Blogs

Why Private Credit AI Stops at Underwriting

Credit managers are putting AI into underwriting first. What portfolio monitoring involves each quarter, why it gets funded last, and three checks a credit team can run before its next valuation.

Written bySurane Jayaraman, Chief of Staff, Needl.ai

Nov 18, 2026 · 15 min read

Why Private Credit AI Stops at Underwriting

Private credit AI is concentrated in underwriting because that is where the deadline, the deal team and a countable time saving sit. In PwC's Global Private Credit Survey 2026, 54% of credit portfolio managers said underwriting is where they are most likely to use AI, and 16% called AI-enabled portfolio management, such as monitoring, a current priority.

Picture a direct lending team on day 45 of the quarter, the date many credit agreements set for quarterly financials. Packages are arriving from more than a hundred borrowers: management accounts, a compliance certificate, a board deck and, from one or two, a waiver request. Each package has to be read against its credit agreement and every amendment before the quarterly marks are set. The memo that won each of those deals was written once. This reading is repeated every quarter for the life of the loan.

This guide looks at where AI is going in private credit, why monitoring gets funded last, what a quarter of monitoring involves, and three checks a credit team can run before its next valuation.

Where are credit managers putting AI?

Both of 2026's main surveys of private credit lenders find AI moving into underwriting faster than into monitoring.

PwC surveyed more than 120 credit portfolio managers in the US, the UK, Europe, the Middle East, Asia and Australia between January and March 2026. Underwriting came first as the place they are most likely to use AI, at 54%. Only 16% called AI-enabled portfolio management, with monitoring as the example, a current priority. The figures come from separate questions in the survey. Together they show where attention is going.

Armentum Partners and Lumonic polled about 150 private credit lenders for their 2026 Private Credit Technology Summit. Use of third-party underwriting tools rose to about 20% from about 10% a year earlier. Use of technology for loan monitoring held at about 50%. In the presenters' words, the growth is "happening at the AI and analytics layer on top."

Where credit managers are putting AI

Two 2026 surveys of private credit lenders. The figures come from separate questions.

SurveyUnderwritingPortfolio monitoring
PwC Global Private Credit Survey 2026, 120+ credit portfolio managers54% most likely to use AI here16% call AI-enabled portfolio management a current priority
Armentum Partners and Lumonic, about 150 lenders, March 2026Third-party tool use rose from about 10% to about 20% in a yearTechnology use held at about 50%

Sources: PwC; Armentum Partners and Lumonic.

Figure 1. Two surveys, one direction: AI use rising at underwriting and little change in monitoring.

The flat monitoring figure is the more useful of the two. Half of lenders already use some technology for loan monitoring, so the open question is what AI adds to the tools, spreadsheets and portals a portfolio team already uses.

Practitioners also doubt what is on offer today. In S&P Global Market Intelligence's 2026 outlook survey, 75% of respondents rated AI ineffective for portfolio monitoring. They cited lack of expertise (49%), data privacy (43%) and model accuracy (38%) as barriers. That survey covers private equity and venture investors, so treat it as a proxy for credit teams.

PwC's own view is more specific. It describes extraction, summarising and monitoring as "relatively mature because the output can be sourced and checked by humans." It also notes that public evidence on early warning systems built for private credit is still thin.

Why does underwriting get the budget first?

Underwriting has an end date and a deliverable, so the case for AI there is easy to make.

A deal team works to an IC meeting. The output is a memo, and the hours saved on it can be counted. John Markell of Armentum Partners told Creditflux in May 2026 that with AI tools the credit memo "can be done in hours versus a week."

Monitoring has none of that shape. It runs for the life of the loan and is spread across compliance certificates, covenant tests, amendments, borrower financials and news. At many firms the work is also split between the deal team that knows the credit, a portfolio monitoring group, the valuation team and fund finance.

The benefit is harder to show as well. A faster IC memo fits in a business case. An early warning on a borrower that has not slipped yet shows up only as a loss that did not happen.

Data is the last reason. Underwriting runs on one data room at one point in time. Monitoring runs on a stream of documents from every borrower, in formats that change from borrower to borrower and sometimes from quarter to quarter. That is a harder problem to buy software for, and the barriers in the S&P survey above are the ones a monitoring tool has to clear first.

