in CRE lessons, weekly live office hours on your own deals, and a vetted room of brokers, operators, developers, and asset managers doing the same. Plus a vetted AI tool database, over 50 Claude skills/prompts for CRE, and a resource vault with copy-and-paste custom builds with Claude.

For those of you who are looking for new opportunities, the firms hiring AI-fluent talent come to us first.

We’re opening up a 7-day free trial for qualified candidates, so if you’re serious about AI, apply below!

Join the CRE AI Institute

5. Three questions worth asking your vendor now

None of these are gotchas. A good vendor answers all three without checking with legal.

1. What of my data can I export in full, in a usable format, without a support ticket?

2. What are you permitted to do with it, and does it feed anything that serves someone I compete with?

3. If I leave in 2028, what comes with me and what stays behind?

Ask before the renewal quote lands. The rep who goes quiet on number two has already answered it.

What I'd actually do this week

Pull the agreement, the DPA, the order form, and any privacy exhibit for the vendor sitting on the most of your data. Paste this into Claude:

you're a vendor-contract analyst for a commercial real estate owner. i'm pasting our agreement with [insert vendor name], plus any DPA, order form, and data/privacy exhibit. work only from these documents.

build a table with one row per category of data we hand them ([insert: rent roll, leases, T-12, unit-level pricing, tenant PII, maintenance logs]) and these columns: what they may USE it for, how long they may RETAIN it, whether they may AGGREGATE it with other customers' data, whether they may ANONYMIZE or de-identify it and what that permits them to do afterward, whether they may TRAIN models on it, whether any of those rights SURVIVE termination, and what we get back or get deleted when we walk.

quote the exact clause and section number in every cell. if the documents don't say, write [NOT IN DOCS]. don't infer from industry practice, don't fill a gap with what's typical, don't soften.

then list the five [NOT IN DOCS] gaps that would cost us the most if the answer came back yes, and draft the email asking the vendor to answer them in writing.

The [NOT IN DOCS] rows are why you run it. A gap in a contract isn't a comfort. It's an option, and the vendor holds it.

[VISUAL: screenshot of Claude returning the data-rights table with two or three [NOT IN DOCS] rows visible]

That prompt is the pocket version. It reads the four documents you happened to have on your desktop. The full vendor data-rights audit (the document list most owners don't know to request, and the clause library of what "aggregated and anonymized" actually permits downstream) is one of the systems we build inside the CRE AI Institute. Same contracts you already signed, with the lights on.

Spread the word!

Know a colleague who signed a platform agreement last quarter and couldn't tell you what's in the data exhibit? Forward them this issue. That's how a lot of people here found us. And if you refer 5 people, you get free access to the copy-and-paste Claude Skill Vault, structured by AI engineers at PSV.

Share

About our team (glad you asked)

If you're new here, my name is Zed. I left my transactions role at PGIM to team up with software engineers from Goldman Sachs and HP building frontier AI for commercial real estate at PSV. My co-founders are deeply technical, with AI research funded by AWS and the National Science Foundation, and one of our advisors was Uber's first head of engineering, now an AI PhD at Berkeley.

Combined, our team has closed over $1 billion in CRE transactions. We're based in San Francisco, in the middle of the AI world, and most people gatekeep what we learn here. I'd rather translate it into quick 10-minute reads for fellow real estate people, for free, forever. (Fun fact: we've got robot dogs in our office!)

If you found this helpful, subscribe so you keep getting it, and follow us so we can keep making it.

Follow PSV on LinkedIn

Follow me on X

Sources

1. RealPage Acquires Cherre, Creating a Trusted AI-Powered Intelligence Platform Across the Full Real Estate Capital Stack (BusinessWire)

2. RealPage Acquires Cherre in Major Proptech Deal (Commercial Observer)

3. RealPage buys AI-focused real estate data firm Cherre (Multifamily Dive)

4. DOJ and RealPage Agree to Settle Rental Price-Fixing Case (ProPublica)

5. DOJ Settles Its Algorithmic Price-Fixing Case Against RealPage (Wilson Sonsini)

Last week a former $10B Bitcoin miner proved it again. It stopped chasing the shiny thing, became a landlord instead, and the name on the 20-year lease was the same AI company half of you run your deals through. If you can see where this is going before the guy across the closing table does, you're still early. Most of your market is still arguing about whether any of this is "real."

Welcome. This is Volume 9 of Keeping up with AI in CRE.

The 30-second version

Anthropic, the company that makes Claude, just signed a 20-year lease on a 401-megawatt data center in Kentucky worth about $19 billion in rent, and Morgan Stanley is raising $3.5 billion to build it.

Below is the picture of the site (The former Century Aluminum smelter in Hancock County of Ohio).

_{Picture Source: Yahoo Finance}

The AI story you keep hearing about is no longer a tech story this week and became a commercial real estate story. The demand is signed, financed, and two decades long.

But there’s a lot more to the story, some of which a big surprise, that any CRE professional should know…

1. Anthropic just became one of the biggest new tenants in the country.

TeraWulf used to mine Bitcoin. Last week it announced a $3.5 billion debt raise, led by Morgan Stanley, to build a 401 MW campus in Hawesville, Kentucky, anchored by a 20-year lease with Anthropic worth roughly $19 billion in contracted rent. First real capacity lands in late 2027, full build by early 2028.

A crypto miner looked at its own business, put down the pickaxe, and became a landlord to an AI company on a two-decade lease. Maybe durable money isn't in the model, it's in the dirt, the power, and the paper underneath it. When a bank underwrites 20-year AI rent as investment-grade income, this is no longer speculative.

So what do you do? If you touch land, industrial, or anything power-adjacent, start treating "proximity to cheap, deliverable power" like you treat proximity to a highway interchange. It just became a true value driver.

2. Every AI data center drags a warehouse in behind it.

Bloomberg reported on July 7 that data-center fever is spilling into the warehouses next door. Meta, Google, and the other hyperscalers are leasing industrial space around their campuses to stage transformers, servers, cooling gear, and construction materials. The rule of thumb the analysts are using: roughly 2 million square feet of spillover industrial demand for every gigawatt of data-center build. With about 100 gigawatts in the national pipeline, that's around 200 million square feet of industrial demand over the next five years.

This isn't a blip. Cushman & Wakefield projects AI will add about 330 million square feet of net-new US real estate demand this decade, and industrial carries the load at +298 million square feet. The takeaway for your deal might be that the data center itself is a specialist asset most of us will never own. However, the warehouse two exits away that ends up leased to a hyperscaler's contractor? That's a plain-vanilla industrial deal you already know how to underwrite. Follow the campuses and read the second-order demand.

3. Smart AI got cheaper again (there's now a $1 tier).

On July 9, after a government capability review, OpenAI put its GPT-5.6 family out to the public: Sol at the top ($5 in / $30 out per million tokens), Terra in the middle ($2.50 / $15), and a budget Luna tier at $1 in / $6 out. That's not an "OpenAI beat Claude" story. It's a price-war story, and the price war is the gift. Every few weeks, near-flagship reasoning gets cheaper across every lab.

Why does a Claude newsletter care? Because it means the cost of running an underwriting pass, a diligence read, or a market-scan agent keeps falling no matter whose logo is on it. The lesson is the same one I keep repeating: don't marry a model. Own the workflow, keep a workhorse, and let the labs fight over who gives you the cheapest smart tokens this month. You win either way.

