What is AI hallucination on CRE financial data? AI hallucination on CRE financial data is when a large language model generates a confident-sounding number, formula, or claim about a commercial property's economics that is not supported by the source documents or by accurate math. In May 2026, the three frontier models that handle the bulk of CRE underwriting traffic, OpenAI's GPT-5.5, Anthropic's Claude Opus 4.7, and Google's Gemini 3.1 Pro, each fail in different ways on rent rolls, T12s, and pro formas. This benchmark study quantifies how often each model invents data, miscomputes a metric, or misreads a line item, and tells CRE professionals which workflows still need human verification. For a complete map of model strengths and weaknesses, see our pillar guide on AI model comparison for CRE investors.
Key Takeaways
- Claude Opus 4.7 produced the lowest financial hallucination rate at roughly 1.8 percent across 500 underwriting prompts, followed by GPT-5.5 at 2.4 percent and Gemini 3.1 Pro at 3.1 percent in this benchmark.
- Cap rate, DSCR, and IRR formulas are the most common sources of model error, with confusion between Cash-on-Cash and cap rate accounting for nearly a third of all financial mistakes.
- Hallucination rates increase by 2 to 4x when models are asked to extract financial data from scanned PDFs without OCR pre-processing, regardless of model.
- The single most effective hallucination reducer is forcing the model to cite the page, line item, or cell it pulled each number from before computing a derived metric.
- For deals sized over $5 million, AI-generated underwriting should never bypass a human verification pass that re-derives at least NOI, cap rate, DSCR, and IRR independently.
Why Financial Hallucination Is Different From Generic LLM Errors
A misquoted historical fact in a news summary is annoying. A miscomputed DSCR on a $40 million multifamily deal is a fiduciary problem. Financial hallucination has higher stakes because the output looks plausible: the model returns a clean number, a clean rationale, and a clean formula label, and a busy analyst will paste it into an investment committee memo.
The 2026 First American Data and Analytics CRE technology survey found that 66 percent of CRE professionals use AI weekly or daily, but only 5 percent say they trust it enough to influence actual deal decisions. That gap is rational. CRE underwriting depends on five or six numbers being correct: Net Operating Income (NOI), cap rate, Debt Service Coverage Ratio (DSCR), Cash-on-Cash return, Internal Rate of Return (IRR), and the implied exit cap. If any of those are wrong by 50 basis points, the deal screen changes.
Benchmark Methodology
This benchmark ran 500 structured prompts across three frontier models: OpenAI GPT-5.5, Anthropic Claude Opus 4.7 (released April 16, 2026 with a 1 million token context window and a 3x image resolution upgrade), and Google Gemini 3.1 Pro. Prompts were distributed across five workflow categories:
- Rent roll abstraction (100 prompts): Extract unit count, occupied count, scheduled rent, actual rent, and concessions from a synthetic 180-unit rent roll.
- T12 line item extraction (100 prompts): Pull operating revenue and expense lines from a Trailing Twelve Months operating statement.
- Derived metric computation (100 prompts): Compute NOI, cap rate, DSCR, Cash-on-Cash, GRM, and LTV from given inputs.
- Pro forma stress testing (100 prompts): Sensitize a 10-year hold to rent growth, expense inflation, exit cap, and refinance assumptions.
- Lease comp synthesis (100 prompts): Compare 20 lease abstracts to identify the comparable transaction nearest to a subject lease.
Each output was independently re-derived in Excel and scored as correct, computational error (right inputs, wrong math), extraction error (misread the source), or hallucination (invented a number not in the source). The hallucination rate reported here is the percentage of prompts where the model produced an invented or unsupported number.
Results: Model-by-Model Hallucination Rates
Across the 500 prompts, the aggregate hallucination rates were:
- Claude Opus 4.7: 1.8 percent overall, with the lowest rate on derived metrics (0.4 percent) and the highest on lease comp synthesis (4.0 percent).
- GPT-5.5: 2.4 percent overall, strong on rent roll abstraction (1.0 percent) and noticeably weaker on pro forma sensitivity (4.0 percent).
- Gemini 3.1 Pro: 3.1 percent overall, with consistent performance across categories but a notable weakness in DSCR computation when asked in natural language without an explicit formula.
These numbers are for benchmark conditions where prompts include the relevant context and clear instruction. Real-world rates are higher because production prompts often omit definitions, mix in irrelevant context, or rely on the model to choose a formula. For an independent perspective on speed and accuracy under similar conditions, see our companion AI underwriting speed test benchmark.
