Are Automated CMAs Accurate? Where They Beat Manual Comps and Where They Don't

CompsAgent Team · August 21, 2026 · 11 min read

This is the accuracy companion to automated CMA vs. manual comps, which covers the time math. Speed is easy to argue about. Accuracy is the objection that actually stops brokerages from adopting, and it deserves a straight answer.

First: an automated CMA is not an AVM

Most "automated CMAs are inaccurate" arguments are actually about AVMs, and the two get conflated constantly.

	AVM	Automated CMA
Output	One estimated number	A comp set, adjustments, context, a range
Human in the loop	None	Agent reviews and signs
Shows reasoning	No	Yes — that's the point
Handles condition	No	Via the agent's walkthrough
Accountable party	Nobody	The agent

An AVM answers "what is it worth." An automated CMA answers "here is the evidence, here is what it implies, and here is where I disagree with it." Zillow's own published accuracy — around 1.74% median error on-market and 7.20% off-market — is a fair benchmark for the first and irrelevant to the second.

Stage by stage

Here's the honest breakdown, and it doesn't go one way.

Stage	More accurate	Why
Retrieving candidate sales	Automated	Queries every match; humans stop when they have enough
Transcribing figures	Automated	No typos, no transposed digits
Computing adjustments	Automated	Arithmetic applied consistently every time
Market statistics	Automated	Full dataset, correctly segmented
Selecting which comps count	Manual	Boundaries, severances, local knowledge
Assessing condition	Manual	Requires having seen the property
Thin-comp markets	Manual	Knowing when to stop rather than pad

Where automation is clearly better

Agents under-rate this half because these errors are invisible. Nobody notices a transposed square footage — the report just quietly carries a wrong adjustment through to the price.

• It doesn't stop searching early. An agent who finds four decent comps at 9pm on a Wednesday stops. A query doesn't get tired at comp five.
• It doesn't fat-finger. Manual comp tables are transcription exercises, and transcription has a nonzero error rate every single time.
• It applies the same logic to every comp. Humans unconsciously adjust harder against comps that disagree with the number they've already formed. This is the most under-discussed accuracy problem in manual CMAs.
• It computes market stats over everything. Segmented absorption and DOM by band, not eyeballed.

The manual error most likely to move your price isn't a bad comp. It's the fourth comp you adjusted a little harder because it didn't fit the number you'd already decided on. Automation is immune to that one, and agents almost never count it.

Where automation is clearly worse

• Boundaries. Distance is a proxy for similarity and a poor one. Software will happily cross a school district line, an arterial or a jurisdiction boundary because the property is 0.4 miles away. This is the most common and most consequential automated error. Screening rules here.
• Below-grade space. If the source record lumps finished basement into total square footage, an automated adjustment inherits the error and multiplies it.
• Condition. Two identical-on-paper homes, one renovated in 2023 and one original 1987, are the same property to a dataset. This is the single largest source of automated error.
• Thin markets. Software asked for six comps tends to return six. Knowing that only two are real — and saying so — is judgment.
• Non-disclosure states. Where sale prices aren't public record, anything leaning on public data has less to work with than it appears to.

The review that closes the gap

Fifteen to twenty minutes, in this order:

1. Boundary check. Look at each comp on a map against the boundaries you know. This catches the biggest error class first.
2. Substitution test. Would that buyer have plausibly bought your subject instead? Anything failing this goes, regardless of specs.
3. Square footage verification. Cross-check against a second source. Note anything below grade.
4. Condition pass. Open the listing photos for each surviving comp. Adjust against what you saw on the walkthrough.
5. Net adjustment sanity check. Anything over ~15% net is telling you it isn't comparable. Drop it.
6. Thin-market honesty. If only three survive, ship three and say so. A padded set is worse than a small one.

Do that and you have the automation's advantages plus your own. Skip it and you've published a dataset's opinion under your name.

Which errors actually cost you

Not all inaccuracy is equal, and this is where the comparison gets interesting.

Error	Typical source	Visible to client?	Cost
Comp across a school boundary	Automated	Yes — instantly	High. Local sellers know.
Transposed square footage	Manual	No	Moderate. Silent price error.
Condition not adjusted	Both	Sometimes	High. Biggest single driver.
Comp set stops at four	Manual	No	Moderate. Unknown unknowns.
Padded thin market	Automated	Sometimes	High. Range looks false.
Adjusting to fit a preconceived number	Manual	No	High and almost never caught.

Automated errors are more visible, which makes them feel worse — a seller spots a wrong-side-of-the-boundary comp immediately. Manual errors are quieter and can be equally expensive. Visibility isn't the same as magnitude, though it is the same as embarrassment.

What to ask a vendor

• Does comp selection use boundaries, or just radius?
• Is below-grade square footage separated from above-grade?
• What happens in a thin-comp market — does it pad, or report the shortfall?
• Is each comp labelled as a confirmed closed sale, a list price, or an estimate?
• Can I remove a comp and have everything downstream recompute?

That last one matters more than it sounds. If removing a bad comp means rebuilding the report by hand, the review step won't survive contact with a busy Thursday. Fuller list in the buyer's guide.

The bottom line

"Are automated CMAs accurate" is the wrong question, because it assumes accuracy is one property of a whole process. It isn't — it's a property of each stage, and the stages split cleanly.

Automation wins retrieval, transcription, arithmetic and statistics. Humans win selection, condition and knowing when the data is too thin. A reviewed automated CMA combines both and is generally the most accurate option available to a working agent. An unreviewed one is the least.

The fifteen minutes is not overhead. It's the part where the accuracy comes from.

Frequently asked questions

Are automated CMAs accurate?

They are more accurate than a human at retrieval and arithmetic and less accurate at comp selection and condition assessment. A reviewed automated CMA is typically at least as accurate as a manual one because it eliminates transcription and calculation errors while the agent still supplies the local judgment. An unreviewed one is not.

Is an automated CMA the same as an AVM?

No. An AVM produces a single estimated value from a statistical model with no human involvement. An automated CMA assembles comparable sales, adjustments and market context into a report that an agent reviews, adjusts and signs. One is a number, the other is an argument.

What errors do automated CMAs make?

Selecting comps across boundaries that matter — school districts, arterials, jurisdiction lines — because distance is used as a proxy for similarity. Treating below-grade square footage as living area. Missing condition differences entirely. And producing a full comp set in a thin market where the honest answer is that few real comps exist.

How long should reviewing an automated CMA take?

Around 15 to 20 minutes. Check each comp against local knowledge, remove anything that fails the substitution test, verify square footage against a second source, and apply condition adjustments from the walkthrough. Skipping this step is what makes automated reports inaccurate.

Related reading

• Automated CMA vs. Manual Comps: What a 30-Agent Office Actually Saves
• Is AI Actually Good at Real Estate Comps Yet?
• CMA vs. Appraisal: The Difference That Costs Deals
• How to Find Real Estate Comps Without Guessing