4. Your AI analyst now works from your pocket.

Anthropic brought Claude Cowork, the "hand it a task and it works in the background" agent, to web, iOS, and Android on July 7. The number that matters: Anthropic says more than 90% of Cowork use is NOT software development. The biggest categories are business operations and content, things like reconciling the quarter's spend and drafting the variance memo, or turning a folder of contracts into a renewals tracker with the risks flagged.

Sound familiar? That's the job description of half your analyst bench. The unlock now is that it's cross-device. You can kick off "normalize this rent roll and flag the outliers" at your desk, close the laptop, and approve the result from your phone between property tours. Don't be the person still copy-pasting docs at 9pm. Hand off one bounded task this week and watch it run without you.

5. Wall Street got nervous about the brokers.

Here's the honest counterweight. Brokerage shares, CBRE, JLL, Cushman, Newmark, took a hit this month as investors started pricing in the fear that AI compresses advisory fees. The exposed layer, per the executives themselves, is the repetitive, public-data work: appraisals, routine research, smaller cookie-cutter transactions. The high-touch, relationship-heavy, complicated deals? Those still need a human who can read a room.

So the question isn't "will AI replace brokers." It's "which kind of broker are you becoming." The one whose value was pulling comps and formatting a BOV is exposed. The one who uses AI to do that in ten minutes and spends the saved hours in front of clients gets stronger. Same tool, opposite outcome. Choose on purpose.

What I'd actually do this week

Before you underwrite your next industrial or land deal, pressure-test whether it sits in the path of AI-driven demand. Paste this into Claude:

you're a CRE market analyst. i'm looking at [asset type] at [address / submarket]. tell me whether this site is positioned to benefit from data-center-driven demand: nearby existing or announced data center campuses, power availability and major substations, transmission access, and any recent hyperscaler or contractor leasing in the area. rank the demand tailwind low / medium / high and explain why. cite a source for every factual claim. if you can't verify something, write [NOT VERIFIED] instead of guessing.

Then go verify the [NOT VERIFIED] items yourself. Claude points the flashlight. You still walk the room.

A quick note

Know a colleague still doing comps by hand and forwarding PDFs at midnight? Forward them this issue. That's genuinely how most people here found us. And if you refer 5 people, you unlock free access to our copy-and-paste Claude Skill Vault, the exact prompts and workflows we deploy for CRE clients.

PSV also runs the CRE AI Institute, the #1 AI training program for commercial real estate: 20+ modules, the full prompt/skills/plugins library, weekly office hours, a vetted network of AI-forward CRE pros, and quarterly in-person events. The first 50 members lock in a founding-member rate forever plus 1:1 time with our team.

Join the CRE AI Institute

About our team (glad you asked)

If you're new here, my name is Zed. I left my transactions role at PGIM to team up with software engineers from Goldman Sachs and HP building frontier AI for commercial real estate at PSV. My co-founders are deeply technical, with AI research funded by AWS and the National Science Foundation, and one of our advisors was Uber's first head of engineering, now an AI PhD at Berkeley.

Combined, our team has closed over $1 billion in CRE transactions. We're based in San Francisco, in the middle of the AI world, and most people gatekeep what we learn here. I'd rather translate it into quick 10-minute reads for fellow real estate people, for free, forever. (Fun fact: we've got robot dogs in our office.)

If you found this helpful, subscribe so you keep getting it, and follow us so we can keep making it.

Follow PSV on LinkedIn

Sources

1. TeraWulf (WULF) Stock Surges on Morgan Stanley-Led $3.5B Financing for Anthropic AI Facility (Parameter), https://parameter.io/terawulf-wulf-stock-surges-on-morgan-stanley-led-3-5b-financing-for-anthropic-ai-facility /

2. Data-Center Fever For Real Estate Spreads to Adjacent Warehouses (Bloomberg), https://www.bloomberg.com/news/articles/2026-07-07/tech-giants-data-center-fever-spreads-to-adjacent-warehouses ; Cushman & Wakefield: AI to Add 330 Million Square Feet of CRE Demand Over Next Decade (Cushman & Wakefield), https://ir.cushmanwakefield.com/news/press-release-details/2026/Cushman--Wakefield-Report-AI-to-Add-330-Million-Square-Feet-of-CRE-Demand-Over-Next-Decade/default.aspx

3. OpenAI Sets GPT-5.6 Pricing at $5 Input, $30 Output per 1M Tokens With Three-Tier Model Family (Crypto Briefing), https://cryptobriefing.com/openai-gpt-56-pricing-tiers/

4. Claude Cowork comes to web and mobile (Anthropic), https://claude.com/blog/cowork-web-mobile

5. AI Anxiety Shakes Brokers (CRE Daily), https://www.credaily.com/briefs/ai-anxiety-shakes-brokers/

Preface

Hi, it’s Zed again, bothering you with another AI in CRE tip. We’re starting the week hot with Micron’s AI infrastructure deal with Anthropic (and a series H investment). This article though, might be a bit hotter for us all in the real estate industry.

Welcome. This is Volume 8 of Keeping up with AI in CRE.

The trap everyone's walking into

I see all of this hype over LinkedIn: "Claude, underwrite this multifamily deal. Make no mistakes." It spits out rent growth, an exit cap, a lease-up curve, expense ratios, a return. Looks rigorous. Feels rigorous. The future is here!!

But no. A model that fills in your assumptions just agrees with you in spreadsheet form. The numbers it "found" are the numbers it guessed, dressed up as analysis. That's a hype-man, not a lawyer. A hype-man tells you the deal is fire.

A good lawyer reads the same deal and says "here are the four things that get you sued." You want the lawyer. I guess you want someone like Harvey from Suits.

So flip the job. Don't ask Claude to build your underwriting. You already did that. Point it at the model you built and tell it to be the investment committee from hell. Its whole job is to find every reason your deal fails, and to attack YOUR assumptions, not invent its own.

Step 1. Hand it your assumptions, not a blank chat.

The adversary needs your actual numbers to attack. Paste in your assumptions exactly as you underwrote them:

• In-place rent and your year-1 to year-3 rent growth

• Your exit cap rate (and the going-in cap for reference)

• Lease-up or stabilization timeline in months

• Expense ratio and any line items you trended

• Loan terms, going-in and stabilized debt service

If Claude has to guess these, you're back to the yes-man. Give it the real ones. Now it has something to bite.

Step 2. Tell it to be a hostile IC, not a helpful assistant.

This is the whole game, and it's a one-line instruction: "You are an adversarial investment committee. Your job is to kill this deal, not approve it." Watch the output completely change Claude’s character.

Think of it like buying a used car from your brother-in-law versus a stranger. Your brother-in-law tells you it runs great. You say you agree.

The stranger you're about to wire $40k to gets the cold, suspicious once-over, the flashlight under the chassis, the "why are you really selling." Your own deal deserves the stranger treatment, because the market is the stranger.

That might be a little confusing, but I tried my best.

Step 3. Make it attack the assumptions that actually move money.

A bad bear case nitpicks the parking ratio. A good one goes straight for the assumptions your return is built on. Force AI there:

• Rent growth. You penciled 3.5% a year. On what comps, in what submarket, against what supply pipeline? What if it's 2%?

• Exit cap. You're exiting tighter than you bought. Why would a buyer pay up in three years? What if rates don't cooperate and you exit 50 to 75 bps wider?

• Lease-up timeline. You assumed 12 months to stabilize. What breaks if it's 18? What's your concession burn in the meantime?

• Expense ratio. Did you trend insurance and taxes at 3% in a market where insurance is up double digits? What's the real number?

The point is not to be a doomer. The point is that these four lines are where deals actually die, and a hype-man will never poke them.

