How to Automate Real Estate Underwriting with AI [Step-by-Step]

A Monday morning at Andover Properties, 2022. Twenty-three OMs sitting in the pipeline from the weekend. Our acquisitions team needed initial underwrites on all 23 before Wednesday's IC pipeline review. That is 160 to 230 hours of analyst time. We had 4 analysts. The math did not work, so we triaged: eyeballed the OMs, picked the 8 that looked interesting, skipped the other 15. One of those 15 turned out to be the best risk-adjusted deal we saw that quarter. A competitor bought it at a 7.2% going-in cap with 200 basis points of rent upside. We missed it because we did not have time to open the PDF.

That deal haunted me. Not because we made a bad call, but because we never got to make a call at all. After $11B+ deployed across 150+ transactions, the single biggest operational bottleneck in CRE acquisitions is the time between receiving an OM and having a populated model. Not the analysis. Not the judgment. The data entry.

The 80/20 Problem Every Acquisitions Team Has

JLL's 2025 Global Real Estate Technology Survey found that 72% of institutional investors plan to increase AI spending in 2026, with deal underwriting ranked as the top automation priority. CBRE estimates the average institutional acquisitions analyst spends 62% of their time on data entry and document processing, leaving 38% for actual analysis. Deloitte's 2025 CRE Outlook reported that firms using AI-driven underwriting workflows are closing deals 40% faster.

Those numbers match what I saw at Andover. Our best analysts, people we were paying $150K+, spent most of their working hours copying numbers from PDFs into spreadsheets. Not doing the work that actually determines whether a deal is a good investment. Data entry.

The technology inflection happened in 2024 to 2025 when large language models became accurate enough to parse CRE-specific documents with high reliability. Before that, OCR tools could extract text but could not understand it. They could not distinguish the broker's pro forma NOI from the actual trailing NOI, or tell gross rent from effective rent after concessions. Today's AI can.

What Your Workflow Actually Looks Like Before AI

I lived this for a decade. Here is exactly what happens:

1. Broker emails a 40-page OM as a PDF attachment on Friday at 4:47 PM.
2. Analyst opens the PDF side-by-side with the firm's proprietary Excel model on Monday morning.
3. Analyst manually locates and extracts: unit mix (page 12), in-place rents (page 14), opex (page 27), tax info (page 31), capex history (page 33), broker pro forma assumptions (scattered across pages 35 to 42).
4. Analyst types each value into the correct cell. Unit count to B5. Asking price to B7. In-place rent by unit type to the rent roll tab, columns C through H, rows 4 through 47.
5. Analyst processes the rent roll separately, which may be a different PDF or buried in pages 15 through 22 of the OM.
6. Analyst processes the T-12, mapping 20+ expense line items from the broker's categories to the firm's standardized categories.
7. Analyst cross-references numbers between documents. Does the OM summary NOI match the T-12 bottom line? Do rent roll totals match the unit mix table?
8. Analyst adjusts assumptions based on market knowledge and investment thesis.
9. Analyst runs scenarios, sensitivity analysis, finalizes the model.
10. Analyst builds an IC memo for Wednesday's pipeline review.

Steps 1 through 7 are almost entirely manual data entry. I have timed this across dozens of deals and hundreds of analysts. For a typical multifamily OM, those steps take 6 to 10 hours. For a complex deal (large rent roll, multiple buildings, renovation program, unusual expense structure), it can be 12 to 15 hours.

The real cost is not just time. A single transposition error in step 4 cascades through your entire model. And when you are screening 50+ deals per month to find the 3 to 5 worth pursuing, the manual process becomes the binding constraint on deal flow.

What AI Automation Actually Means (and Does Not Mean)

I have seen too many pitches that overpromise on this. Let me be direct.

AI-automated underwriting does NOT mean:

• Push a button and get an investment recommendation.
• Replace the analyst's judgment on exit cap rate assumptions.
• Eliminate the need for submarket expertise.
• Produce a final IC memo without human review.

It DOES mean:

• AI reads your deal documents and understands what it is reading.
• AI extracts structured data from unstructured PDFs with high accuracy.
• AI populates your existing Excel model, mapped to the correct cells.
• AI cites every value back to its source document and page so you can verify.
• Your team reviews, adjusts assumptions, applies investment thesis, and makes the decision.

AI compresses steps 1 through 7 into a single automated step that takes minutes. Your analysts spend their time on steps 8 through 10: assumptions, analysis, and the IC memo. That is the work where CRE expertise, market intuition, and investment judgment actually matter.

