Automated and Manual Underwriting Fail in Opposite Directions

Automated underwriting vs manual is framed as a contest with a winner. It is not. The two approaches fail in opposite directions, and the direction of the failure is the useful property. Automated underwriting fails silently, at scale, and in the same direction every time, which makes it systematic and therefore correctable. Manual underwriting fails loudly, one deal at a time, and differently on each occasion, which makes it inconsistent and hard to correct. A firm that understands this stops picking a side and starts designing the handoff between them.

Key Takeaways

• A systematic error is a solved problem waiting for someone to look. A random error is a permanent tax.

• Automation's danger is that the same mistake is applied to 400 deals before anyone notices. That is also why it is fixable in one pass.

• Manual review's danger is that two analysts reach different answers on the same file and neither can reconstruct why.

• Fannie Mae has run both paths side by side for decades: automated risk assessment through Desktop Underwriter, with defined recommendation types that route certain files to human judgment.

• The performance metric that matters is not accuracy. It is the exception rate and what happens to the exceptions.

How does automated underwriting fail?

Automated underwriting fails quietly and consistently. A rule that classifies a reimbursement line into the wrong expense category will misclassify it identically on every deal it touches, producing a small, stable, invisible bias. Nothing errors out. The output looks normal. The bias only surfaces when someone compares the model to realized results.

The word for this is systematic. It is the more frightening failure at first glance, because it scales: one wrong assumption in a screening rule reaches 400 deals in a quarter. It is also the more tractable one, because a systematic error has a single cause, and finding it repairs every instance at once.

Consider three real shapes this takes. A parser reads "gross potential rent" from a rent roll that reports it before concessions, so economic occupancy is overstated on every multifamily deal. A default expense ratio is applied where the operating statement is unreadable, so buildings the seller starved look identical to buildings the seller ran. A trailing statement is annualized from nine months without seasonality adjustment, so every winter-heavy asset looks cheaper to operate than it is. Each error is stable, directional, and invisible on any single deal.

How does manual underwriting fail?

Manual underwriting fails loudly and unpredictably. An analyst working at 9pm on the fourth deal of the week transposes a figure, misses an amendment, or applies a rent mark from memory rather than from comparables. The error is visible when caught and untraceable when not, because the reasoning lived in someone's head rather than in the file.

Human error is not directional. Two analysts underwriting the same asset will disagree, and the same analyst will disagree with herself in March and in September. That variance is often defended as judgment. Some of it is. Most of it is fatigue, sequence effects, and the absence of a written standard.

Read the last two rows together. Automated error is expensive to detect and cheap to fix. Manual error is cheap to detect, if anyone checks, and expensive to fix, because the fix is a person's habits. Neither is better. They are different maintenance obligations.

What does pairing automated and manual underwriting look like?

Pairing means automating to a stated confidence threshold, routing everything below it to a person, and treating the routing rate as a managed number. The machine handles the volume it can support with evidence. The human handles the residue. Neither is asked to do the other's job.

This is not a novel design. Fannie Mae's Selling Guide offers lenders two paths for comprehensive risk assessment, automated through Desktop Underwriter or manual, and the automated path returns defined recommendation types, including referrals that put a file in front of a person rather than forcing a machine verdict. The system has run at national scale for decades on the premise that automation should be allowed to decline, not required to decide. Commercial real estate has been slower to adopt the premise, though the workflow structure is identical to straight-through processing versus exception handling in an extraction workflow.

The threshold is the design decision. Set it too high and everything routes to humans, which reproduces the capacity constraint automation was meant to relieve. Set it too low and systematic error flows into the pipeline unexamined. The correct threshold is the one where the cost of a missed error equals the cost of a review hour, and that point differs between a first-pass screen and a signed letter of intent.

Why is the exception rate the real performance metric?

Because accuracy is a claim about a test file and the exception rate is a measurement of your own production. It tells you what share of deals the system could not handle with evidence, which is the honest description of how much work automation absorbed and how much it handed back.

Every figure below derives from the stated inputs. A team screens 400 deals a quarter. Manual first-pass review costs 2.5 hours per deal. Automated review costs 20 minutes per deal on straight-through files and 1.5 hours on an exception, because the reviewer inherits partial work and has to verify it.

Two readings matter here. First, even a poor 50 percent exception rate saves 63 percent of the hours, which is why teams tolerate weak systems for years without measuring them. Second, the marginal value of improvement is steepest at the bad end: moving from 80 percent to 50 percent saves 140 hours, while moving from 15 percent to 5 percent saves 46. Firms tend to chase the second improvement because it sounds like precision work, and ignore the first because it looks like admitting the system is weak.

Track the exception rate by document type and by cause. An exception rate that is 40 percent on leases and 6 percent on operating statements is not a system-wide accuracy problem. It is a lease parsing problem with a name and an owner. The same logic applies when evaluating a purchase, which is why underwriting software should be judged on speed you can audit.

Which decisions should never be automated?

The ones where the input is a forecast rather than a fact. Rent growth, exit cap rate, business plan sequencing, capital structure, and the price you will not exceed are opinions about the future. Automating them does not remove the opinion. It hides whose opinion it is and removes the person who can be asked to defend it.

There is a standards-body version of this line. The Appraisal Standards Board adopted Advisory Opinion 41 on April 23, 2026, holding that technology can assist analysis but does not relieve the professional of responsibility for a credible result, and that an automated valuation output is not by itself an appraisal. Acquisitions has no equivalent enforcement, which makes the discipline voluntary and therefore easier to skip.

The practical rule: automate the fields, never the assumptions. A rent roll line is a fact with a page number. A rent mark is a forecast with an author. Systems that blur the two produce documents where the reader cannot tell which is which, and that ambiguity is where both failure modes compound instead of canceling. The distinction is sharpest in multifamily underwriting, where the rent roll decides the deal.

Frequently Asked Questions

Is automated underwriting more accurate than manual underwriting?

It is more consistent, which is not the same thing. Consistency means the same input produces the same output, and it makes errors findable. Accuracy against reality still depends on whether the rules encode a correct view, which is a human question.

What exception rate should a firm target?

Target the trend rather than a number. A rate that falls quarter over quarter while the underlying deal mix holds steady means the system is learning your documents. A flat rate means nobody is working the exception queue back into the rules.

Does pairing automated and manual underwriting slow the process down?

No, if the routing is automatic. The delay comes from unrouted work, where a person reviews everything because there is no threshold telling them what to skip. A stated threshold with automatic routing is faster than either pure approach.

Conclusion

The automated versus manual argument persists because both sides are describing a real failure and neither is describing the other's. Machines are wrong the same way every time. People are wrong differently every time. Those are complementary defects, and a system built to exploit that lets the machine handle everything it can evidence, sends the rest to a person, and keeps score on how often that happens. Stop asking which one is better. Ask what your exception rate was last quarter, what caused it, and whether anyone owns bringing it down.

Related

• Straight-through processing vs exception handling in the extraction workflow

• Underwriting software gives you speed you can audit or speed you cannot

• Multifamily underwriting is decided in the rent roll