Step 4. Rank the risks by how much they cost you if you're wrong.

This is the step that separates a worry-list from underwriting. A list of 15 general risks is useless. You need to know which one wrecks the deal. So make Claude run the sensitivity: for each assumption it attacks, how many points of IRR or equity multiple do you lose if it's wrong?

This is triage. When you walk into the ER, they don't treat patients in the order they arrived. They treat the one bleeding out first. Your exit cap might cost you 600 bps of IRR if it moves, while your expense ratio costs you 40 (common sense of course).

Make sure you tell Claude to rank them, stress-test, and run evals for everything.

Bonus: Make it steel-man the bear case (the part everyone skips).

A cheap critic throws twenty objections and hopes one sticks. You don't want volume, you want the strongest version of "this deal is bad." So tell Claude: "Now build the single most credible bear case. Not the scariest. The most likely to actually happen."

This is the difference between a heckler and a real opponent. A heckler yells anything. A worthy opponent makes the one argument you can't easily answer, the one that keeps you up at night. If you can beat the steel-manned bear case, you have a deal. If you can't, you just saved yourself a very expensive education.

Here's the prompt I use. Make sure to paste your assumptions in first.

you are an adversarial investment committee reviewing a deal i want
to approve. your job is to find every reason it fails. do not be
polite and do not fill in missing assumptions with optimistic
defaults.

here are MY assumptions: [paste in-place rent, rent growth by year,
going-in and exit cap, lease-up timeline in months, expense ratio
and trended line items, loan terms]

do this in order:
1. attack each of my assumptions above. for each, say what i'd have
   to believe for it to hold, and what a skeptical buyer would assume
   instead.
2. rank the assumptions by impact: if i'm wrong on this one, roughly
   how much IRR or equity multiple do i lose? order them worst-first.
3. build the single most CREDIBLE bear case (most likely, not
   scariest) using my numbers.
4. then adjudicate: which attacks are real and diligence-able vs noise.
   output a numbered diligence list i can act on this week.

cite the source for any external number (comp, cap rate, supply
figure). if you don't have a source, write [NOT IN DOCS] instead of
estimating. do not make up data.

Quick Note

If you want the full version of this, with the adversarial-IC prompt as a saved skill, plus 20+ modules on running deals through Claude without it lying to you, that lives in the CRE AI Institute, the #1 AI training program for commercial real estate.

You get the prompt and skills library, weekly office hours, a vetted CRE network of tech-forward professionals, and quarterly in-person events. The first 50 members lock a founding-member discount forever and get 1:1 implementation time.

Very unbiasedly, I'd tell you it pays for itself the first time it talks you out of one bad deal :)

Join the CRE AI Institute

A bonus on the bonus: Adjudicate into a diligence list you can act on this week.

The adversary's job is to attack. Your job is to judge. Don't accept the bear case wholesale and don't dismiss it because it stings. Sort every objection into two buckets:

• Real and diligence-able. "Verify the insurance quote, you trended it too low." Good. That's a phone call this week.

• Noise. "What if a meteor hits the submarket." Thanks, discard.

What survives is a grounded diligence list: the three or four things you genuinely don't know yet, ranked by how much they move your return. That's not a model that agreed with you. That's a model that made you earn your conviction. A deal you can defend to a hostile room is a deal you can defend to your LPs.

A 30-second checklist before you wire anything

• Did you hand Claude YOUR assumptions, not let it invent them?

• Did you tell it to kill the deal, not approve it?

• Did it attack rent growth, exit cap, lease-up, and expenses specifically?

• Did it rank risks by impact on your return, worst first?

• Did it steel-man the most credible bear case, then adjudicate it?

Wow, you read the whole thing. Top 10% of readers, you'll go far. Want to know where your shop actually stands on this stuff?

Take AI Readiness Quiz

And if you'd rather we just set the adversarial workflow up with you, book a call with our team.

Work With PSV

About Our Team (glad you asked)

If you're new here, my name is Zed. I left my transactions role at PGIM to team up with software engineers from Goldman Sachs and HP and build PSV. My co-founders are Gedeon Baende (previously Goldman Sachs) and Aidan Scudder (previously Kochava). Our AI research has been funded and recognized by AWS and the National Science Foundation (two papers, both pre-transformer). One of our advisors was Uber's first head of engineering, now an AI PhD at Berkeley. We're backed by Jason Calacanis' LAUNCH and Founders, Inc., and we're based in San Francisco.

Combined, our team closed over $1B in CRE transactions. Fun fact: we got robot dogs in our office.

Most people gate-keep this stuff to stay ahead. I've decided to give the basics away to fellow CRE professionals, for absolutely no cost, forever.

Oh, and refer 5 people and you get free access to our copy-and-paste Claude Skill Vault, the adversarial-IC prompt included.

More On Our Team

Quick Summary

• A model that fills your assumptions just hands your optimism back as a clean table. That's a hype-man, not a lawyer.

• Flip the job: give Claude YOUR real numbers and tell it to be a hostile IC that kills the deal.

• Make it attack the assumptions that move money: rent growth, exit cap, lease-up timeline, expense ratio.

• Rank every risk by how much IRR or equity multiple you lose if you're wrong. Triage, worst-first.

• Make it steel-man the single most credible bear case, then adjudicate into a real diligence list you can act on this week.

Want more of these?

Subscribe so the next volume hits your inbox, and follow PSV and me (Zed Truong) on LinkedIn and X for the in-between stuff. A deal that survives a hostile room is a deal worth doing. That's where conviction comes from.

Subscribe now

Follow PSV on LinkedIn

Follow me on X

Preface

Claude’s newest model Fable 5.0 lasted 3 days before being shut off by the U.S government. The update was extremely powerful: long, hard tasks like underwriting a 25-page model are now executed much more accurately. However, everything good comes with a cost. In this case, it was more tokens.

5 prompts in and your marriage to Fable switches to an ugly divorce. But in this scenario, unless you pay more for usage, your partner Claude takes all of the estate. We’re here to make sure you at least have a prenup set up.

Welcome. This is Volume 7 of Keeping up with AI in CRE.

To catch you up…

On June 9, 2026, Anthropic released Claude Fable 5, its most capable widely released model, built for "the most demanding reasoning and long-horizon agentic work." Buzzword for the AI perform better on tasks that usually take longer.

The headline spec for CRE: a 1 million-token context window and up to 128k output tokens. In plain English, that means the model can read an entire OM, a full lease stack, and 12 months of financials in one pass (without forcing your team to chop everything into pieces).

Although significantly better than Opus 4.8, Fable 5’s price tag reflects its capabilities at $10 per million input tokens and $50 per million output, roughly double Opus’ rate. Just like you having a daily reliable car and a pleasure ride car, you should use both models to achieve the best mileage.

Ready for the full advice?

Join the CRE AI Institute

Five things to do stop hitting your Claude limit at 9pm.

1. Know what’s eating your usage and clear it.
The OM, T-12, rent roll, and leases can take up 90%+ of your token load. To make sure that number stays below 80%, make sure that you always clear context (if using Claude Code) or start a new chat if you’re switching to a different task.

2. Right-size the model.
Have you ever seen a director work on data entry? Neither have I. So don’t use Opus 4.8 for a simple question that you can Google (or Gemini?). Multi-step tasks like underwriting, red-flag passes, and IC memos go to Opus 4.8, and the rest of easy tasks can be Haiku/Sonnet. Don’t be that guy/gal.

