Method / Input acceptance
AI can help locate and transcribe evidence, flag inconsistent inputs and prepare a model mapping. Each field still needs a defined source, unit, period and acceptance owner. A number can be accurately copied from a document while being inappropriate for the model cell it fills. Use the input audit to identify the next action before the model obscures that distinction.
Use the fictional worked example and editable worksheet to apply this method to your own documents.
The useful distinctions
Keep the document, page, reporting period and original wording with the extracted value. The published IC-memo workflow describes analyst review of structured deal evidence; its relevance is the review step, not automatic acceptance of a property’s financials.
Source 1A calculated value should retain its formula and accepted inputs. An assumption should retain its owner and reason. The InnvestFL case supports consistency between a model and investor materials; it does not make a shared assumption an observed outcome.
Source 1Apply an explicit check at the point an AI output enters the team’s working file. NIST’s voluntary AI risk framework is useful governance context; the field-level acceptance method here is our practical application for CRE document workflows.
Source 1Observation to next action
| Evidence in the file | What it describes | Next review |
|---|---|---|
| Extracted contractual figure | What a named document states | Confirm the correct party, unit, period, amendments and target field. |
| Calculated figure | A formula applied to selected inputs | Review the formula, units, missing values and input acceptance. |
| Assumed scenario | A deliberate analytical choice | Name the owner, rationale and sensitivity; keep it labelled as an assumption. |
Work through the file
Use this sequence with the documents available for your file. Keep a reference with each observation and turn unresolved items into explicit requests.
For each model field, name the expected unit, period and acceptable evidence. Separate in-place operations from a proposed future scenario. In the fictional rent exercise, the destination expects annual collected rent while the source supplies monthly scheduled charges. Record that mismatch before calculating anything. A multiplication by twelve would change the time unit without converting scheduled rent into cash received.
Store the source value as written and the exact locator before normalization. Retain document edition, page or table, relevant parties and the period covered. Preserve ambiguous signs, units and date ranges for review rather than silently repairing them. If the fictional extraction has no page reference, request the original and leave the field unaccepted until a reviewer can reproduce the reading.
Write conversion, aggregation and mapping rules beside the destination field. Keep the inputs and their evidence statuses attached to the formula. The fictional versions use different occupied-unit populations, so their disagreement cannot be resolved by choosing the higher total. Ask the model reviewer to reconcile the population and period first. If an annualized estimate is later used, preserve its assumptions and estimated status.
Give each conflict to an owner and record whether the field is accepted, held or replaced by a separately labelled scenario. Retain the accepted version, reviewer and reason alongside the original source. A missing field stays open; it does not become zero or a plausible value because a completed model looks better. Recheck dependent calculations when an accepted input changes, rather than updating only the visible summary.
Before using the output
A parcel record alone does not provide the operating revenue and expenses required for that calculation. Use the relevant financial documents and review their period and definitions.
Ask AI to return original wording, document locator, unit, period and proposed destination for each value, then flag incompatible definitions or populations. A document reviewer confirms the extraction and the underwriting lead accepts its use. A model confidence score does not replace either check. The worksheet organizes their requests; it does not run extraction or accept inputs automatically.
No. It exports the evidence and acceptance tasks that should precede use of inputs in your own reviewed model.
From the page to your next task
Build an editable model input and transformation acceptance ledger, retain document references and open decisions, then export your team's working file.
Scope: One review file using the method in the accompanying guide. The example is fictional.
Edition 2026-09-30.1
Put the reviewed inputs to work
Once inputs have sources, units, periods and reviewer acceptance, use this instruction pack to draft extraction, cash-flow and return handoffs. Its Excel files are illustrative; reconcile the inputs, approve assumptions and independently check formulas before relying on an output.
Explore the existing resource ↗Evidence and scope
NextAutomation · Checked 2026-09-29
Florida; hospitality and mixed-use development workflow. Asset context: hospitality, mixed-use development.
NextAutomation · Checked 2026-09-29
Unspecified CRE investment team; no geographic proof. Asset context: commercial real estate, unspecified asset mix.
NIST · Checked 2026-09-30
Voluntary guidance for managing AI risks. We apply the general need to evaluate and govern outputs to document-supported CRE workflows; this is our editorial application.
