Putting AI into your comp work does not change what a good comp set is. It changes how cheaply you can build a bad one. A general assistant will assemble, adjust, and narrate a set of comparables in minutes, and it will do it with the same fluent confidence whether the comps are sound or invented. That is why the discipline matters more now, not less: the framework you use to select comparables is the thing that keeps AI’s speed from becoming a liability. This piece lays out a five-part selection framework a non-appraiser can run under deal pressure, and marks exactly where AI helps and where it has to be kept on a short leash.
Why the framework matters more with AI in the loop
Comp selection has always rewarded discipline and punished shortcuts. What is new is that the shortcut now looks like finished work. Ask a general assistant for “recent office comps in my submarket” and it returns a clean table of addresses, dates, and cap rates that reads exactly like a set you spent an afternoon assembling. Some of those transactions may not exist, and the output carries no visible signal that separates a sourced comp from a fabricated one. The burden of that distinction moves entirely onto your process.
A firm without a selection framework used to be slow. A firm without one in an AI-assisted workflow is fast and wrong, which is worse, because speed hides the error until it surfaces in front of a lender or an investment committee. The framework below is not new appraisal theory; it is the same comparable-selection logic appraisers have used for decades, restated as a repeatable sequence and annotated for where a language model belongs in each step. The point is to put AI exactly where it adds speed without letting it make the judgments that determine a value.
The five-part comp selection framework
Every defensible comp set moves through the same five stages, whether a person or a model does the legwork:
- Define the subject — pin down what you are actually valuing before you look at anything.
- Source the pool — gather candidate transactions from real data, not from a model’s memory.
- Screen for comparability — cut the pool to genuinely comparable properties.
- Adjust with evidence — normalize the survivors to the subject with support you can show.
- Defend the set — assemble the result so it survives an outside reader.
AI can accelerate three of these stages and must be supervised in all five. Treat the sequence as a gate: a comp does not advance until it has passed the prior step.
Step 1: Define the subject before you look at a single comp
The most common comp error starts before any comp appears. If you have not specified the subject property precisely, you have no standard to screen against, and a fuzzy standard is one an AI will happily satisfy with loosely relevant transactions. Write down the property type, submarket, size, vintage, class, tenancy, and the specific value question, in-place rent, market rent, exit cap, or price per square foot, that the comps have to answer.
This step is human work with AI as a scribe at most. A general assistant can turn a rambling deal description into a structured subject profile, and it can flag missing attributes (“you have not specified building class”). What it cannot do is decide which attributes matter for this asset, because that judgment depends on how the market prices the specific property, which is exactly the expertise you are bringing. A clear subject definition is what lets every later step reject a comp for a stated reason rather than a hunch.
Step 2: Source the pool from data, never from memory
This is the stage where AI-assisted comp work most often goes wrong, and the rule is absolute: comps come from a data source, never from a language model’s memory. A general assistant has no proprietary transaction database. Asked to produce comps directly, it generates plausible, specific, and frequently fictional transactions, the failure mode that makes untrained AI use dangerous in valuation rather than merely unhelpful.
Your candidate pool has to originate in a real source, and which source depends on the comp type. Sales comps live in platforms like CoStar and Crexi; verified lease terms, base rate, escalations, free rent, and tenant-improvement allowances, come from an exchange like CompStak, which sources them from executed leases rather than public filings; automated multifamily rent comps come from tools like HelloData. Matching the source to the comp type is its own buying decision, and we work through the full landscape in our guide to the best AI comp tools for commercial real estate. The through-line is constant: the data layer supplies the transactions, and the reasoning layer only ever works on data you handed it.
Once the pool is sourced, AI earns its keep. Point a general assistant at an export and it will deduplicate, standardize inconsistent fields, and structure a messy set into a workable grid in seconds, safely, because every row traces to a record you can open.
Step 3: Screen for genuine comparability
With a sourced pool in hand, screen it against the subject definition on the criteria that actually drive value: location within the same submarket or a truly analogous one, recency (generally transactions within the last six to twelve months, tighter in a moving market), physical similarity in size, age, class, and condition, and the same property use and income profile. A comp that fails on a value-driving attribute comes out, no matter how convenient it is to keep.