3. Strip the OM before it hits the chat.
A 140-page OM is maybe 40 pages of substance and 100 pages of marketing. In a perfect world, you can just insert everything into Claude and let it run. But if you’re close to your weekly limit, might be worth it to take a little time stripping out irrelevant pages that are “fluffed up” for marketing.

4. One job per chat. Load the deal once.
Don’t run a three-hour everything-thread. Normalize the rent roll, then start fresh for comps, then again for the memo. Just like how you don’t want a guy who’s been doing acquisitions his whole life to all of a sudden handle asset management work, don’t do the same to Claude. It’ll “overload” with stale assumptions, and eventually context loss if the chat gets too long.

5. Extract once, then work against the summary.
I should be gatekeeping this secret sauce. But because the Knicks won, our team can be a bit more charitable here. Instead of having to re-insert a large file into Claude multiple times, you can read the full deal one time, and collapse it into a one-page brief, Next time, you can use that same condense brief instead of re-inserting a 140-page document and burning tokens.

Here’s a quick prompt on how to do that on the operator side:

Looking for a library of prompt vaults that are more than just workflow automation?

Join The CRE AI Institute

Bonus step: a 30-second checklist before a Claude run

Wow you read the whole thing. Made it to the top 10% of readers - you’ll go far! Very unbiasedly, I’d tell you to join the CRE AI Institute to level up you AI knowledge if you really are this curious. Anyways, here’s the tip.

Before you start a deal session, run the checklist:

1. Stripped? Cut the materials to only relevant pages (if necessary).

2. Right model? Lighter model for extracts, Opus 4.8 for reasoning, Fable 5 only for true whole-deal reads.

3. One job? New chat per task, not one endless thread.

4. Loaded once? Deal docs in a Project, not re-pasted.

5. Extract first? Build the one-page brief, then work against it.

Thirty seconds up front might buy you a full uninterrupted session. Run this on your next live deal and you'll feel the difference the first night you don't get cut off.

And the honest limit: none of this is about being cheap. It's about not getting rationed in the middle of real work. The capacity is there, on Opus 4.8 today and Fable 5 when it's back in front of you. Token discipline is just how you make sure the tool is there when the memo's due, instead of tapping out at 9pm like mine did.

Want the prompts we actually deploy? If you refer 5 people, you get free access to our copy-and-paste Claude Skill Vault: the exact prompts and workflows we deploy for CRE clients, including the Deal Brief extractor above and the full underwriting suite it feeds.

About Our Team (glad you asked)

If you’re new here, my name is Zed. I left my transactions role at PGIM to team up with software engineers from Goldman Sachs and HP building frontier AI solutions for commercial real estate at PSV.

My business partners are deeply technical with AI research funded by AWS and the National Science Foundation (even prior to transformer architecture becoming widely popular). One of our advisors was Uber’s first head of engineering and now an AI PhD at Berkeley.

I’ve been lucky enough to learn from the best talent in the AI capital of the world. Most people gate-keep this information to stay ahead, but I’ve decided to give the basics away to fellow CRE professionals, for absolutely no cost, forever.

My commitment to you is translating our team’ deep technical knowledge of the constantly changing AI industry into quick 10-minute reads everyone in real estate can understand.

If you’re reading interested in working with us - book a call with our team 👉 HERE.

If you found this volume helpful, I’d really appreciate a subscribe and follow on LinkedIn so we can keep making great content for you all!

Follow PSV on LinkedIn

About Us. Read This.

PSV is the leading AI partner for commercial real estate. Our team has institutional CRE background at PGIM & CBRE, combined with deep AI expertise from ex-engineers at Goldman Sachs and Kochava. We build the AI systems that brokers, owners, and sponsors actually use on live deals.

About Us

We’ve come a long way

Let me tell you about the moment I stopped trusting AI on deals.

About 2 years ago, I watched dropped a deal room into Co-Pilot and ask it to pull the in-place rents and build a quick T-12 summary. Forty seconds later: a beautiful, confident, perfectly formatted table.

But every number was wrong.

Not "rounding error" wrong. Invented wrong. Rents that didn't appear anywhere in the document. A cap rate from a different deal. One line item it just made up, because one file was scanned crooked on page 44 and the model would rather hallucinate than admit it couldn't read it.

Imagine if I would’ve sent it to my MD.

That, not "is the AI smart enough," has always been the real problem in CRE. They just lied to you with total confidence when they weren't sure. And in our world, a confident lie is worse than a blank cell.

In the last 12 days, that quietly changed. Anthropic have shipped two new models, Opus 4.8 and Fable 5, and buried in the release notes of one of them is the single most important sentence for anyone who underwrites for a living.

Is it just marketing hype to drive attention to their IPO? We’ll break it down for you.

What actually happened (the 12-day one-two punch)

Here's the timeline, because it moved fast:

1. May 28, 2026. Claude Opus 4.8 ships. Same price as before ($5 per million input tokens, $25 per million output), but a meaningfully sharper model with four new operational features (Anthropic's announcement).

2. June 9, 2026. Claude Fable 5 ships. A brand-new tier above Opus. Anthropic's words: "our most capable widely released model, built for the most demanding reasoning and long-horizon agentic work" (model docs).

So in under two weeks, the ceiling moved twice. The reflexive take is "great, it's smarter, who cares." That's the take that keeps you exactly where you were.

But the interesting part isn't the IQ bump. It's actually what they chose to fix.

The number that matters more than any benchmark

Buried in the Opus 4.8 notes is this line, verbatim from Anthropic:

Opus 4.8 is "around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked."

That specific stat is measured on code, not underwriting. I'm not going to dress it up as a CRE benchmark. It isn't one. But read what it's actually describing: the model got dramatically better at catching and flagging its own mistakes instead of letting them slide by.

That is the entire ballgame for us.

Think about what kills AI adoption on a deal team. It's never "the model can't add." It's the made-up rents. The 2021 cap rate it pulled from nowhere (lol), or the line item it confidently fabricated rather than saying "page 44 is unreadable can you check”…

A model that is structurally less willing to let its own errors pass unremarked is a model you can finally put near an underwriting file. Pair that with Fable 5's other new behavior (it now has safety classifiers that can decline a request outright and returns a "refusal" instead of bluffing), and you're looking at the first generation of these tools built to say "I'm not sure" out loud.

In CRE, that is worth more than a hundred confident, wrong tables. I mean would you rather have an analyst that tells you when they make a mistake or one that never checks his work but still sends it anyways?

The 1M context window: run multi-step tasks

Here's the workflow change you can make this week, and it's the one almost nobody is using right.

Both Opus 4.8 and Fable 5 run a 1 million token context window with up to 128,000 tokens of output per response. A context window is essentially the amount of tokens it would run until the AI starts losing memory (to keep it high-level).

In plain English: that's roughly a 700-750 page document the model can hold in its head at once.

Most people in our industry are still working like it's 2024: screenshotting the field into pieces, feeding it page by page, stitching the answers back together by hand, and praying the model didn't forget the unit mix by the time it got to the financials.

Stop chopping. The whole deal fits now.

Drop the entire package in (the full OM, the rent roll, the T-12, the loan terms, the PSA redline) and ask cross-document questions you literally couldn't ask before:

• "The OM claims $2.10/sf in-place rent. Does the rent roll actually support that, unit by unit? Show me every unit that doesn't."

• "Reconcile the T-12 expenses against the OM's pro forma. Flag every line where the pro forma assumes an improvement with no stated basis."

• "Cross-check the lease abstracts against the rent roll. Which tenants have an option or a co-tenancy clause the OM didn't mention?"

