Where Models Fit in Research Workflows

Placement before scale

Teams often introduce models by asking what the tool can generate. InvestCAD asks where the tool may sit without collapsing the difference between draft and record. Placement is a workflow decision: assisted steps remain reversible; human veto remains mandatory before circulation.

Fit zones

• Ingest and labeling — propose source types, detect duplicates, suggest capture metadata.
• Extraction — pull tables, quotes, and checklist candidates with citations.
• Draft comparison — highlight divergences across document versions.
• Out of scope — recommendations, forecasts, invented metrics, attractiveness rankings.

Editorial gates

A gate is a named review step with a reject rule. Example: no figure enters the archive without a source fragment or an explicit “absent” mark. Another: no evaluative adjective survives without an anchor. Gates are boring; they are also what keeps FinTech research tooling from becoming an advice engine by accident.

If a model output cannot be rejected easily, it was placed too far downstream.

Logging assisted drafts

Audit-friendly workflows keep assisted drafts distinguishable from approved text. That may be as simple as a separate file, a status field, or a watermark in the working copy. The goal is not theater; it is the ability to reconstruct how a claim arrived.

1. Define allowed model tasks in writing.
2. Ban prompt patterns that request potential, returns, or “what to do.”
3. Require human confirmation before any client-facing or public circulation.
4. Archive rejected suggestions that invented numbers — they train the team’s eye.

Closing

Fit is not a feature list. It is a map of where acceleration helps and where fluency becomes risk. Teams that document that map before scaling tools keep research methods intact when volume rises.