- contact@insightcreditai.com
Eight agents: seven work each request in order, one watches the loan book. They read, spread, reconcile, value, grade and draft. They don’t clear checks, record mitigants, write to borrowers or decide.
Seven agents work a credit request in order; the eighth watches the loan book. Every agent writes what it found and why. When a person needs to act, it prepares the decision and waits.
| Agent | Step | What it does and prepares | What it never does |
|---|---|---|---|
| Intake agent | Collect and spread | Reads every document into fields with their lines; prepares a request to the borrower for what is missing | Invent a figure, or send the request |
| Spreading agent | Collect and spread | Builds the spread and reconciles it with the returns, agings, bank statement and rent roll; prepares a request to explain the differences | Adjust the statements, or clear a check |
| Collateral agent | Analyze | Values the collateral from the appraisals and the aging, computes LTV and coverage; prepares an appraisal order when one is stale | Supply a value, or order the appraisal |
| Credit analyst agent | Analyze | Computes the scorecard grade and explains what drives it; prepares a request to explain a decline | Pick a grade or override one |
| Policy agent | Recommend | Runs the credit policy controls; drafts a mitigant for each referred exception from the figures | Record a mitigant, or justify a block |
| Memo writer | Recommend | Drafts the credit memo from the engines’ figures | Add a number or a conclusion that isn’t in them |
| Supervisor | Recommend | Checks what blocks a decision and the authority the request needs; prepares the submission | Submit, approve or decline |
| Covenant monitor | Monitor the book | Tests every covenant on the quarterly reporting; prepares watchlist additions and opening the overdue annual reviews | Put a borrower on the watchlist, or take one off |
Every action an InsightCredit agent can take is hard-capped at L2, and the tenant ceiling (CREDIT_AGENTS_MAX_LEVEL) can only lower that. An administrator sets each agent’s level per action within the cap; agreement and override rates are tracked for every action, so the evidence for a change is on the screen.
| Level | Name | What the agent may do |
|---|---|---|
| L1 | Draft | Prepares a draft; a person does the work. Where every action starts in the demo |
| L2 | Ask first | Prepares the decision and asks a person before it acts. The ceiling for every InsightCredit action |
| L3 | Act and tell | Acts, tells the person, and lets them pull it back. Locked |
| L4 | Act | Acts on its own. Locked |


Reading, spreading, ratios, reconciliations, grades, valuations, controls and covenant tests are deterministic code. Where a language model is used, it reasons over numbered facts and phrases what the agent found; a sentence that carries a number or a name not in the facts is thrown away and the engine’s own wording stands.
DSCR, leverage, global cash flow, LTV, coverage, scorecard points and covenant tests are calculated, not generated.
Every line an agent writes refers to a document line, a spread figure, a check or a control.
With the model switched off, every agent runs on its rules and says so on each step.
What an agent prepares lands on the case and on the analyst’s or officer’s home. Accepting carries it out (a request is drafted and held, a mitigant recorded, the memo saved, the request submitted); rejecting records that a person disagreed; asking sends a question back. Each answer is recorded with who and when.


The copilot answers questions about one credit request with read-only tools, citing what each answer rests on. Every figure and name in an answer is checked against what the tools returned. It suggests; you decide.

