AI for Net Effective Rent Analysis: 4 CRE Options

By Jude Lee · · Comparison

Two tenant-rep brokers comparing landlord lease proposals and a spreadsheet in a conference room

The job: four proposals, four ways of hiding the concession

A tenant-rep assignment for 18,000 SF generates a stack of landlord responses. One is full-service gross with a base year stop. One is triple net with a $45/SF TI allowance amortized at an undisclosed rate. One offers ten months free on a 126-month term, spread rather than front-loaded. One escalates at CPI rather than a fixed 3%, referencing the index published by the U.S. Bureau of Labor Statistics — which makes the escalation an assumption, not a number.

The analyst’s job is to make those comparable: normalize rentable area (buildings measured under different editions of the BOMA floor measurement standards can produce different load factors for the same usable space), build a cash flow per year, subtract concessions, and report both a straight-line average and a discounted NER. The discount rate choice alone can reorder the ranking, which is why the memo has to show the rate, not just the answer.

This is adjacent to — but not the same as — drafting the LOI or lease proposal itself. That’s about narrowing options. NER analysis is about defending the recommendation.

Where AI genuinely helps, and where it quietly breaks

Split the job into two halves, because they have opposite risk profiles.

Extraction — reading a proposal letter and pulling term, commencement, base rent schedule, escalation basis, free rent months, TI allowance, expense structure, base year, options, and parking — is what modern language models do well. Messy formats and inconsistent labels are the normal case, and pattern-matching beats regex here.

But extraction is not risk-free, and its failure modes are specific. Concessions hide in footnotes and asterisks below the term sheet table. Free rent is frequently granted in the landlord broker’s cover email and never restated in the attached proposal. “Net” means different things across markets and even across two proposals on the same deal — sometimes excluding management fees, sometimes not. Scanned pages, photographed pages, and handwritten margin markups degrade extraction quality without announcing that they have. So the extraction table needs field-level spot checks against the source document, not just a sanity check on the final math.

Arithmetic — building the cash flow, applying escalations, discounting — is where models generate plausible-looking numbers with no audit trail. The fix isn’t a better prompt. It’s making the model call a calculator: write and run code, or invoke a deterministic tool that you wrote and tested. If your workflow can’t show the formula behind each cell, it isn’t ready to go in front of a client.

Option 1: The Excel template and no AI at all

A well-built workbook — one tab per proposal, a normalization layer, an NPV row — is still the honest baseline. It’s auditable, every broker can read it, and it costs nothing beyond the hours to maintain.

It wins when deal volume is low, when proposals arrive in a stable format, or when the analyst is already fast. Its weakness is entirely upstream: someone still reads eight pages and types 30 fields per proposal, and typos in that step are invisible downstream.

Option 2: AI features inside CRE and lease-admin platforms

Three product categories already touch this data: lease administration and abstraction systems (Visual Lease, MRI, Yardi), brokerage CRM and deal platforms (Buildout, Apto, ClientLook), and market data providers (CoStar). The concrete feature to look for is document abstraction — uploading a lease or proposal and having critical dates, base rent schedules, escalation clauses and expense structure land in structured fields you can export rather than retype.

Capabilities and pricing here move quickly, so check each vendor’s own current documentation before you buy — including against this article, last reviewed January 2026. The pattern worth internalizing: platform AI is strongest where the platform already owns structured data, and weakest where your workflow deviates from the vendor’s model of the job. It wins when you already pay for the platform and want zero new vendors to vet. It struggles when you need a custom NER convention — your firm’s discount rate policy, your treatment of amortized TI, your client’s preferred presentation — that the product doesn’t expose.

Option 3: A general AI assistant with code execution

Drop the proposal PDFs into Claude or ChatGPT, ask for a normalized extraction table, review it, then ask it to build and run the cash flow in code and output a spreadsheet. The code-execution part matters: it converts “the model asserted $31.40” into “here is the script that produced $31.40.”

Before anything gets uploaded, deal with confidentiality. Landlord proposals, LOIs and client requirement documents are routinely covered by NDAs or client data policies, and consumer and enterprise tiers of the same assistant can differ on whether inputs are retained or used to improve models. Read the tier’s current data-retention and training terms rather than assuming, and consider redacting party names, addresses and client identity before upload — the math works identically on “Building A” and “Landlord 2.” If your firm handles institutional client data under contractual restrictions, confirm the approach with your compliance lead or counsel before it becomes routine.

The method’s other limits: nothing is remembered between sessions unless you package it, it doesn’t know your comp set, and consistency depends on whoever is typing. Packaging your house method as a reusable skill — a written spec for how your firm normalizes and presents NER — removes most of the variance, the same way standardized marketing skills remove variance from listing collateral.

