Automating a Bad Buy Box Only Makes You Wrong Faster

Automation does not improve buy box criteria quality. It removes the last thing protecting a firm from its own bad criteria, which is inconsistency. A principal who says the firm buys nothing built before 1985 still takes the call when a broker sends a 1978 building at the right basis. Software does not take the call. It applies the rule to every deal, with no exceptions and no memory of why the rule exists. A buy box is a hypothesis about which deals deserve attention, and encoding an untested hypothesis turns a tendency into a policy.

Key Takeaways

• A buy box is a set of hypotheses about which deals deserve attention, and almost no firm has tested whether those hypotheses are correct.

• Human screening is inconsistent, and inconsistency is an accidental safety valve. Automation removes it, so criteria quality becomes the only variable left.

• A criterion earns the right to auto-decline only if its rejections are almost never wrong and the fact is verifiable from the offering memorandum. Everything else belongs in the sort order, not the filter.

• Criteria written in one market regime keep running after it ends. MSCI data reported by Colliers shows garden apartment volume down 21 percent year over year in Q2 2026 while industrial and hospitality each rose 27 percent.

• Testing a criterion is cheap: pull the deals it alone rejected, underwrite a sample, count how many would have cleared.

What makes a buy box criterion good or bad?

A good criterion rejects deals the firm would never have bought and almost nothing else. A bad criterion correlates loosely with something the firm cares about and kills real fits as collateral damage. The difference is measurable, and most firms have never measured it, because rejections leave no record behind.

Take two rules from the same buy box: no assets built before 1985, and debt service coverage of at least 1.25x at the lender's stressed rate. Only one is a fact about the deal.

Vintage is a proxy. The firm does not object to 1978 construction. It objects to deferred capital expenditure, obsolete systems, and layouts that will not lease. Vintage predicts those imperfectly. A 1978 building with a 2019 roof, new mechanicals, and a repositioned lobby is the asset the rule was written to find, and the rule kills it.

Debt service coverage is a computed property of the deal at stated assumptions. When it fails, the deal fails. Sorting proxies from facts is the most useful pass to run over a buy box before any of it is automated.

Where do bad buy box criteria come from?

They accumulate. Nobody sits down and writes a bad criterion. A buy box collects rules from fund documents, from one painful deal, from a capability the firm no longer lacks, and from market conditions that ended two years ago. Nothing removes them, so the box only ever tightens.

The last row is the quiet killer, because it fails without a visible error. A return threshold calibrated when appreciation did half the work declines deals that pencil on income alone. CBRE put the average U.S. cap rate near 6.3 percent in its Q2 2026 capital markets figures. A buy box built against 2021 pricing screens a market that is no longer there.

How do you test whether a criterion is any good?

Attribute every rejection to the specific criterion that killed it, then underwrite a sample of the deals each criterion rejected on its own. The share that would have cleared underwriting is that criterion's error rate. Run it once and the buy box stops being an opinion.

Work an example from stated inputs. A firm screens 1,240 offering memorandums over eighteen months, logging the criterion behind each rejection. The vintage cutoff rejected 214 deals on its own. Pull 40 at random and underwrite them properly: seven clear the return threshold, a 17.5 percent error rate, implying roughly 37 real fits across the 214. The debt coverage floor rejected 180 on its own. Pull 40: one clears, a 2.5 percent error rate, implying about 5 fits across the 180.

Same buy box, same firm, same eighteen months. One rule is wrong seven times in forty. The other is wrong once. Automating both at the same confidence is the error, and it stays invisible until somebody attributes rejections to criteria and pays for the models.

That cost is knowable in advance: eighty deals at four analyst hours each is 320 hours, or $40,000 at a fully loaded $125 per hour. Set against a rule discarding 37 fits, the test is cheap. Most firms will not spend it, which is why false negatives stay the expensive error in automated screening: nobody funds the measurement that would make them visible.

Which criteria should be automated at all?

Only the ones verifiable from the document and almost never wrong. Everything else belongs in the ranking, not the filter. The distinction matters because a filter destroys information while a ranking preserves it. A deal sorted to position 400 can still be retrieved. A deal auto-declined is gone.

Most buy boxes put all four rows in the first category, which is what makes automating them dangerous. A queryable buy box makes criteria enforceable, and enforceability is what a bad criterion should not have. Structure is still necessary, as covered in what a buy box is and why a memo version cannot screen a deal. Applied to unvalidated rules, it converts occasional error into consistent error.

What does a stale criterion cost when the market moves?

It costs the entire segment the criterion excludes, for as long as nobody revisits it. Buy boxes are written in one market and enforced in another. That gap is invisible from inside the screening process, because the screen reports what it rejected and never what the rejection was worth.

The current divergence makes the point. MSCI Real Capital Analytics data summarized by Colliers put U.S. investment volume at $113.7 billion in Q2 2026, up 9 percent year over year. Beneath that number, industrial and hospitality volume each rose 27 percent while multifamily stayed flat and garden apartment volume fell 21 percent. Pricing split the same way: the RCA CPPI rose 0.9 percent across all property types while apartments fell 1.7 percent.

A buy box written in 2021 around garden multifamily in the Sun Belt now points at the softest part of a recovering market. Nothing about it broke. The market moved underneath it. A human screening those deals would have felt the drift. An automated screen produces no such signal, which is why a quarterly audit of the offering memorandums a firm received and what it did with them catches stale criteria when nothing else will: it looks at the rejected pile rather than the advanced one.

Frequently Asked Questions

What is a bad buy box criterion?

A rule that rejects deals the firm would have bought. Most bad criteria are proxies: they stand in for something the firm cares about, like deferred capital expenditure, without establishing it. They look precise and reject real fits as collateral damage.

Should a firm fix its buy box before automating deal screening?

Yes, at least for the criteria that will auto-decline. Consistency amplifies whatever error the criteria already contain. Criteria never tested against the firm's own rejections should reduce a deal's score rather than eliminate it.

How often should acquisition criteria be reviewed?

Quarterly for thresholds tied to market pricing, such as going-in yield and debt coverage floors. Annually for structural criteria like asset class and geography. Any criterion that survives two years untested should be treated as unvalidated.

Does AI make deal screening criteria better?

No. A model applies criteria, it does not judge whether they are the right ones. Extraction and matching remove the labor. The judgment in the buy box is the firm's, and automation inherits it exactly as written.

Conclusion

The conversation about screening automation is about speed, and speed was never the binding constraint. A firm that screens 1,240 offering memorandums against criteria it has never tested owns a fast, consistent, defensible process for reaching the wrong conclusion. That is worse than the manual version, which at least had a principal who sometimes ignored the rule and was sometimes right to.

The work that pays is unglamorous: attribute every rejection to a criterion, underwrite a sample of what each criterion killed, demote the rules that turn out to be proxies. Do it once and the firm knows which parts of its buy box are facts and which are habits. Automate only the facts. As precision becomes the real constraint in deal screening, the edge belongs to firms that can say why each rule in the box is there.