Where Should a Private Credit Firm Start With AI?

1. The Short Answer

A private credit firm should start with AI on three workflows its credit team repeats every quarter: borrower package intake and spreading, covenant tests against the credit agreement, and the first draft of the credit memo. An analyst checks each one before anyone relies on it.

Portfolio monitoring comes next, once those three run well on the firm's own documents.

For a lean direct lender, AI for private credit starts with the borrower package: the financial statements and compliance certificate each borrower sends every quarter. The tool can be an AI assistant such as Claude or ChatGPT on a business plan, the version each vendor sells to companies.

These three pass the test in PwC's Global Private Credit Survey 2026, of more than 120 credit portfolio managers: "If a task is document grounded and verifiable, AI is usually ready now." Each starts from your own documents and can be checked page by page.

2. The First Three Workflows

In each one, AI reads and drafts, and an analyst checks the result against the source. Anthropic's credit underwriting example, written for commercial bankers, puts it this way: "Claude pulls spreads, runs ratios, and structures the writeup while you own the credit decision."

Take them in this order: the covenant test runs on the spread, and the memo draws on both.

Proskauer's 2026 private credit survey shows why the covenant test belongs on the list: "46% of U.S. lenders say they will not do deals without covenants, up from 35% last year."

3. Then Monitoring

Portfolio monitoring means watching the whole book between covenant tests: what changed at each borrower and which names belong on the watchlist. It comes fourth because it runs on what the first three produce.

PwC's survey points the same way: 54% of respondents "are most likely to use AI in the underwriting process," and "only 16% of respondents view AI-enabled portfolio management (for example, monitoring) as a current priority." Both figures measure plans and priorities. Neither counts current use.

A watchlist built on unchecked spreads repeats their errors across the whole book.

4. What a 10 to 30 Person Lender Does in 90 Days

At a lender of 10 to 30 people, plan the first 90 days around one quarter of borrower packages:

1. Weeks 1 and 2: rules and a test set. Name one owner on the credit team, with time set aside each week. Write a one-page data rule on which borrower files may go into AI. Buy a business plan, never personal accounts, and pick three closed deals the team knows well.
2. Weeks 3 to 6: spreads, then covenant tests. Run the test deals through the tool and compare its spreads with your analysts' own. Save the instructions that work, then build each deal's covenant list and recheck one compliance certificate.
3. Weeks 7 to 10: memo drafts and training. Draft a memo per test deal in your house template, then train the rest of the credit team on redacted copies of the same work.
4. Weeks 11 to 13: a live quarter. Run the next quarter's borrower packages through all three workflows, analyst checks included. Then decide on monitoring and on any specialist software.

No survey publishes hours per workflow for direct lenders, so time one analyst's spread, covenant test and memo draft in week 1 and again in week 12. For scale, J.P. Morgan says its own machine learning cuts data extraction "which would typically take around 45 minutes for one company, to just 30 seconds."

Whether that owner later becomes a full-time hire is the subject of Does a Private Credit Fund Need a Head of AI?

The goal is what PwC sees at firms already using AI: "reviewing more deals with the same headcount."

5. Tools First, Software Later

The first AI tools a direct lender needs are general assistants such as Claude or ChatGPT, working on the documents the team already has.

Specialist software earns its cost when a job needs a system of record, one place that many people and the auditors trust:

• Covenant tracking across the whole book, with test history and alerts. Claude can build the covenant list and recheck a compliance certificate, but keeps no test history. Lumonic, Allvue, Chronograph and 73 Strings sell covenant monitoring.
• Portfolio monitoring the whole firm reports from, such as S&P Global's iLEVEL, "your private markets software for portfolio monitoring," in S&P's words.

Loan administration always sits in a system of record, such as S&P Global's WSO or Hypercore, or with an outside administrator, and AI reads its exports. None of these vendors publishes a price.

Test any vendor on your own credit agreements before you sign, with the scorecard in How Should a Private Credit Firm Choose AI Software?

6. What It Costs

Licenses are the small line. Claude Team and ChatGPT Business each cost $25 a standard seat a month, or $20 billed annually, and by default neither Anthropic nor OpenAI trains its models on business-plan data.

Assume a credit team of 12 on Claude Team, billed annually: 12 x $20 x 12 = $2,880 a year.

The other cost is time: the owner's hours each week, and the analysts' hours in training and checking.

7. Where to Start

1. List the borrower packages due next quarter and who spreads them.
2. Name one owner on the credit team, with time set aside each week.
3. Write the data rule and buy a business plan.
4. Run three closed deals through the spreads, covenant tests and memo drafts.
5. Put next quarter's live packages through all three, then decide on monitoring.

The Claude version, step by step, is in How to Implement Claude at a Private Credit Firm.