That last one catches the stuff that blows up deals in DD. The model can only do it because it's holding the lease stack and the rent roll and the OM simultaneously, not three separate conversations you have to referee.

Dynamic Workflows: the inbox-to-underwriting agent is now a product feature

The second Opus 4.8 feature that matters: Dynamic Workflows. Anthropic's framing is that Claude can now "plan the work and then run hundreds of parallel subagents in a single session."

Translate that out of engineer-speak. Last year, an agent that watched your inbox, pulled every OM as it landed, ran each one through your underwriting template, and handed you a ranked shortlist by Monday. That was a step-by-step custom build that other companies charge $5-$10K per step for.

It's now closer to a setting than a system.

The shape of the workflow is as follows:

1. New deal emails hit your inbox with OMs and rent rolls attached.

2. The agent pulls each one, runs it against your box: your assumptions, your return thresholds, your red-flag list.

3. It spins up a separate subagent per deal so ten OMs get underwritten in parallel, not in a queue.

4. You get a ranked shortlist with the three deals worth a real look (and why), plus the seven that died and what killed them.

You're now reviewing the model's kill decisions instead of manually going through everything. Fun right?

Is your company looking to deploy an AI initiative? This is exactly the kind of system we build for CRE teams at PSV + AI advisory, training and education. If this sounds interesting, you can book a call with our team here and we'll map it to your actual processes.

Fable 5 or Opus 4.8: which one do you actually use?

Most people make the mistake ofefaulting to the newest, most expensive model for everything because it's the "best."

Don't.

Look at the pricing side by side:

• Opus 4.8: $5 per million input tokens, $25 per million output. Same as the last version. There's also a fast mode at $10/$50 that runs ~2.5x faster and is three times cheaper than fast mode used to be.

• Fable 5: $10 per million input, $50 per million output. Roughly double Opus.

Fable 5 is the new top gear. But a top gear is for the climb, not the parking lot. Use it where the reasoning is genuinely hard and long-running: a full DD sweep across a 12-document package, a multi-step development feasibility model, a portfolio-wide re-underwrite. Things where being more right is worth paying double.

For the daily grind (summarize this OM, draft this LOI, pull these rents, clean up this email), Opus 4.8 is the workhorse, and at half the price with a fast mode, it's the one you'll actually run all day.

The operators who win the cost game aren't the ones using the fanciest model. They're the ones who know which model the task deserves. Picking the $50 model for a job the $25 model nails is just lighting money on fire with extra steps.

The one prompt to run this week (catered to operators this time)

The honesty upgrade only helps you if you ask for it. By default, these models still try to be helpful and complete, which means they'll still reach for an answer before they'll flag a gap, unless you tell them the gap is the answer you want.

So tell them. Paste your deal docs into Opus 4.8 or Fable 5 and run this:

You are my underwriting QA partner, not my cheerleader.

I've attached [the OM, rent roll, and T-12] for a deal I'm reviewing.

Pull the in-place numbers I need to underwrite it: in-place rents,
economic occupancy, T-12 NOI, and the pro forma assumptions.

Hard rules:
1. For EVERY number, cite the exact source: document name + page/line.
2. If a number is unclear, illegible, or not stated anywhere, DO NOT
   estimate it. Write "NOT FOUND" and tell me where I have to look.
3. At the end, give me a "Soft Spots" list: every place where the
   document contradicts itself, the pro forma assumes an improvement
   with no stated basis, or you had to make a judgment call. Rank them
   by how much they'd move my valuation if I'm wrong.

I would rather have ten honest "NOT FOUND"s than one confident guess.

That last line is doing real work. You are explicitly rewarding the model for admitting uncertainty, which is exactly the behavior this generation finally got good at. Run it once and watch the "Soft Spots" list. The first time it hands you back the three assumptions the OM was quietly leaning on, you'll feel it click.

That's the dopamine. That's the moment AI stops being a party trick and starts being a second set of eyes you'd actually let near a deal.

What to do differently this week

1. Stop chopping documents. Put the whole deal package in one context and ask cross-document reconciliation questions.

2. Demote your daily model. Use Opus 4.8 (with fast mode) for the grind; save Fable 5 for the genuinely hard, long-running analysis.

3. Run the QA prompt above on your next OM and read the "Soft Spots" list before you read anything else.

4. Reframe what the model is for. It's not here to give you the answer. This generation is finally good enough to tell you where the answer is soft, which, on a live deal, is the part that actually saves you.

The ceiling moved twice in 12 days. The teams that notice why (not just that it got smarter, but that it finally got honest) are the ones who'll be underwriting circles around everyone still chopping up OMs by hand.

Learning something new from this post? Forward it to one person on your deal team who's still pasting OMs page by page. (And if you refer 5 people, you get free access to our copy-and-paste Claude Skill Vault: the exact prompts and workflows we deploy for CRE clients.)

Want to know where your team actually stands? Take our 2-minute AI Readiness Quiz.

More on our team: Combined, our team has closed over $1B in CRE transactions and built AI systems at Goldman Sachs and Kochava. Fun fact: we've got robot dogs in our office. 🦾

If you got value here, hit the orange Subscribe button below and restack this so another operator stops trusting confident, wrong tables.

Zed

Follow me on LinkedIn for the in-between-volumes stuff.

About Us - Read This

PSV is the leading AI partner for commercial real estate. Our team has institutional CRE background at PGIM & CBRE, combined with deep AI expertise from ex-engineers at Goldman Sachs and Kochava. Current clients includes teams at Cushman & Wakefield, Westlake Urban, and more.

For tech-forward CRE firms, we deploy AI in 2 main areas:

• Custom AI systems for CRE firms handling the deal cycle end-to-end (sourcing, underwriting, closing, asset management, property management, disposition).

• AI advisory and strategy consulting for mid-market companies and major institutions.

If this sounds like what you need, put in a request to work with us → HERE.

Otherwise, welcome. This is volume 5 of keeping up with AI in CRE.

1. Generic AI tools do not cut it in CRE

You uploaded an OM with financials to Claude CoWork, asked it to underwrite, and got something back that looked clean until your MD asked for the sources. 2021 cap rates. Made up rents. Vacancy that has nothing to do with your submarket. Oof.

It’s not really your fault, nor the way you prompt. Generic AI is trained on the public internet; meanwhile, CRE underwriting is not. Your firm's comp set, your debt terms, your reversion math, your asset management playbook - none of it lives anywhere these models can see by default.

The tech-forward firms getting real value are running custom architecture on top of foundational models like Claude/ChatGPT. Think of it as a system that knows your templates, your data sources, and how your team actually underwrites.

Instead of a generic template, it uses your model, your IC memo, and your triangulation method. This allows AI to shift from being an intelligent assistant, to something closer to a junior analyst.

Now, we NEVER recommend replacing humans entirely with AI. The tool is supposed to be a highly useful co-pilot, not an auto-pilot. As we all know, CRE is a relationship industry, fueled by trust. Human-to-human trust.

The human judgment layer remains a very important factor, and unlike the rest of Silicon Valley, we believe in a future of AI companionship, not AI replacement.

Onto the next!

2. The smartest operators are pushing AI down, not up

Most firms we talk to are thinking about AI at the OpCo level: how do we save on memos, cut analyst hours, automate reporting, etc.

That’s great. But there’s more value. Not up the chain — down to the asset level. Intelligence such as:

• flag tenants likely to vacate 6 months out

• monitor utility usage at the property level

• watch local activity and signal when to push rents

• write the leasing updates and investor letters straight off the data

OpCo work is a cost story. Asset-level work is an NOI story. There are no more cap rate compressions to bail us out this cycle. Buckle up.