What does a quarter of monitoring involve?

For a sponsor-backed direct lending book, each quarter means reading every borrower's package against its own credit agreement, then carrying the result into ratings, marks and investor reports.

One quarter of monitoring

Six stages for a sponsor-backed direct lending book, repeated for every borrower.

StageWhat arrives or happensWhat the team checks
Reporting packageManagement accounts, the compliance certificate, the board deck and budget, plus a borrowing base certificate on asset-based facilitiesWhether it is complete and on time, and whether the format or the auditor has changed
Covenant testThe borrower's own calculation of leverage, coverage and liquidity, signed by an officerThe ratios recomputed on the agreement's EBITDA definition, with its caps, pro forma rules and amendments
Amendments and waiversRequests, executed amendments, repricingsWhat each one changes in the next test
PIK and payment termsPIK elections, interest deferrals, extension requestsWhether PIK was in the original deal or added later
Ratings and marksWatchlist changes, internal ratings, Level 3 fair value under ASC 820, the valuation committee and often a third-party valuation providerThe support behind each mark
Investor reportingQuarterly reports to LPsConsistency with the marks and with ILPA's templates

Illustrative quarter for a sponsor-backed direct lending book.

Figure 2. One quarter, six stages, repeated for every borrower. Packages arrive by email, lender portals, agent sites and shared folders.

The definition decides the number

Covenants are tested on the credit agreement's definition of EBITDA, with its permitted add-backs and caps, as amended. The add-backs are large. Lincoln International reported that EBITDA adjustments were nearly 25% of adjusted EBITDA across the companies it values in Q3 2026. A team that only copies the certificate's figures takes the borrower's reading of the definition.

Amendments change next quarter's test

Every amendment and waiver changes what is tested next. Lincoln reported amendment activity up 8% quarter on quarter in Q1 2026, even as its covenant default rate fell.

PIK added after close is a signal

Lincoln found PIK interest in 10% of the loans it valued in Q3 2026. Of those, 54% were what it calls bad PIK, introduced after a deal closed without it. That is 6% of all loans. On those investments, loan-to-value has moved from about 45% at close to about 75% today, and Lincoln's view is that this PIK use "suggests there would have otherwise been more defaults."

Investors want the same numbers in a standard form

ILPA's updated Reporting Template applies from the Q1 2026 reporting period for funds still in their investment period. Its new Performance Template applies to funds that started operating on or after January 1, 2026.

Why do the numbers point to monitoring?

The book is growing, and most of today's credit stress first appears as a request or an amendment in a borrower's documents.

More loans, the same team

PwC's base case has private credit growing from more than $2 trillion in assets to $3.4 trillion by 2030. Every loan written today adds years of quarterly packages, covenant tests and marks. There is no reliable public benchmark for how many borrowers a portfolio analyst covers, so we do not quote one. Few portfolio teams expect their headcount to grow at the pace of the book.

Stress arrives as a change in terms

Fitch's U.S. Private Credit Default Rate reached a record 6.3% for the 12 months to Q3 2026, as reported by MonitorDaily. Interest deferrals and PIK introductions made up 50% of the default events, and maturity extensions under stress made up 38%. Uncured payment defaults were 8%. Most of the events Fitch counted were changes to terms, which a lender sees first in a request, an amendment or a certificate.

Other measures are lower because they count differently.

Three default rates for the same quarter

Q3 2026. Each measure counts different events.

MeasureQ3 2026What it covers
Fitch U.S. Private Credit Default Rate6.3%, trailing 12 monthsIncludes PIK introductions, interest deferrals and extensions under stress
Proskauer Private Credit Default Index2.41%746 senior secured and unitranche loans, on Proskauer's definition of default
Lincoln International covenant default rate2.5%, size-weightedCovenant defaults across the private companies Lincoln values

Sources: Fitch Ratings via MonitorDaily; Proskauer; Lincoln International.

Figure 3. Three default rates for the same quarter. Which events count as stress depends on the definition, and the same is true of a team's own watchlist.

Stress has moved between sectors

Lincoln identified elevated stress in parts of software, including marketing technology and human capital management, and in healthcare staffing and behavioral health. It found 9% of first lien and unitranche debt valued below 90% of par in Q3 2026. Fitch, by contrast, recorded the lowest default rate among its large sectors in technology software, at 0.6%. Teams need to read each name on its own documents.