Here a general assistant is a strong first-pass filter and a weak final judge. Handed your criteria and the pool, it will rank candidates by similarity and explain each ranking, surfacing the strongest and flagging the marginal ones far faster than you would by hand. Where it stumbles is the judgment call at the margin: whether a Class B comp two submarkets over beats a Class A comp next door depends on how tenants and buyers actually substitute in your market, and a model reasoning from general patterns gets that wrong often enough that you cannot delegate it. Use the AI to narrow and to argue; make the inclusion decision yourself. The same discipline that separates a real market signal from filler in an AI-drafted market report applies here, a distinction we unpack in reading AI market reports for what is signal and what is filler.
Step 4: Adjust with evidence, not vibes
Adjustment is where AI is most seductive and most dangerous. Ask an assistant to adjust a comp set to the subject and it will return tidy percentage adjustments for size, age, location, and condition, each stated with total assurance. The direction of those adjustments is usually right. The magnitude is a guess dressed as a calculation, because the model is pattern-matching, not deriving the adjustment from market evidence.
The professional standard is the bar to hold yourself to even if no appraiser is involved. USPAP’s sales comparison approach requires that adjustments be supported by market-derived evidence, typically paired-sales analysis, not by assertion. An AI-proposed adjustment is a hypothesis. It graduates to a real adjustment only when you can point to the paired transactions, the survey, or the rent-roll evidence that supports the number. Let the assistant draft the adjustment grid and explain its reasoning; then replace every magnitude you cannot defend with one you can, or mark the comp as a directional check rather than a valuation input. On rent comps specifically, automated tools compress the survey work dramatically, and the trade-off between that speed and manual precision is worth understanding before you lean on it, which is the subject of our look at automated rent comps versus doing them by hand.
Step 5: Defend the set
A comp set is finished when it can survive a skeptical outside reader: an investment committee, a lender’s credit team, or an appraiser reviewing your assumptions. Defensibility is a property of the whole set, not of any single comp. It means every included transaction traces to a source record, every exclusion has a stated reason, every adjustment has evidence behind it, and the resulting value sits inside the range the comps bracket rather than outside it.
AI is useful for the last mile: a general assistant will draft the “why these comps” narrative your memo needs in a fraction of the time, pulling from the grid and reasons you have already validated. That is the right use of the reasoning layer, drafting an argument from vetted inputs, not manufacturing them. A lean shop that runs this discipline produces a comp set as defensible as an institutional one without an appraisal department, the kind of edge a small firm gains from doing the fundamentals well with AI on the tools, a theme in our broader work on how small CRE firms out-operate larger competitors. For where comp selection fits inside the wider workflow, the deal-analysis playbook for lean teams puts the pieces in sequence.
Where AI actually earns its place
Read across the five steps and a clean division of labor emerges. AI is a force multiplier on the mechanical work and a liability on the judgment. Assigning each task to the right owner is what makes the whole framework safe to run fast.
| Task | AI does this well | A human must own this |
|---|---|---|
| Define the subject | Structure a description, flag missing attributes | Decide which attributes drive value |
| Source the pool | Deduplicate, standardize, structure an export | Choose the data source; supply the data |
| Screen comparability | Rank candidates, explain rankings, flag outliers | Make the inclusion or exclusion call |
| Adjust | Draft the grid, propose direction, explain logic | Set and defend every magnitude with evidence |
| Defend | Draft the comp narrative from vetted inputs | Sign off on the set and the value |
That split holds across every step: AI compresses the assembling, normalizing, ranking, and drafting; the person keeps the deciding. A firm that draws the line here gets most of the speed with none of the exposure.
When the framework strains: thin markets and unique assets
The framework degrades gracefully where AI tends to fail silently. In a thin market, a rural submarket, a rare asset class, a quiet period, there may be too few genuine comps to select from at all. A disciplined process makes that scarcity visible: the screen empties out, and you know you are in weak-comp territory. An AI asked for comps in the same market will often fill the vacuum with distant or fabricated transactions rather than report that the data is not there.
The response is the one an experienced appraiser reaches for. Widen the geography or time window deliberately and adjust harder for the added difference, lean more on the income approach where sales comps run out, and state the thinness of the set as a limitation rather than paper over it. The framework does not manufacture certainty the market does not support; it tells you honestly when to trust the comps less, which is precisely the moment an unsupervised AI will tell you the opposite.
