Updated 4 November 2026.
Short answer: on five real offering memorandums, a general AI model and Реалист Underwriting (fast screening) each found all 23 key issues in our answer key. The general model did it in about 1.5 minutes of model time per memorandum on clean text; Реалист took a median of 83 minutes. What Реалист added is its own NOI, a cross-check against records on file, a verdict on the asking price and broker questions ranked by value, each tied to a page and a figure. The general model made one wrong number in five answers. This is a small experiment, not a benchmark. For the checks themselves, see как проверить цифры в инвестиционном меморандуме.
The table
| Incoming offering memorandum, 5 real listings | Аналитик вручную | Person + general AI model | Реалист Underwriting (fast screening) |
|---|---|---|---|
| Key issues found (of 23) | not run | 23 | 23 |
| Secondary issues found (of 17) | not run | 17 | 12, plus 4 partial |
| Wrong numbers | not run | 1 | 0 |
| Вопросы к брокеру | not run | 15 per memorandum, specific, not ranked | 11 to 15, ranked by value, with page and figure |
| Own NOI and verdict on the asking price | not run | not given | 5 of 5 |
| Your time | about 120 min* | about 30 min* | about 20 min* |
| Machine time | нет | about 1.5 min | median 83 min |
*Estimates, not measured. Footnote: experiment on five offering memorandums from open broker listings, answer key fixed before the runs, one run per cell. The "person + general AI model" column is a general AI model (Claude) given the text of each memorandum with the same prompt, no web and no comparables; no person was timed, so the times for people are estimates. The prompt is published below. Реалист machine times include queueing. Small sample: read it as a measured experiment, not a benchmark.
Per memorandum
Key issues found:
| Меморандум | Реалист | General AI model |
|---|---|---|
| 34-unit multifamily, Kansas | 5 of 5 | 5 of 5 |
| 20-unit multifamily, Texas | 5 of 5 | 5 of 5 |
| 216-unit multifamily, Massachusetts | 4 of 4 | 4 of 4 |
| 37-unit multifamily, Georgia | 4 of 4 | 4 of 4 |
| 72-unit multifamily, Georgia | 5 of 5 | 5 of 5 |
Where the general model is not worse
- It found every key issue and every secondary issue on the five memorandums.
- It took about 1.5 minutes of model time per memorandum, against a median of 83 minutes for Реалист.
- It covered 30 of the 31 key broker questions, in one pass, with long due-diligence lists (litigation, environmental, contract terms).
- It also caught a loss-to-lease column that does not add up, which was not in the key.
- For a 30-minute first pass on clean text it is a fair tool.
Where Реалист is ahead, and where it is weak
- Ahead on what comes with the answer: its own NOI against the one in the memorandum, a verdict on the asking price, and a cross-check against records on file. On the 72-unit Georgia memorandum it showed an own NOI of $171,436 against $185,926 in the T-12. On the 37-unit Georgia memorandum it flagged a property record at 1,394 square feet built in 1946, where the memorandum said 26,594 and 1984.
- Ahead on the questions: ordered by value, each with its page and figure.
- Not shown here: how a person gets a 20 to 30 page memorandum into a chat. The general model was handed clean text from each PDF. Pasting a long memorandum can hit size limits or lose tables and images; we did not test that.
- Weak on time: 70 to 100 minutes of machine time per memorandum, so it is not instant.
- Weak on coverage: 12 of 17 secondary items found, plus 4 partial. It missed per-unit rent comparisons, a bonus loft and demographics.
Метод
- Five real offering memorandums from open broker listings. No client documents were used.
- The answer key was written before any run: rent roll, offering memorandum and T-12 mismatches, occupancy, loan term, tax, risks, and the questions an experienced underwriter would ask. The file hashes were fixed with it.
- The same prompt for both AI columns:
Attached is an offering memorandum for a N-unit apartment property at ADDRESS. Please review it for me. Summarize the deal, check the numbers for errors or inconsistencies, tell me what looks risky or missing, and write a list of questions I should send to the broker.
- Scored by item against the key: issues found, wrong numbers (recomputed from the figures each answer quoted), and questions that match the key.
- The general-model answers came from one session of a general AI model on the extracted text of each memorandum. That session had seen the earlier headline counts, not the key items, before writing; the key was opened after all five answers were written. The same session scored them.
- The analyst column is the catalog estimate of two hours for a light screening, not a measured run. Times for people in every column are estimates.
- One run per cell, five memorandums.
Вопросы и ответы
Can an AI model screen an offering memorandum?
Yes, as a first pass. In this experiment a general AI model found 23 of 23 key issues in about 1.5 minutes of model time per memorandum, and Реалист found 23 of 23 in a median of 83 minutes.
What does Реалист add over a general model?
Not more key issues in this test. It adds its own NOI, a cross-check against records on file, a verdict on the asking price and broker questions ranked by value with page and figure. More on where a general chat model stops: can ChatGPT underwrite a commercial real estate deal.
How do I send a memorandum for screening?
Forward the offering memorandum. It comes back screened, with a list of questions for the broker, like this broker questions sample. Запросить демо to try it on one from your inbox.