BDC disclosures show the same pressure

PitchBook LCD put PIK at 8.2% of total interest income for the largest listed BDCs in Q1 2026, its lowest level since Q4 2023. Octus counted debt on non-accrual at 1.9% of BDC debt investments at cost in Q2 2026, down from the prior quarter but 58% higher in dollar terms than a year earlier. At the 10 largest non-traded BDCs, redemption requests averaged 13% of NAV in Q1 2026 and 14% in Q2, against the 5% level at which managers can limit withdrawals, according to With Intelligence.

Regulators are asking about marks

The Financial Stability Board's May 2026 report warned that "valuation opacity and reliance on private credit ratings can amplify strains in stress." The SEC Division of Examinations' fiscal 2027 priorities, published October 23, 2026, say examiners will assess "the reasonableness of their valuation processes and methodologies" at private fund advisers. They will also review whether advisers' controls cover risk at the level of both the fund and each investment.

Three checks to run this quarter

Each check takes an afternoon, uses your own files and shows where the monitoring work actually goes.

How long until a compliance certificate is checked against the credit agreement?

Measure the days from receipt to a completed recomputation under the agreement's definitions, with the latest amendment applied. The certificate is the borrower's calculation. The check is whether its EBITDA ties to the management accounts, whether each add-back fits the agreement's categories and caps, and whether a step-down is coming.

A hypothetical example shows why the gap matters. A borrower certifies Consolidated EBITDA of $20.0 million, including $4.0 million of cost-savings add-backs. The agreement caps those add-backs at 15% of EBITDA before the adjustment, which is $2.4 million. Net debt is $100 million and the maximum leverage is 5.50x.

One certificate, two readings

Hypothetical borrower. Net debt $100M, maximum net leverage 5.50x.

ItemAs certifiedRecomputed under the agreement
Cost-savings add-backs$4.0M$2.4M (capped at 15% of $16.0M)
Covenant EBITDA$20.0M$18.4M
Net leverage5.00x5.43x
Room before the 5.50x test failsEBITDA can fall about 9%EBITDA can fall about 1%

Hypothetical example for illustration.

Figure 4. Hypothetical borrower. The certificate passes either way, and the watchlist conversation is very different.

How many watchlist names did your team find first?

Tag each watchlist addition with its trigger and its source. Borrower-initiated triggers include waiver requests, amendment requests, PIK elections and requests to extend a maturity. Given Fitch's breakdown, these are often the default events themselves.

Team-initiated triggers include headroom shrinking across two quarters, add-backs growing faster than reported EBITDA, a late package or a request to move the reporting deadline, revolver draws, an auditor change, and customer or litigation news. If most additions come from the borrower, the team is learning on the borrower's schedule.

How long would it take to show an examiner how a mark was supported?

Pick three Level 3 positions at random and time it. A complete answer links the mark to the reporting package used, the covenant calculation, any amendment in force, the internal rating, the valuation memo and the third-party provider's report where there is one. Given the SEC's stated focus on valuation processes, this test is worth running before an examiner runs it for you.

What to look for in AI for monitoring

A monitoring tool earns its place when a reviewer can check its work quickly, on the firm's real documents, inside the firm's own controls.

Six things to test in a pilot

Apply them to any monitoring tool, on one quarter's real packages.

TestWhat good looks like
Source traceabilityEach figure opens the passage or the spreadsheet cell it came from, so a reviewer can check it in seconds. An answer that no document supports is marked as such.
The agreement's logicCovenants are recomputed on the agreement's own definitions, including amendments, with the bridge from reported to covenant EBITDA shown. A tool that only extracts the certificate's numbers repeats the borrower's view.
Real borrower documentsTest on actual packages: inconsistent PDFs, multi-tab workbooks, a conformed agreement beside separate amendments, and packages that arrive late or incomplete.
Changes since last quarterNew and edited files are flagged, and the previous review stays readable beside the new one.
The firm's own environmentBorrower data is confidential and often material non-public information. Ask where the tool runs, whether it trains on your data, and whether answers stay within what each person is already allowed to see.
Human reviewThe analyst owns the spread and the rating, and the committee owns the decision. The tool should make that review faster and leave a record of what was asked and answered.