If this sounds interesting, you can book a call with our team here.

3. AI is no longer a tool. It’s core business strategy.

A year ago, AI was a productivity hack. Today it’s a board-level conversation.

The largest firms now have AI committees, AI budgets, and AI hires reporting straight to the C-suite. Blackstone, JLL, Cushman, CBRE all making public commitments. Anthropic itself just announced an enterprise services arm backed by a consortium with $1T+ in AUM.

The future is already here.

We see AI as a necessity, not a nice-to-have. If the big institutions are the ones moving first, we all should sleep with one eye open. Adapt or die.

4. You might be losing without knowing it

This one might sting a bit.

Most firms we talk to say some version of “yeah we are playing around with Claude” or “our analyst uses ChatGPT for emails.”

That’s not adoption. That’s dabbling.

Asking Claude to build a generic model is one thing. Having AI agents watch your inbox, pull OMs and rent rolls as they hit, run them through your own underwriting template using your preferred data source, and email the file back to you before lunch is a completely different thing.

And that’s one workflow.

The edge compounds when you have an AI operating system instead of one single-task agent. Forward-thinking groups are looking at end-to-end coverage:

• deal sourcing

• underwriting

• memo creation

• capital raising and funding

• closing

• asset management

• disposition

Humans stay in the loop for the qualitative judgment that actually moves the needle. AI handles the rest.

The firms doing this are not posting about it. They are quietly compounding while everyone else is still picking a subscription. Welltower is a prime example of successfully building tech in the shadows (take a look at this article here).

Ready to go all-in on AI for CRE?

Pacific Software Ventures is the leading AI developer and consultant for tech-forward real estate firms, run by ex-PGIM, CBRE, and Goldman Sachs with published AI research.

We’re applying Palantir’s AIP business model to CRE — going 10 miles deep with each client and building for AI intelligence that actually gives you an edge.

Check out our work below.

Our Website

Learning something new from this post?

Share the article with a member on your team, or a friend in real estate who you think should read this! (5 referrals result in free access to our fully ready copy-and-paste Claude Skill Vault)

Share

More on our team

Combined, our team have closed over $1B in CRE transactions and have published two research papers on AI, funded by AWS and the National Science Foundation.

We’re backed by one of the best tech companies and figured in the world, including Uber’s first head of engineering, Jason Calacanis’ LAUNCH pre-accelerator, Founders, Inc, and more. Current clients include teams at Cushman & Wakefield, Industry RE Partners, Westlake, and more.

Our team is also located in San Francisco, so whatever you just heard about AI, we’ve tackled it here first. Fun fact: we got robot dogs in our office!

About Us

Want more of these?

I’ll share the most raw, up-to-date instructions on AI applications so CRE professionals can stay ahead of the curve. If you feel that you’re constantly behind with no time to learn, subscribe and share with a friend and I’ll make sure you’re all caught up in the world of AI.

Subscribe now

If you enjoyed this article, feel free to share it below and follow me on X!

Share

About Us - Read This

PSV is the leading AI partner for commercial real estate. Our team has institutional CRE background at PGIM & CBRE, combined with deep AI expertise from ex-engineers at Goldman Sachs and Kochava. Current clients includes teams at Cushman & Wakefield, Westlake Urban, BaySpring, and more.

For tech-forward CRE firms, we deploy AI in 4 main areas:

• AI education & training for executives and employees

• Forward-deployed AI engagement where our engineers automate workflows inside your company’s existing AI tool stack (Claude, ChatGPT, CoPilot, among others)

• Custom AI employees built for leading CRE firms handling high volumes of deals

• Fractional AI advisory and strategy consulting

If this sounds like what you need, put in a request to work with us → HERE.

Otherwise, welcome. This is volume 4 of keeping up with AI in CRE.

Anthropic’s Big Reveal

It’s an exciting and honestly a bit uncertain time for us all in CRE. Every time we blink, it feels like Anthropic or OpenAI is releasing something new that could change how our industry works.

The reality is most busy real estate professionals do not have time to keep up with all of it. That’s why our research team spends all day deep in AI for commercial real estate and pulls out what actually matters, so that you don’t have to.

In quick 10-minute reads, we’ll break down the most relevant CRE technology news and practical tips, without any buzzwords so any non-technical professional can understand.

So what is all the hype behind this new Claude Opus 4.7 release?

There are 3 main things you need to know.

1. Opus 4.7 is substantially better at agentic workflow

2. You can now assign a budget to every task

3. Vision gets an upgrade to higher-resolution

Individually, none of these are really game-changing. But when combined together, they are extremely powerful when it comes to CRE use cases. When you use custom architecture on top of foundational models like Claude, AI gets pushed from being a helpful assistant to an actually reliable junior analyst.

Don’t worry, we’ll break it down for you below.

What “better at agentic workflow” actually means

You’ve probably tried using AI to underwrite a live deal. It works for one step, then forgets your assumptions the next step. 2021 cap rates in 2026 environment. Yikes.

If you’ve ever tried something like this:

• Upload an OM, T12, rent roll

• Upload your internal model

• Tell Claude “make no mistakes” and pray it works

Then you might be in a bit of a better situation this time. In this release, Anthropic improves the model’s ability to stay consistent across multi-step tasks.

Opus 4.7 scores a 64.4% on the Vals Finance Agent v1.1 benchmark (a 3.7% increase from Opus 4.6). Numbers don’t lie - Claude by itself won’t be perfect.

However, we’ve found that combining custom architecture with foundational models brings much better accuracy - typically in the 90%+ range for CRE underwriting. This is typically achieved with a mix of agent teams, and some proprietary tech.

One of our clients ran a 25-page life science model through a custom AI employee we built. No #REF, no #ERROR, and it remembers assumptions across different markets. We have our own way of pulling comps (I know you’d ask).

If this sounds interesting, you can book a call with our team here.

Task budgets — what to do here

This might sound boring to you if you’re working at a large institution who’s burning serious cash on API-based billing. For the rest of us, it’s not as simple.

Before this, AI would either overthink simple deals or underthink complex ones. No in between.

You’d run a workflow and either:

• burn unnecessary cost on a straightforward deal

• or get a half-baked output on something that actually needed depth

Now you can assign a “budget” to the entire workflow.

Thinking, tool calls, outputs. Everything. The model sees the limit and paces itself.

We’ve been doing this on our end already through custom orchestration. This just makes it more native. The typical results are:

• more predictable cost per deal

• more consistent output quality

• no runaway workflows in the background

Because a simple suburban multifamily deal should not get the same level of processing and cost as a 25-page life science development. Onto the next!

What about vision upgrade?

CRE data is messy. Everyone knows this.

Rent rolls come in 10 different formats.. T12s are half clean, half digging for line items that never existed. And sometimes decks love to hide key numbers in random places.

Older models would miss things. But now with higher-resolution vision, Claude gets better at reading dense documents.

We’ve seen this matter most when it comes to:

• large rent rolls with small font

• scanned PDFs

• screenshots from brokers or internal systems

• large files that you pray won’t crash your computer & unsaved excel file (happened to me)

Combined with PSV architecture, you get very close to an AI analyst - handling market research, underwriting, memo creation and more for your team.

Is your company looking to deploy an AI initiative?

Pacific Software Ventures is the leading AI developer and consultant for tech-forward real estate firms. We help companies:

1. 3-5x deal flow and work output with 24/7 custom AI employees

2. Upskill your team on AI tools through hands-on, in-person training

3. Implement end-to-end workflows and integrations within your existing Claude/Co-Pilot/ChatGPT subscription

Check us out by clicking the orange button below.