Option 4: A custom agent connected through MCP

MCP (the Model Context Protocol) is an open standard for giving an AI assistant governed access to your systems and tools. For this workflow, a custom MCP server would expose a small, specific set: get_requirement(client), list_proposals(requirement_id), get_lease_comps(submarket, size_band), and critically calculate_ner(cash_flow, discount_rate, method) — a function you wrote and can point to when a client asks how the number was derived.

That calculator carries real ongoing cost. It needs a regression suite of known-answer cases — deals you already modeled in Excel and agree on — run every time the code changes, and a named owner inside the firm who updates it when your discount rate policy, TI amortization treatment or area normalization rule changes. An unowned calculator drifts silently from house convention.

With that in place the agent does the multi-step job: pull proposals from the deal folder, extract terms, call your calculator, compare against your comp set, draft the client memo, and flag the assumptions it had to make. The mechanics of standing one up are covered in our walkthrough on connecting Claude to CRE systems via a custom MCP server.

General assistant + code execution
Live in an afternoon. No integration work. Good for a few assignments a month. Analyst drives every run, uploads files manually, and owns consistency and confidentiality review.
Custom agent + MCP server
Weeks of build plus ongoing maintenance and a named owner for the calculator. Repeatable output, your NER conventions enforced in code, reads directly from your deal folders and comp database.

How to choose between the four

There is no single best AI tool for commercial real estate, and any list that names one is selling something. The useful question is per-job: what’s the format variance of the input, how high are the stakes of a wrong number, and how often does this happen?

  1. Count the reps

    If you run NER analysis fewer than a couple of times a month, stop at Option 1 or 3. Custom build economics don’t work at low volume.

  2. Check what your platform already does

    Read the current docs for the CRM or lease system you already pay for. Free capability beats a new line item.

  3. Separate extraction from math

    Whatever you pick, require that arithmetic runs in code or a tested function — never as model prose.

  4. Write the house method down

    One page: discount rate policy, straight-line vs. NPV presentation, TI amortization treatment, area normalization rule. That document is your prompt, your skill file, and your training material.

  5. Run parallel for a month

    Analyst does it the old way, agent does it the new way, compare. Promote only after the deltas are explainable.

Two rules of thumb that don’t belong here

Searches for CRE heuristics often land on the 2% rule — the residential shortcut that monthly rent should be roughly 2% of purchase price. It’s a single-family investing filter, not a commercial underwriting method, and it says nothing about tenant-rep lease economics. Commercial analysis runs on cap rates, discounted cash flow, and NER.

The 30% rule in AI is similar folklore: there is no formal industry standard by that name, and the phrase gets used for several unrelated things. My own opinion, stated as opinion: do not set a fixed percentage review rate on financial output. Review 100% of client-facing numbers, and let the AI eliminate the typing, the chasing, and the first draft.

The value isn’t the model producing the number. It’s the model producing the draft, the assumptions list, and the source citations — so a human can check the number in five minutes instead of building it in two hours.

Model the payback with your own numbers

Don’t take anyone’s published savings figure. Build it yourself. The worked example below uses illustrative assumptions only — replace every one of them.

12
Assumption: NER comparison rounds per year (one a month) — substitute your actual count
Illustrative assumption
4 hrs
Assumption: analyst hours per round, extraction through memo — time one real round before using this
Illustrative assumption
12 × 4 × R
Annual desk cost formula, where R is your loaded hourly rate
Formula, not a finding

Then estimate what fraction survives automation — extraction and first-draft memo, realistically, not review or negotiation strategy. Compare that against the build and maintenance cost of whichever option you’re considering, including the calculator owner’s time. The fuller return is reallocation: analyst hours redirected to tour scheduling, canvassing and client contact, plus the errors a tested calculator prevents. If the arithmetic only works when you assume 90% automation, you’ve assumed your way into a bad decision.

What stays human, permanently

The discount rate is a judgment call. So is whether a landlord’s TI number is real money or a figure that will be clawed back in construction management fees. So is telling a client that the lowest NER is the wrong building. An agent that reads eight proposals, normalizes them, runs your tested calculator, and hands you a memo with every assumption flagged is a genuinely good use of this technology. An agent that emails a client a ranking without a broker reading it isn’t a workflow — it’s a liability waiting for a rounding error.

Not sure where to start?

Get a free automation audit: we map your deal pipeline, marketing, and back-office workflows and show you what's worth automating — before you spend a dollar.

Get a free automation audit

Related guides