Applies to any tool.

Figure 5. Six things to test in a pilot, with any tool.

Where AI stops

AI will not know what the sponsor said on last week's call. It cannot fix a package that was never sent, and it has less to test on a covenant-lite loan. Models also make mistakes, which is why every figure needs a source a person can open. The value is in getting the reading and the first covenant test done early enough for judgment to be applied while there is still time to act.

Where Needl.ai fits

Needl.ai pairs portfolio monitoring with data room diligence. The two run side by side, because the credit agreement a deal team reads before close is the same one the portfolio team checks every quarter after it.

Before close

In AI Due Diligence, Needl.ai reads the data room and checks the projections against the underlying model and source data. On credit deals it sets out a covenant analysis for the lender from the credit agreement: the covenant register, headroom, claim ranking across the capital structure, the EBITDA bridge, and amendments and waivers. A first-pass assessment comes back around 30 minutes after the data room is loaded, and a cited IC memo follows in the firm's own structure.

After close

In Private Credit Portfolio Monitoring, each borrower is its own room, synced from Google Drive or OneDrive or uploaded directly.

  • The package, read. Management accounts, compliance certificates and board decks are read in full, and workbooks tab by tab, with the sheet and range cited for spreadsheet figures.
  • The covenants, tested. A covenant register lists each covenant with its level and test frequency. The ratios are recomputed on the agreement's own definition of EBITDA, with the headroom on each test and how far earnings can fall before one fails.
  • The certificate, checked. The certificate's own calculations are set against the figures in the accounts, so a difference shows at once.
  • The history, kept. Amendments, waivers and repricings sit in one history, and new or edited files are flagged after each review.
  • The signals, graded. Trackers on the borrower, its sponsor and its sector grade news, litigation and filings high, medium or low, with the reason.
  • The review, drafted. A borrower review is drafted in your memo structure, and citations open the passage with the quote highlighted.

Needl.ai runs in your own AWS or Azure account, as a dedicated single-tenant instance, or as Needl.ai SaaS. Every search respects the permissions each person already holds on the source documents, and Needl.ai does not train models on customer data. See Trust & Security for certifications.

Needl.ai gives no view on whether to lend. It frames the position, and the committee decides.

FAQs

Why do private credit firms use AI in underwriting before monitoring?

Underwriting has an IC date, a deal team and a memo whose time saving can be counted, so the business case is simple. Monitoring runs for the life of each loan across many documents and teams, and its payoff is a loss avoided. In PwC's 2026 survey, 54% of credit portfolio managers said underwriting is where they are most likely to use AI. Needl.ai covers both stages.

What does AI do in private credit portfolio monitoring?

It reads each borrower's reporting package, spreads the figures with their sources, recomputes covenants on the credit agreement's definitions, tracks amendments and waivers, and flags news and filings on the borrower and its sponsor. On Needl.ai, each borrower is its own room and the review is drafted in the firm's memo structure. See Private Credit Portfolio Monitoring.

How can AI check a compliance certificate?

By recomputing the borrower's ratios from the management accounts under the credit agreement's own EBITDA definition, with its add-back caps, pro forma rules and amendments, and setting the result beside the certified figures. A difference in Covenant EBITDA can turn comfortable headroom into a near breach. Needl.ai shows both figures side by side, each with its source, so the analyst can decide.

Can AI decide whether to lend or how to mark a loan?

No. The analyst owns the spread and the internal rating, the valuation committee owns the mark, and the investment committee owns the decision to lend. AI does the reading and the first covenant test, and it should show the source behind every figure so a person can check it. Needl.ai gives no view on whether to lend.

Can Needl.ai run inside our own cloud account?

Yes. Needl.ai runs in your own AWS or Azure account, as a dedicated single-tenant instance, or as SaaS. Every search respects the permissions each person already holds on the source documents, and it does not train models on customer data. See Trust & Security for the certifications.

Private CreditPortfolio MonitoringCovenantsDirect LendingAI

Keep reading

Related on Needl.ai: Private Credit Portfolio Monitoring

See it on your own documents

Bring the filings, reports or data room your team works with, and see Needl.ai answer with a source on every claim.

Book a demo