An incomplete deal needs a next question

An incomplete deal needs a next question. It shouldn't disappear because a screening workflow converted an empty field into a failed criterion.

- Sasha Deneux

The Take: Faster screening needs a visible review queue

NIST's January 2023 AI Risk Management Framework places evaluation in the context where a system will be used. For deal screening, that means testing the incomplete submissions and ambiguous documents an acquisition team actually receives.

Processing more broker emails can produce more structured records. The design question is what happens to each record once it exists. A queue that produces more confident labels without explaining them simply moves the review problem downstream.

Consider an OM with an unclear occupancy figure. The workflow might read the wrong column, fail a buy-box threshold, and mark the property as a pass. If nobody sees the evidence behind that decision, the team has lost a potentially relevant deal without making a conscious choice.

I want three distinct outcomes: a verified mandate mismatch, insufficient evidence, and ready for analyst review. Those states tell people what to do next. They also prevent the score from hiding the difference between an unattractive asset and an incomplete submission.

The same discipline applies to sourced properties. A public record can justify another research step without establishing that the owner wants to sell. Keep the observation and the conclusion separate as the property moves toward contact research.

Before expanding volume, inspect the records the process would have rejected. Review time is one useful measure; false rejection is another. Write down which errors changed an outcome and which only changed formatting. That gives the team a practical basis for improving the queue before sending more opportunities through it.

The Teardown: Give every screened deal a next action

Use a practice set of previously reviewed opportunities, including attractive deals, clear mismatches, and incomplete files. Write down the expected outcome before running the workflow. That keeps the evaluation from shifting to accommodate whatever the model produces.

Version the mandate. Capture asset type, geography, size, and other agreed criteria with their definitions. Distinguish a preference from a hard constraint. A new mandate version should not silently rewrite the rationale for an older screening decision.

Record evidence beside the criterion. Save the extracted value, the relevant file and location, and the reporting period. A field that is present can still be wrong. Review material values and reconcile them with the source instead of relying entirely on warnings or confidence labels.

Separate incomplete from unsuitable. When a critical input is missing or contradictory, route the record to a clarification queue. Name the question and assign an owner. When a mismatch is supported, show the criterion and evidence so the analyst can confirm or override the result.

Make the handoff concrete. A record ready for review should carry the documents, unresolved items, proposed next action, and the mandate version used. Draft any broker follow-up for review. Processing the inbound file does not itself authorize sending a message or passing on the opportunity.

Inspect the rejected pile. Have the reviewer check proposed rejections as well as promoted deals during the pilot. Record wrong extractions, misapplied rules, missing evidence, and overrides separately. If a revised rule changes the outcome, preserve the earlier decision and its reason.

For off-market research, use the same handoff principle: a dated observation, a verified property identity, a current owner-research status, and a next check. Avoid allowing a research clue to turn into a claim about an owner's intentions as it passes between tools.

The CRE screening playbook describes the intake review. Apply the same evidence standard to sourced properties before their research records enter the acquisition pipeline.

Signal

Reference desk: three ideas to bring into an intake pilot.

• January 2023: retain the test conditions. NIST's AI Risk Management Framework calls for documenting evaluations and their limits. Keep the mandate, document set, and model version with your results. Otherwise, a changed outcome can be difficult to explain.
• July 2024: verify generated evidence. NIST's generative AI profile recommends checking sources and citations. Open the occupancy table before accepting a rejection based on occupancy. The review needs to cover declined opportunities as well as the shortlist.
• April 2013: preserve the steps behind a result. W3C's PROV overview describes how information about an output's origins supports assessment. Keep the extracted observation, applied screening rule, and reviewer decision together so a later override has a clear starting point.

From NextAutomation

We build sourcing and intake workflows around your buy box and analyst handoff. Book a conversation with an example of a deal that was difficult to classify. The off-market sourcing pack provides a starting structure for research and review.

Give the uncertain deal an owner and a next question.

Next edition: evaluating the data path and the assumptions before buying another AI tool.

- NextAutomation Team