The one rule that keeps AI-assisted comps safe
If you adopt nothing else, adopt this: an AI may assemble, rank, and explain comps, but it may never source or finalize them. Every comp must originate in a real transaction record, and every value that leaves your shop must have a human owner who can defend each inclusion and each adjustment. That single rule contains AI’s two failure modes, fabricated transactions and false-confidence adjustments, while keeping all of its speed. Comp errors do not stay contained: a mis-selected comp moves a value a lender sizes a loan against and an investor commits capital on. The framework is what lets a lean firm run comps fast and still stand behind the number.
FAQ
What is a comp selection framework?
A comp selection framework is a repeatable sequence for choosing comparable transactions that produces a defensible value. The version in this article has five steps: define the subject property, source a candidate pool from real data, screen it for genuine comparability, adjust the survivors with market evidence, and assemble a set that survives outside review. Following the sequence in order means every comp is included, excluded, or adjusted for a stated reason rather than a hunch.
Can I use ChatGPT or Claude to pick comps?
Use a general assistant to organize, rank, and explain comps, never to supply them. A language model has no proprietary transaction database, so asking it directly for comps invites confident, specific, fabricated transactions. Feed it a real export and it will deduplicate, normalize, rank, and draft the narrative quickly and safely. The rule: the AI reasons over data you hand it; it never invents the data.
How does AI change comp selection in commercial real estate?
AI changes the cost of the mechanical work, not the standard for a good comp. It compresses the hours spent assembling, standardizing, ranking, and narrating a set, which lets a lean firm produce more sets faster. It does not change what makes a comp comparable or an adjustment defensible, and it introduces new failure modes, fabricated transactions and overconfident adjustments, that the selection framework exists to catch.
What are the failure modes of AI-generated comps?
There are three. First, fabricated transactions, when a model produces comps from memory instead of data. Second, false-confidence adjustments, when it states an adjustment magnitude with total assurance but no market evidence behind it. Third, thin-market breakdown, when it fills a genuine data vacuum with distant or invented comps rather than reporting that few real ones exist. A disciplined selection process surfaces all three; unsupervised AI hides them.
How do I make an AI-assisted comp set defensible?
Hold it to the professional standard even if no appraiser is involved. Every included comp must trace to a source record, every exclusion needs a stated reason, and every adjustment must be supported by market-derived evidence such as paired-sales analysis, the bar USPAP’s sales comparison approach sets. Let AI draft the grid and narrative, then replace any adjustment magnitude you cannot support and confirm the value sits inside the range the comps bracket.
Which data sources should AI-assisted comps come from?
Match the source to the comp type. Sales comps come from platforms like CoStar or Crexi; verified lease terms come from an exchange like CompStak, which sources rate, escalations, free rent, and tenant-improvement data from executed leases; automated multifamily rent comps come from tools like HelloData. County records and your own closed deals are also valid. The point is that the transactions originate in real data, after which an assistant can safely structure and reason over them.
Does the framework work in thin markets or for unique assets?
Yes, and its main value there is honesty. In a thin market or for a rare asset, a disciplined screen empties out and tells you there are too few genuine comps, whereas an AI asked for comps will often fill the gap with distant or fabricated ones. The response is to widen geography or time deliberately, adjust harder for the added difference, lean more on the income approach, and state the scarcity as a limitation rather than hide it.
What should a human never delegate to AI in comp work?
Never delegate three decisions: which attributes drive value for the subject, whether a given comp is genuinely comparable, and the magnitude of every adjustment. These are the judgments that determine the value, and a model reasoning from general patterns gets them wrong often enough that they must stay with a person. Delegate the assembling, normalizing, ranking, and first-draft narrative freely; keep the deciding.
Key takeaways
- AI does not change what a good comp set is; it lowers the cost of producing a bad one that looks finished, which is why a selection framework matters more with AI in the loop, not less.
- The framework is five gated steps: define the subject, source the pool from real data, screen for comparability, adjust with evidence, and defend the set. A comp advances only after passing the prior step.
- Comps must originate in a data source, never in a language model’s memory. AI reasons over data you supply; it never invents transactions.
- The safe division of labor is consistent: AI assembles, normalizes, ranks, and drafts; a human decides which comps belong and owns every adjustment magnitude with market-derived evidence.
- The framework degrades gracefully in thin markets by making comp scarcity visible, exactly where unsupervised AI fabricates certainty the market does not support.
Not sure where AI belongs in your firm’s underwriting and comp work, or where it is quietly introducing risk? A short assessment maps that against your deal volume, property types, and markets faster than any tool comparison. Book your free AI-readiness assessment →
Dirk Jan van Veen, PhD
