Yes — if the committee treats it as one interrogated input, not a stamped conclusion. Check the same things you would check in any opinion of value: sourced comparables, a disclosed confidence level with a stated reason, and what happens when a figure is wrong. It is an opinion of value, not an appraisal, and should never stand alone.
The question underneath the question
"Is the AI good?" is not what a committee actually asks. A committee asks the same thing about a broker opinion of value regardless of who or what produced it: where did each number come from, how confident is the preparer and why, and what happens if a piece of it turns out to be wrong. An AI-generated opinion of value passes or fails that test on the same grounds a human-prepared one does. The interesting question is not whether a model can produce a plausible-looking number — most can — but whether the process behind it survives the same interrogation a committee already runs on every opinion of value that crosses the table.
What a careful committee typically checks, before relying on any output
This applies to an opinion of value from any preparer, ours included, and describes practice rather than a specific lender's or regulator's policy.
- Where each comparable came from. Public record, transaction data, or an internal database — and whether the sample size behind the conclusion is stated next to the number, not buried. A single comparable dressed up as a market conclusion is a common failure mode, AI-generated or not.
- A confidence or risk figure with a stated reason, not just a score. A number on its own ("85/100") tells a committee nothing about what it's missing. What tells a committee something is the preparer naming what's thin — an unresolved zoning question, a small transaction sample, missing occupancy data — and what would need to happen to move the number.
- Which inputs are observed and which are assumed. Rent, expenses, growth rates, and cap rates are frequently assumptions dressed as facts in a polished narrative. A committee typically wants those two categories kept visibly separate, not blended into a single clean number.
- A correction record. Has this preparer ever caught and disclosed its own error, and what changed afterward? A shop that has never disclosed an error is not necessarily a shop that has never made one — it may just be one that doesn't show its work.
- An as-of date, on the underlying market data and on the run itself. An opinion of value ages; a committee checks how old the inputs are before relying on the conclusion.
- Reproducibility. Can the committee open the underlying calculation — a workbook with visible formulas — or only a formatted narrative or a scanned document nobody can check?
- What the preparer is licensed to call it. An opinion of value prepared without an appraisal licence is not an appraisal. A committee benefits from knowing, up front, which one it is holding.
How our own practice answers that checklist
We don't ask a committee to take our word for the list above; the clearest way to show it is a real deliverable, not a description of one. In one internal underwriting model — a workbook built for a Tampa multifamily property — every input cell was classified as one of three things, and nothing else was allowed on that classification tab: sourced from a public record, assumed, or computed from the other two. That classification is what let the workbook's own reviewers separate what the model observed from what it had to guess, line by line, rather than trusting a summary paragraph.
The same workbook is also the clearest real example of the correction side of the practice. An early draft defaulted its purchase-price input to the property's county tax-assessed value — a figure that exists to calculate property tax, not a number anyone would pay for the building. Because the income side of the model and the price side were now built on two different bases, the arithmetic connecting them reported the gap between those bases as if it were investment return: a levered return far outside a defensible range for the asset, on evidence that looked, at a glance, like a good deal.
The error was caught by an independent recalculation, not by a reader noticing the number looked too good. The fix had three parts, and all three are visible in the file itself rather than described in a cover note: the price was rebuilt off the property's own income at a stated going-in capitalization rate, an assessed-value comparison that would have repeated the same basis mismatch on the exit side was removed, and four automated checks were added directly to the workbook — on yield, on the exit cap relative to the entry cap, on debt service coverage, and on the operating expense ratio — each of which would have fired on the very first draft. The point of building the checks into the file, rather than into a memory a future analyst has to have, is that the same mistake now trips a visible flag on the next run instead of depending on someone remembering last time.
That is what "confidence disclosure and correction practice" means here in practice: label what's observed versus assumed, build the sanity checks into the deliverable so a wrong number announces itself, and when an error is found, show the fix rather than quietly replace the number.
The honest limit — what a committee should not do with our output alone
An opinion of value we produce is not a substitute for an appraisal, and whether a given decision calls for a licensed appraisal instead is a determination for the committee and its own policy, not something this page can settle. Beyond that:
- Don't rely on a headline confidence score without reading what it's attached to — the reason behind the number is the part worth checking, not the number itself.
- Don't treat a single run as evidence the process is error-proof. The workbook example above is worth something precisely because the error is visible in it, not because the process never produces one.
- Below a handful of comparable sales, a point estimate should widen into a range rather than hold false precision — ask whether the output you were given did that, or quietly kept a single number anyway.
- Don't assume the output already reflects documents specific to your deal — a rent roll or lease-expiration schedule can't be built from public sources; if you haven't supplied your own operating data, the opinion of value is working from what's publicly observable and stated assumptions, not from your property's actual books.
- Don't extend one deliverable's discipline into a blanket claim about accuracy across every property type or market. A workbook checked line by line for one Tampa asset says something about how that workbook was built; it is not a claim about every output this process will ever produce.