Our Website

Learning something new from this post?

Share the article with a member on your team, or a friend in real estate who you think should read this! (5 referrals result in free access to our fully ready copy-and-paste Claude Skill Vault)

Share

More on our team

Combined, our team have closed over $1B in CRE transactions and have published two research papers on AI, funded by AWS and the National Science Foundation.

We’re backed by one of the best tech companies and figured in the world, including Uber’s first head of engineering, Jason Calacanis’ LAUNCH pre-accelerator, Founders, Inc, and more. Current clients include teams at Cushman & Wakefield, Industry RE Partners, Westlake, Bayspring and more.

Our team is also located in San Francisco, so whatever you just heard about AI, we’ve tackled it here first. Fun fact: we got robot dogs in our office!

Work With Us

Want more of these?

I’ll share the most raw, up-to-date instructions on AI applications so CRE professionals can stay ahead of the curve. If you feel that you’re constantly behind with no time to learn, subscribe and share with a friend and I’ll make sure you’re all caught up in the world of AI.

Subscribe now

If you enjoyed this article, feel free to share it below and follow me on X!

Share

Follow me on X

Preface

You can skip this if you’re a regular. If you’re new here, my name is Zed. I left my transactions role at PGIM to team up with 2 software engineers from Goldman Sachs and HP building frontier AI solutions for commercial real estate at PSV.

I’ve been lucky enough to learn from the best AI talent in San Francisco, including Uber’s first head of engineering, who’s one of our close advisors. My commitment to you is translating our team’ deep technical knowledge of the constantly changing AI industry into quick 10-minute reads everyone in real estate can understand.

All I ask is if you can share this article with someone who will find it helpful.

Welcome. This is volume 3 of keeping up with AI in CRE.

Why staying informed about AI matters more than you think

If you're allocating capital, underwriting deals, or advising clients in CRE right now, the gap between the firms that talk about AI and the firms that are actually positioned for it is widening fast. The ones ahead are the companies actually implementing it, not blindly, but correctly.

In this quick 10-minute read, we'll break down what the biggest institutions see as the real opportunities in AI — and where they're quietly pulling back

If you’re not sure where your firms stands, feel free to take our AI knowledge quiz below to see how you’re doing in comparison to peers in the industry.

The TLDR Summary

A few patterns jumped out after reading all 20:

• Everyone agrees data centers are the trade: Blackstone, PwC, Brookfield, Hines, and Morgan Stanley all rank digital infrastructure as their #1 or #2 opportunity for 2026.

• Sentiment is splitting between AI infrastructure and AI implementation: The firms bullish on data centers are the same firms admitting their operational AI rollouts are stalling without the right guidance.

• Only MetLife and a handful of others are publicly modeling the downside: Specifically, the 50K–4M white-collar jobs AI may have already eliminated, and what that means for office and multifamily leasing.

• Gateway markets with AI hubs are the consensus winner: Manhattan, San Francisco, and Dallas: CBRE, JLL, and Morgan Stanley all flag them as leading the recovery.

All of the articles will be linked below for your due diligence.

The 20 Outlooks, Unfiltered

• Blackstone: “All-in on AI infrastructure.” Link: https://www.blackstone.com/insights/article/real-estate-enters-the-next-phase-of-the-cycle/

• PGIM: “US growth is riding AI capex.” Link: https://www.pgim.com/us/en/institutional/insights/market-portfolio/outlooks

• CBRE: “AI cities drive the $562B investment rebound.” Link: https://www.cbre.com/insights/books/us-real-estate-market-outlook-2026

• JLL: “Only 5% of firms hit their AI goals.” Link: https://www.jll.com/en-us/insights/market-outlook/global-real-estate

• Cushman & Wakefield: “AI capex anchors the 2026 recovery.” Link: https://www.cushmanwakefield.com/en/united-states/insights/united-states-outlook

• Brookfield: “Asia-Pacific skips legacy models for AI fulfillment.” Link: https://www.brookfield.com/views-news/insights/real-estate-outlook-investing-through-next-cycle

• Morgan Stanley: “AI to unlock $34B in real estate efficiency gains.” Link: https://www.morganstanley.com/insights/articles/ai-in-real-estate-2025

• Deloitte: “AI honeymoon over — 27% hit implementation walls.” Link: https://www.deloitte.com/us/en/insights/industry/financial-services/commercial-real-estate-outlook.html

• PwC: “Big Tech to spend $660B on AI buildout in 2026.” Link: https://www.pwc.com/gx/en/industries/financial-services/real-estate/emerging-trends-real-estate/etre-global-outlook.html

• KPMG: “70% of NYC execs grow CRE footprint despite AI.” Link: https://kpmg.com/us/en/media/news/new-york-perspectives.html

• Hines: “Digital infrastructure leads the cycle turn.” Link: https://www.hines.com/news/hines-2026-outlook-global-real-estate-stabilizes-and-enters-a-new-growth-cycle

• Nuveen: “Six straight quarters of gains — AI sectors lead.” Link: https://hub.ipe.com/asset-manager/nuveen-real-estate-homepage/2026-real-estate-outlook-navigating-the-real-estate-resurgence/10134736.supplierarticle

• Colliers: “AI adoption reshapes every CRE sector.” Link: https://www.colliers.com/en/research/nrep-outlook-2026

• Newmark: “AI already dictates where CRE demand lands.” Link: https://www.nmrk.com/insights/thought-leadership/the-ai-driven-evolution-of-commercial-real-estate

• Savills: “AI now rivals macro as the top force in RE.” Link: https://impacts.savills.com/market-trends/the-themes-that-will-shape-global-real-estate-in-2026.html

• Marcus & Millichap: “AI and fiscal stimulus are the two tailwinds.” Link: https://seekingalpha.com/article/4862027-marcus-and-millichap-inc-mmi-discusses-2026-outlook-and-strategies-for-maximizing-real-estate

• LaSalle: “AI-driven demand breaks the Groundhog Day cycle.” Link: https://www.prnewswire.com/news-releases/lasalles-isa-outlook-2026-americas-real-estate-poised-to-break-its-groundhog-day-cycle-302632691.html

• Goldman Sachs Asset Management: “AI is one of five catalysts for 2026 alpha.” Link: https://am.gs.com/en-us/institutions/insights/article/investment-outlook

• MetLife Investment Management: “AI erased up to 4M white-collar jobs.” Link: https://investments.metlife.com/insights/real-estate/2026-commercial-real-estate-outlook/

• CBRE Investment Management: “GenAI is redefining US office demand.” Link: https://www.cbreim.com/insights

Want to learn how others are implementing AI in their CRE business? Got a project that you’re dying to deploy?

Pacific Software Ventures is the leading AI developer and consultant for tech-forward real estate firms. We help companies:

1. 3-5x deal flow and work output with 24/7 custom AI employees

2. Upskill your team on AI tools through hands-on, in-person training

3. Implement end-to-end workflows and integrations within your existing Claude/Co-Pilot/ChatGPT subscription

Our Website

Learning something new from this post?

Share the article with a member on your team, or a friend in real estate who you think should read this! (5 referrals result in free access to our fully ready copy-and-paste Claude Skill Vault)

Share

More on our team

Combined, our team closed over $1B in CRE transactions and have published two research papers on AI, recognized by AWS and the National Science Foundation.

We’re backed by one of the best tech companies and figured in the world, including Uber’s first head of engineering, Jason Calacanis’ LAUNCH pre-accelerator, Founders, Inc, and more.