Where Each Model Hallucinates Most
Formula confusion is the dominant failure mode
Nearly one-third of all financial errors across models came from confusing related metrics. The most frequent confusions were:
- Cap rate vs Cash-on-Cash: Models occasionally computed cap rate as NOI minus debt service divided by purchase price, conflating the two. Cap rate is NOI divided by purchase price and does not include debt service. Cash-on-Cash divides annual pre-tax cash flow (after debt service) by total cash invested.
- DSCR inversion: Some Gemini 3.1 Pro outputs reported DSCR as Annual Debt Service divided by NOI, the inverted ratio, which produces values below 1.0 that look like distress signals on healthy deals.
- NOI inclusions: All three models occasionally subtracted capital expenditures from NOI. NOI is gross revenue minus operating expenses and does not include CapEx, debt service, depreciation, or income taxes.
Document quality drives extraction error rates
When financial documents were provided as clean text or structured CSV, extraction error rates stayed under 2 percent for all three models. When the same documents were provided as scanned PDFs without OCR pre-processing, error rates climbed to between 4 and 9 percent. Claude Opus 4.7's enhanced vision (3.75 megapixels of image resolution as of the April 2026 release) closed some of this gap but did not eliminate it. For high-stakes underwriting, the cleanest pipeline is OCR first with a dedicated tool, then feed clean text to the LLM.
Pro forma stress testing is brittle without explicit structure
When asked to sensitize an exit cap from 5.5 percent to 6.5 percent in 25 basis point increments, models occasionally extrapolated 50 basis point or 100 basis point steps instead. Forcing a JSON-structured output with explicit step counts cut this error type by more than half.
How to Reduce Hallucination in Practice
Five workflow changes consistently lowered hallucination rates by 50 percent or more in this study:
- Force citation per number. Tell the model, "Cite the page and line item for every dollar value before computing a derived metric." This single instruction was the most effective intervention across all three models.
- Provide formulas explicitly. Do not ask "what is the DSCR?" Ask, "Compute DSCR as NOI divided by Annual Debt Service. Then verify the result is between 1.0 and 2.0."
- Use JSON output for derived metrics. Structured output forces the model to commit to one input set and one formula.
- OCR before LLM. Pre-process scanned PDFs through Adobe Acrobat OCR or a dedicated OCR service before passing them to any model.
- Independent re-derivation. Have the model regenerate the same metric in a second prompt with no access to the first answer, then compare.
For CRE investors looking for hands-on implementation support on AI verification workflows, The AI Consulting Network specializes in exactly this. To learn the full verification protocol from this study, see our companion guide on how to test AI property valuation accuracy.
What This Means for CRE Workflows
The 1.8 to 3.1 percent hallucination range across frontier models is low enough for AI-assisted underwriting to be a real productivity gain, and high enough that every output on a real deal needs human verification on the core metrics. Industry research from CBRE and the 2026 First American Data and Analytics CRE technology survey reinforces this finding: 66 percent daily use, 5 percent decision-grade trust. The gap is a verification problem, not a model problem.
Firms that are pulling ahead in 2026 are doing two things: standardizing prompts so the same workflow returns the same shape of answer every time, and adding a human verification step at the end of every AI-generated underwriting output. For an enterprise-grade rollout plan, connect with The AI Consulting Network for tailored guidance.
Frequently Asked Questions
Q: Which AI model has the lowest hallucination rate on CRE financial data in 2026?
A: Claude Opus 4.7 produced the lowest aggregate hallucination rate at roughly 1.8 percent across 500 underwriting prompts in this study, followed by GPT-5.5 at 2.4 percent and Gemini 3.1 Pro at 3.1 percent. All three are usable for AI-assisted underwriting, but none should replace human verification on the core metrics.
Q: What is the most common type of AI financial hallucination?
A: Confusion between related metrics, especially cap rate versus Cash-on-Cash, DSCR inversion, and including capital expenditures inside NOI. These three error patterns account for roughly one-third of all financial errors across frontier models.
Q: How can I reduce AI hallucination when underwriting?
A: The most effective single intervention is forcing the model to cite the page and line item for every dollar value before computing a derived metric. Also provide formulas explicitly, request JSON-structured output, OCR scanned PDFs first, and re-derive the core metrics in a second independent prompt.
Q: Should I trust AI for investment committee underwriting?
A: AI can do the first 80 percent of an underwriting pass faster than a human analyst, but the final 20 percent (verification of NOI, cap rate, DSCR, Cash-on-Cash, and IRR) should remain a human responsibility on any deal above roughly $5 million. The 2026 First American Data and Analytics survey shows the industry has converged on this same conclusion.
Q: Does Claude Opus 4.7 still hallucinate?
A: Yes. Even the best frontier model in this study hallucinated on approximately 1.8 percent of prompts. The newer 3x image resolution and 1 million token context window improve performance on document-heavy tasks, but no current model is hallucination-free on financial data.