Step 1: Audit Your Template (One-Time, 3 to 5 Hours)

Before you automate anything, you need to know exactly what you are automating. At Andover, our template had evolved over 8 years and 200+ deals. Fourteen tabs, roughly 300 input cells, some formatting quirks that only made sense if you knew the history. Your template probably looks similar.

• Catalog every input cell. Walk through each tab and identify every cell where raw data from documents gets entered. Asking rents, opex, unit counts, square footages, tax assessed values, capex line items.
• Separate inputs from formulas. Most well-structured institutional models already color-code inputs (blue font is the convention). If yours does not, add it now. AI should only populate input cells, never overwrite formulas.
• Document the source for each input. "Unit count: OM summary, usually page 2 or 3." "In-place rent per unit: rent roll, varying page." "Real estate taxes: T-12, line 14." This becomes your mapping configuration.
• Flag judgment-call inputs. Some cells require human judgment, not extraction. "Market rent growth assumption" is not a number you pull from the OM; it is a number your analyst determines from Yardi Matrix data and submarket research. Flag these so the AI leaves them blank.

This is a one-time effort. It will also improve your model documentation for onboarding new analysts, which is a side benefit nobody talks about.

Step 2: Pick Your Tool

I wrote a full comparison of available tools (see Best AI Tools for CRE Underwriting [2026]). The short version: you need four things. CRE document intelligence (can it parse an OM?). Excel-native integration (does it live in your spreadsheet?). Custom model mapping (does it work with YOUR template?). Provenance citations (can your MD click a cell and see where the number came from?).

General-purpose Excel AI tools like Endex or Shortcut are strong for generic finance work but cannot process CRE documents. If OM-to-model population is your bottleneck, you need a CRE-specific tool. Budget 1 to 2 weeks for evaluation and vendor demos.

Step 3: Configure Your Document-to-Model Mapping (One-Time, 4 to 8 Hours)

This is where the work from Step 1 pays off. Mapping tells the AI which extracted data points go into which cells.

• Map OM summary data to your deal overview tab. Property name, address, total units, total SF, asking price, in-place cap rate. Straightforward one-to-one mappings.
• Map rent roll fields to your rent roll tab. Unit number, unit type, SF, beds/baths, lease dates, current rent, market rent, concessions. This gets complex if your template's column order differs from typical OM formats.
• Map T-12 expense line items to your opex tab. This is where it gets tricky. The broker might list "Repairs and Maintenance" as one line. Your model might break it into "Interior R&M" and "Exterior R&M." Configure rules for category mismatches.
• Define handling for missing data. What if the OM does not disclose insurance costs? Does the cell stay empty? Get flagged with a comment? Get populated with a Green Street benchmark? Define these rules upfront.

At Andover, the initial mapping took us about 6 hours. After that, every deal used the same configuration. One-time setup that persists across all future deals.

Step 4: Run Your First AI-Assisted Underwrite Against a Known Deal

Start with a deal you have already underwritten manually. This gives you a known baseline.

• Upload the OM from the completed deal. Let the AI extract and populate your template.
• Compare every extracted value against your manual underwrite. Cell by cell. Flag discrepancies.
• Check provenance citations. Click 10 to 15 populated cells at random. Verify the cited source page actually contains that number.
• Time the process. Upload to fully populated model. Compare against the hours your analyst originally spent.
• Document issues. Missing data, incorrectly mapped values, formatting problems. These inform configuration refinements.

On the first run, expect 90% to 95% accuracy on a well-formatted digital PDF OM. The remaining edge cases are usually unusual formatting or category mismatches. Each iteration improves accuracy.

Step 5: Build Validation Rules

Trust but verify. Even with high extraction accuracy, you want automated sanity checks.

• Cross-document checks. Does the total unit count from the rent roll match the OM summary? Does the T-12 NOI match the OM's stated NOI? These are the errors that catch analysts when they are tired and rushing on a Thursday before IC. Let the AI flag them automatically.
• Reasonableness checks. Is the per-unit rent within 2 standard deviations of the submarket average per Yardi Matrix? Is the expense ratio between 35% and 55% for multifamily? Is the going-in cap between 4% and 9%? Configure bounds from your own experience.
• Completeness checks. Are all required input cells populated? Which ones still need manual attention?
• Provenance audit. Randomly sample 5 cells per deal and verify citations. Over time, as you build confidence, reduce the sampling rate.

Step 6: Scale with Automation Workflows

Once you trust the core pipeline (typically after 10 to 15 deals), start building automation around it.

