AI portfolio monitoring for private lenders: what the agents do, and what a person still does.

AI portfolio monitoring is software in which AI agents run the recurring steps of a lender's monitoring cycle. The agents collect borrower reporting, spread the financials, map them into the monitoring model, test covenants, check compliance certificates and draft the monitoring note. A person reviews each output and approves it, and every figure is traced to the page it came from.

This guide answers for private lenders: private credit funds, non-bank and SME lenders, venture debt, property and real estate debt funds, and asset-backed and direct lenders. It is written for the credit or operations team that runs the cycle. The question is a practical one. Which steps does the AI do, which stay with the analyst, and what should a lender check before it buys any of it?

The monitoring cycle has nine steps, and most of them are routine.

Every lender runs some version of the same cycle for each loan, monthly or quarterly. The borrower's pack arrives, or does not. Someone extracts the financials and types them into the model. The ratios recalculate. Someone tests the covenants, reads the compliance certificate, compares this period with the last, writes a note and sends it for sign-off.

Laid out as steps, the cycle is: reporting collection, financial spreading, model mapping, covenant testing, a data consistency check, the compliance certificate check, trend analysis, the monitoring note and the portfolio view. Seven of the nine are the same work every period, on a different pack. That is the part the agents take.

What an agent does at each step.

• Reporting collection. Chases each borrower's pack by email and portal against the reporting calendar, logs what has arrived and lists what is late before the cycle starts.
• Financial spreading. Reads the financials in the shape they arrive, PDF, Excel or an accounting export, and spreads them into the lender's standard fields. Each figure is linked to the page it came from.
• Model mapping. Maps the spread into the lender's own Excel monitoring model line by line, recalculates every ratio and flags any cell that moved for a reason it cannot trace.
• Covenant testing. Tests every covenant against the register, reports the headroom or the breach and sizes the cure where one is needed.
• Data consistency. Compares this period's financials with every prior period and flags restated figures, and line items or categories that were not there before.
• Compliance certificate. Checks the borrower's certificate against the facility agreement and the lender's own calculation, and flags where the three disagree.
• Trend analysis and the note. Writes up what moved and why across financials, covenants and correspondence, then drafts the monitoring note in the lender's template with each statement cited.

The portfolio view is the ninth step and is not work at all. It is every loan's cycle on one screen: what was received, what was tested, what is late and what to watch, each figure linked back to the run that produced it.

What stays with a person.

The agents produce evidence. They do not approve anything. A named analyst reviews each output and signs it off before it reaches a committee, a borrower or an investor, and the approval is recorded with who signed and when.

That line is not a design preference. For Australian private credit managers inside the FSC, clause 6.2.4 of FSC Standard No. 30 says personnel who originate, acquire, dispose of or manage an asset should not have sole authority to approve its valuation. An automated cycle can produce the valuation evidence. It cannot be the approval. The same logic applies to a covenant waiver, a watchlist decision or a restructure. The agent sizes the cure. The credit team decides whether to grant it.

The judgement work also stays. Why leverage drifted. Whether a borrower's explanation holds. What the trend across three periods means for the loan. The agents put the figures, the sources and the history in front of the analyst. The analyst does the thing the lender is paid for.

Six checks before you buy.

Vendors describe these products in similar words. These are the questions that separate them.

• Does every figure trace to a page? Ask to see a spread figure clicked back to the borrower's pack. A total with no source is a number the analyst has to rebuild.
• Does it test covenants from the facility agreement or from the borrower's certificate? A tool that checks the certificate's arithmetic is a tracker. A tool that recalculates from the definitions in the agreement finds the case where the two disagree.
• Does it run where the team already works? A monitoring model in Excel, a reporting inbox in Outlook, an AI tool such as Claude or Copilot. A new system of record is a migration project, and migrations are where monitoring tools stall.
• Is training switched off, and where is the data hosted? Get both in writing. For an Australian lender, hosting location and the AI providers used should be agreed in scoping and documented in the engagement agreement.
• Is there a record a reviewer could follow two years later? Every run logged, every approval recorded, every change versioned. Clause 4.3.1 of Standard 30 asks for records sufficient to demonstrate governance in operation, and that is what an audit trail is for.
• Has it been tested against your own past cycles? Before go-live, the agents' output for prior periods should be compared with the monitoring work the team actually did. Where they differ, someone should be able to say which was right.

What it changes, in figures that have been published.

Public figures are rare in this category because most vendors do not publish them. In Levercon's own published case study, a $500 million corporate private credit manager cut monitoring from a full day per loan to about two hours. The team assessed four times as much data per cycle and moved from four loans per analyst to twelve. Those are one lender's numbers, and they are disclosed here because this guide is written by the vendor that produced them.

The pattern behind the numbers is the one to look for. The time came out of spreading, mapping and testing. It went back into origination and into the analysis that the cycle had been squeezing out.