Our team is also located in San Francisco, so whatever you just heard about AI, we’ve tackled it here first. Fun fact: we got robot dogs in our office!

Work With Us

Want more of these?

I’ll share the most raw, up-to-date instructions on AI applications so CRE professionals can stay ahead of the curve. If you feel that you’re constantly behind with no time to learn, subscribe and share with a friend and I’ll make sure you’re all caught up in the world of AI.

Subscribe now

If you enjoyed this article, feel free to share it below and follow me on X!

Share

Follow me on X

My co-founders’ work on AI

Gedeon Baende (previously at Goldman Sachs): here.

Aidan Scudder (previously at Kochava): here.

Preface

You can skip this if you’re a regular. If you’re new here, my name is Zed. I left my transactions role at PGIM to team up with 2 software engineers from Goldman Sachs and HP building frontier AI solutions for commercial real estate at PSV.

I’ve been lucky enough to learn from the best AI talent in San Francisco, including Uber’s first head of engineering, who’s one of our close advisors. My commitment to you is translating our team’ deep technical knowledge of the constantly changing AI industry into quick 10-minute reads everyone in real estate can understand.

Welcome. This is volume 3 of keeping up with AI in CRE.

To win, do a little more than others

Most real estate investors still rely on:

• broker relationships

• stale databases

• content no one watches

• word of mouth (powerful but tough to scale)

That’s not enough anymore. To win at this game, you’ll have to run the extra mile. Do more research, enrich more data, be more personal, and look for signals that others aren’t paying attention to.

In under 10 minutes, we’ll show you exactly how to:

• find properties before they’re marketed

• figuring out who actually owns them

• reaching out with context

…all without spending thousands on third parties, or having to act like Juliano Massarelli on cold calls. For those of you that pay attention, there will be a bonus at the end!

The Stack & Goal

• Perplexity → find signal

• Claude (CoWork) → organize + execute

• Skip tracing tool (optional) → speed up contact info

You’re building a system that:

1. finds properties early

2. maps to real people

3. reaches out with context

Step 1. Be Specific Or Don’t Even Bother

If your criteria sucks, your output will suck. You need a well-defined deal box.

Example of what I’d actually run for a commercial real estate lead:

• Asset type: Industrial

• Size: 20–100K SF

• Location: East Bay Area

• History: Owned 10+ years

• Sales activity: No recent activity.

Save this - you’ll use it for memory later.

Step 2. Hunt With Perplexity

Drop your buy box into Perplexity. The goal is not to find the properties that perfectly fit your box. It’s to find inconsistencies that acts as selling signals.

• Partial information: ownership isn’t fully clear, property details show up in fragments across sources & pieces exist, but no clean story.

• Weird signals: long hold periods with no visible activity, assets that look “off” for the submarket & signs of neglect, mismanagement, or inertia.

• Incomplete data: missing leasing info, outdated listings or stale marketing, no clear operator presence.

That’s where deals come from.

Example:

find industrial properties in the bay area that fit this criteria: [insert your buy box]. prioritize long-term ownership, low leasing activity, or signs of underperformance. Return everything to me in a digestible forma. Always check sources, cite them, and do not make up data without auditing and verifying. My goal to acquire these properties (or to get them to sell as a broker).

Step 3. Build A List in Claude

Drop results into Claude:

[insert context into previous conversation] and turn this into a table with property, ownership clues, and a priority score based on my box and [insert investment strategy]. keep a running list so we don’t duplicate.

Each week, you can add new properties, remove duplicates, and rank with the help of Claude.

Essentially a simple, smart, self-improving CRM.

Step 4. Identify the Key Parties (The Part Everyone Gets Wrong)

Most of you can take care of the rest from here and just stop at finding the LLC name. Maybe you’ll use a skip tracing tool to call the potential owner.

That will only give you mediocre results. You’re baking one potentially right person and time window.

Use Perplexity to look for patterns, not answers:

• same entity across multiple assets

• repeated names in filings, articles, leasing docs

• company tied to the mailing address

• individuals mentioned in any transaction history

So when you run into roadblocks when data is stale, there’s still a story that can be pieced together. You’ll have more contacts to chat with, more relationships to make, and thus more deals to close.

Cast a wide enough net, in a good enough place, and eventually you’ll catch a fish.

Step 5. Get Contact Info + Context

To get email domains for free, you can scrape using either Claude or Perplexity for:

• company website

• public leasing docs

• press mentions

or just use skip tracing tools like DealMachine or Reonomy for faster, more accurate results.

Then, use Perplexity to look for:

• LinkedIn

• past deals

• podcasts / interviews

• company focus

• geography

• anything they care about

These information will be used as context for outreach.

Step 6. Use Claude CoWork to Run Outbound

Now feed all the context above into Claude (CoWork):

Example Prompt:

write a short, natural outreach email referencing the property and ownership. keep it simple with light, relevant personalization.

Then, create 2-3 follow up email chains (just prompt Claude) to do so. Make sure you connect your gmail (or email account) to Claude under connectors before doing so.

Final Step: Repeat Weekly - That’s It!

We’ve actually automated this entire process 24/7 using Open Claw + custom plugins so you don’t have to constantly copy and paste prompts between Claude and Perplexity (along with proprietary tactics we’ve used to find leads). If you’re interested, we our team can set up a custom system for you! Book a call here if that’s the case.

Learning something new from this post?

Share the article with a member on your team, or a friend in real estate who you think should read this! (5 referrals result in free access to our fully ready copy-and-paste Claude Skill Vault)

Share

Want to see where your business stands with AI knowledge? Take a short quiz.

Pacific Software Ventures is the leading AI consultant for tech-forward real estate firms. We help companies:

1. Train executives and teams on how to actually use AI effectively

2. Implement end-to-end workflows and integrations within Claude

3. Build custom AI systems for firms seeking a proprietary edge.

This is an example of an autonomous agent we’ve built:

Want us to build you a custom AI system that gets results even better than this? Book a free AI audit here and we’ll show exactly where your tech stack stands.

Take AI Readiness Quiz

Want to work with us?

Talk To Our Principals

More on our team

Combined, our team closed over $1B in CRE transactions and have published two research papers on AI, recognized by AWS and the National Science Foundation.

We’re backed by one of the best tech companies and figured in the world, including Uber’s first head of engineering, Jason Calacanis’ LAUNCH pre-accelerator, Founders, Inc, and more.

Our team is also located in San Francisco, so whatever you just heard about AI, we’ve tackled it here first. Fun fact: we got robot dogs in our office!

Work With Us

Quick Summary

• Most CRE professionals are still waiting for deal flow instead of building systems to generate it

• Perplexity is used to scan for inconsistencies in the market, not to find perfect deals

• Claude becomes your memory layer, turning scattered data into a compounding pipeline

• The goal is not to find the owner, it is to identify and reach the actual decision maker

• The real edge is consistent outbound, turning leads into conversations every week

Bonus tip: when you find one good owner, map every other property they own and reach out across the entire portfolio, not just one asset. A good relationships goes a long way.

Want more of these?

I’ll share the most raw, up-to-date instructions on AI applications so CRE professionals can stay ahead of the curve. If you feel that you’re constantly behind with no time to learn, subscribe and share with a friend and I’ll make sure you’re all caught up in the world of AI.

Subscribe now

If you enjoyed this article, feel free to share it below and follow me on X!

Share

Follow me on X

My co-founders’ work on AI

Gedeon Baende (previously at Goldman Sachs): here.

Aidan Scudder (previously at Kochava): here.