Before You Automate a CRE Workflow, Get Specific About the Work

An AI system produces an underwriting summary in minutes. The analyst then spends an hour checking its numbers, reconciling conflicting documents, and rebuilding the assumptions.

To know whether AI saves time, measure the whole job: producing the draft, checking it and making corrections.

That is why an AI project should start with what slows the whole process down. In commercial real estate, that may be document review. It may also be a missing rent roll, an ambiguous policy, or a file waiting for approval. Each requires a different change to the process.

Find out where the time goes

A process map becomes more useful when every stage has a timestamp.

Consider an illustrative borrower-package workflow taking 80 business hours from receipt to first-pass completion. Four hours involve active work. The other 76 involve waiting for documents, clarification, or internal review.

If AI halves the active work while waiting remains unchanged, completion takes 78 hours. Handling time falls 50%. The borrower gets an answer just 2.5% sooner.

The two improvements solve different problems. Less handling time releases analyst capacity. Faster turnaround requires addressing the delays that hold up completion.

Record receipt, first-touch, clarification, handoff, and completion times. Pair the logs with operator interviews: a timestamp shows that the file waited; the analyst explains why.

Look at the median and the 90th percentile, the time within which 90% of cases finish. The difficult packages can disappear inside a reassuring average.

Focus on the step that causes the delay. If incomplete packages create the longest delays, earlier completeness checks may matter more than faster memo drafting.

Resolve the ambiguity before automating it

Two analysts can read the same operating statement and calculate different net operating income. One excludes a nonrecurring expense. The other follows a policy that requires approval for that adjustment.

Giving both analysts a faster extraction tool leaves the disagreement intact.

Define the accepted output and the rules behind it: the reporting period, permitted adjustments, authoritative documents, and exceptions requiring approval. BOC Group’s process-modeling framework supports this discipline: document the detail needed for the intended use. Read the framework.

Then assign the work appropriately. AI can locate a figure and its supporting passage. Software can calculate a ratio from validated inputs. An accountable reviewer approves the adjustment. This follows Anthropic’s guidance to use the simplest architecture that meets the task. Read the guidance.

If reviewers disagree about what “correct” means, settle that disagreement before measuring the model’s accuracy.

A system can be 98% accurate and catch nothing

Imagine your accounting team receives 100 invoices. Ninety-eight are legitimate. Two are duplicates of invoices already paid, each for $50,000. These figures are illustrative.

An AI reviewer approves all 100.

It made the correct decision on the 98 legitimate invoices. Its accuracy score is 98%. But it missed both duplicates, allowing $100,000 in duplicate payments through this review.

The score looks good because most invoices were fine to begin with. The system gets credit for approving routine invoices even though it failed to do the job it was deployed for: catching payments that should not be made.

For this task, start with a different question: Of the duplicate invoices present, how many did the system catch? In this case, zero out of two.

Then measure the dollars at risk and the time spent investigating legitimate invoices flagged by mistake. Together, those measures tell you whether the review is protecting cash at an acceptable operating cost.

Keep the timer running through review

For an acquisition memo, measure the time spent preparing the inputs, generating the draft, checking sources and making corrections. Include any rework requested by the next reviewer. The relevant comparison is the effort needed to reach an accepted memo, not just a first draft.

Compare assisted and unassisted cases of similar complexity, using the same quality standard. Where practical, randomly assign cases and have reviewers assess outputs without knowing how they were produced. Giving one analyst the same package twice creates a familiarity advantage that can be mistaken for an AI benefit.

Separate results by task and experience level. A 2025 study of 5,172 customer-support agents found a 15% average productivity increase from AI assistance, with larger benefits for less experienced workers. The most experienced and highest-skilled workers saw small speed gains and small quality declines. That is evidence from one support operation, not a CRE forecast, but it shows why an average is insufficient. Read the study.

Test incomplete files and conflicting versions alongside routine cases. Passing 60 tests without an error does not prove the system is error-free. Even if those cases are independent and representative, the underlying error rate could still be about 4.9% at the upper end of a one-sided 95% confidence interval. That calculation also assumes the failure rate stays constant.

A clean test run can still leave material uncertainty. Keep difficult cases in the evaluation set and rerun it when the system changes.

Decide how the business will use the time saved

Consider an illustrative team processing 200 packages monthly. Staff time falls from 120 to 55 minutes per package, including review and corrections.

That releases approximately 217 hours a month. If staff time costs $75 per hour, including benefits and other employment costs, those hours represent $16,250 of work capacity before the costs of the AI system.

The payroll bill has not automatically fallen by $16,250.

The business only benefits if that time is put to use: more opportunities screened, faster borrower responses, reduced overtime, or avoided hiring. The relevant outcome depends on the team’s constraint. More underwriting capacity will not create more completed deals if the bottleneck is sourcing or committee availability.

Subtract implementation, software, model usage, maintenance, and oversight costs. Adjust for the share of work actually eligible for the system and how consistently it is used.

Count the time released, then verify what the business did with it. Report capacity, cash savings, and revenue effects separately.

Use the right company rules and records

A company brain earns its place when it resolves a specific uncertainty: which underwriting policy applies, which drawing revision governs the estimate, or whether an entitlement condition was proposed or adopted.

Retrieving a relevant document is only part of the job. The system must check whether the document is approved, when it applies and who is allowed to see it. Otherwise, an old committee exception can become today’s assumed policy, or an estimate can inherit exclusions from the wrong project.

Assign an owner to the governing rules. Record corrections, review them, and test changes before applying them to future work.

Treat source authority as part of accuracy. A correct quotation from an obsolete policy can still produce the wrong recommendation.

Redesign the operation, then measure what changed

Return to the file that takes 80 hours to complete. Halving its four hours of analysis leaves 76 hours of waiting untouched. To materially improve turnaround, the intervention has to reach the missing documents, unclear rules, and approval queues that hold the file up.

That is how Mason approaches the work: understand the operation, redesign how work moves through it, and build measurement into the change from the start.

Employee interviews and a review of workflow records help reveal where work changes hands, why it gets stuck and which rules were never written down. Current project records and approved rules give each step the evidence and instructions it needs. Automation inside existing systems can check completeness at intake, prepare the evidence, and route exceptions to the person authorized to resolve them. The process owner retains control.

Capture the baseline during discovery. After deployment, combine system data with staff feedback to establish whether waiting time, review effort, errors, and cost actually fell. When results get worse, find out why and use what you learn to improve the process.

Success means the team can complete the work more reliably and at a lower total cost. Ongoing measurement tells you whether that improvement holds as people, policies and AI models change.