• Email-triggered intake. Forward your acquisitions inbox so that when a broker sends an OM, the AI automatically processes the attachment and queues a populated model for analyst review. At Andover, this alone would have saved 15 to 20 minutes per deal just on downloading, saving, and organizing PDFs.
• Batch processing for the Monday morning OM pile. Upload all OMs received during the week in a single batch. AI processes them overnight. Your team has populated models waiting Monday morning before the pipeline review.
• Screening alerts. Configure rules to flag deals that meet your criteria: cap rate above 6.5%, 150+ units, Sun Belt MSA, value-add opportunity. Get notified when a qualified deal hits the pipeline.
• CRM/pipeline integration. Feed extracted deal metadata (property name, location, unit count, asking price, cap rate) into your deal tracking system automatically.

Budget 2 to 4 hours to set up each workflow. These compound. After 3 months you will wonder how the team functioned without them.

Step 7: Expand by Property Type

Do not try to automate everything on day one. Start narrow. Prove the value. Expand.

• Month 1: Highest-volume property type. If you are primarily multifamily, start there. Get OM parsing, rent roll extraction, and T-12 mapping dialed in.
• Month 2: Second property type. Industrial OMs look different from multifamily: tenant roll instead of rent roll, NNN lease structures, TI/LC schedules. Configure the mapping for industrial-specific data points.
• Month 3: Add automation workflows and connect market data sources.
• Month 4+: Additional property types (retail, office, mixed-use). Refine validation rules based on error patterns. Train new team members on the AI-assisted workflow.

Before and After: The Numbers

These numbers are based on institutional acquisitions teams processing 15 to 60 deals per month:

Where AI Gets It Right and Where It Does Not

I have been building and testing AI for CRE underwriting for years. Here is an honest accounting.

AI Handles Well

• Structured data extraction. Unit mixes, rent rolls, opex line items from PDFs. When I was manually processing a 380-unit rent roll for a portfolio deal in Atlanta, it took my analyst a full day. AI does it in minutes.
• Model population. Placing extracted values into the correct Excel cells with proper formatting. Solved problem when the mapping is configured correctly.
• Cross-referencing. AI catches inconsistencies between documents faster than manual review. Does the OM's stated NOI of $3.8M match the T-12 net income? AI flags a $47K discrepancy instantly.
• Scenario generation. Once the base case is populated, AI rapidly generates sensitivity tables across exit caps, rent growth, renovation costs, and leverage structures.

AI Still Falls Short

• Assumption quality. AI cannot tell you whether 3.5% annual rent growth is aggressive for Class B multifamily in suburban Raleigh. That requires knowing the submarket's supply pipeline from CoStar, employment trends, and historical absorption.
• Qualitative judgment. The best deal I ever sourced at Andover came through a relationship with a property manager who knew the seller was motivated. AI cannot assess sponsor quality, PM reputation, or the political dynamics of a submarket's entitlement process.
• Broker pro forma skepticism. Every CRE professional knows the broker's pro forma is aspirational. AI will faithfully extract those numbers. It takes a human to know that "Year 3 stabilized NOI" assumes 98% occupancy in a submarket that has never been above 94%.
• Edge case documents. Poorly formatted OMs (especially from smaller regional brokers), scanned handwritten docs, and unusual layouts can still trip up extraction. Provenance citations are your safety net here.

The right mental model: AI handles the data layer (extraction, population, cross-referencing, scenario generation) at machine speed. Humans handle the judgment layer (assumptions, market context, investment thesis, relationships) with years of pattern recognition. Together, you get better outcomes than either alone.

Start With 5 Deals This Month

If you take one thing from this, let it be this: start with a single property type and a single document type. Prove the value on 5 real deals. Then expand.

1. Pick your highest-volume property type. Probably multifamily if you are reading this.
2. Pick one document type. Start with OMs. Most data points, highest manual processing time.
3. Run 5 real deals through the AI pipeline alongside your manual process. Compare accuracy and time savings on each.
4. Measure everything. Time per underwrite. Extraction accuracy. Errors caught. You need hard data to make the business case to your partners or IC.
5. If the numbers work, expand. In my experience, the ROI case becomes obvious after the second deal.

For tool comparisons, see Best AI Tools for CRE Underwriting [2026] and Endex AI vs Shortcut AI vs RealQuant.

The firms that adopt AI-automated underwriting in 2026 will screen every OM in their pipeline, submit LOIs before competitors, and put analyst hours toward actual investment analysis instead of data entry. The question is how many deals your team will pass on this quarter because nobody had time to open the PDF. Request early access to RealQuant and run your first 5 deals.