AI Disruption Risk in Software Loans: A Review Framework for Private Credit Lenders

1. The Short Answer: Large Exposure, No Arrears Yet

Business development companies (BDCs) have lent around $115 billion to software firms, about a fifth of their lending, according to a BIS Bulletin of July 14, 2026. Almost none of it is behind on payments yet. Review each borrower for what an AI agent could replace, how it prices, what data it owns and how its loan is built.

The same bulletin calls software revenues "highly vulnerable to AI." It also found that "Less than 1% of software loans are behind on payments" (BIS Bulletin No 128), and KBRA has called widespread AI-driven defaults unlikely.

So the risk sits ahead of you. Covenant tests look backward, and the BIS found that "neither BDCs nor their equity investors have priced software exposure differently." By the time a borrower trips a test, its options and yours have narrowed.

The aim of a review is to know which borrowers are exposed before the marks move.

2. The Numbers: How Much Software Debt Lenders Hold

Every major source puts software among the largest exposures in private credit. The shares differ because the BIS, Octus, KBRA and Moody's define software differently and count different pools of loans.

• BIS, July 2026. BDCs' software loans are "about a fifth of all their lending and over 80% of their fast-growing technology portfolios" (BIS Bulletin No 128).
• BIS, March 2026. In private credit direct lending, "Outstanding loans to SaaS firms increased from almost $8 billion in 2015 to over $500 billion, or 19% of total direct loans" by the end of 2025 (BIS Quarterly Review, March 16, 2026).
• Octus, 2025 and 2026. Software at "approximately 29% as of Sept. 30, 2025" across 155 BDC portfolios (Octus), then "approximately 32% at cost and fair value" across 175 at the end of March 2026 (Octus).
• KBRA, February 2026. 17% of borrowers by count and 22% ($224 billion) of total debt exposure in KBRA's own assessment portfolio, where software is the second-largest sector (KBRA).
• Moody's, April 2026. Software is the largest BDC sector exposure, "accounting for roughly 25 per cent of portfolios on a median basis," according to Moody's as reported by Alternative Credit Investor.

So about a fifth and approximately 32% can both be right.

The BIS also found that "about 60% of software lending now goes to firms that borrow from seven or more BDCs, up from under 10% in 2015," so trouble at one borrower lands in many books. The IMF's Global Financial Stability Report of April 14, 2026 flagged "the substantial exposure of direct lending to software borrowers facing potential AI disruption."

However you count it, software is too big to review as one sector line.

3. What Has Shown Up So Far, and What Hasn't

Equity markets moved first. Software stocks "collapsed by almost 30% between October 2025 and February 2026," according to the BIS Quarterly Review, which found that BDCs with high software exposure "have performed around 5 percentage points worse" since October 2025 than those with low exposure.

Credit has moved less. Octus put the average price of BDC software holdings at "97.8% of par as of Sept. 30, 2025," with "Fewer than 6%" of software loans marked below 90% of par. By the end of March 2026, loans priced below 90% of par had "increased to 8.4% of all loans" (Octus). As reported by Forbes, PitchBook LCD data showed "nearly ten percent of top BDC investments tied to distressed borrowers by mid-2026," and software loans and second-lien tranches "accounted for 89% of the year-over-year increase in distressed debt."

The share of BDC loans using payment-in-kind interest rose "from approximately 6 percent to roughly 10 percent by early 2026," the Federal Reserve Bank of Boston found in the filings of 168 BDCs. For software, the BIS warns, PIK "can postpone the point at which strains become visible."

The other side is just as clear. The BIS wrote that "Borrowers' revenue uncertainty posed by generative artificial intelligence has not affected these loans yet," and that modest debt levels and secured lending "may limit spillovers." Few loans need refinancing soon: Octus found "just 9.7% of software loans mature prior to 2028," though "35% of software principal is still set to mature in 2028 and 2029" (Octus).

KBRA flagged 165 of 495 borrowers as higher AI risk, 41 of them with near-term maturities. If all 41 defaulted, its default rate by count would rise to 4.8% from 3.4%, and KBRA called its findings "strong evidence that widespread AI-driven defaults are unlikely." As reported by Forbes, "Fitch expects most rated BDCs to keep meeting asset-coverage requirements even if every software investment were written down by 50%."

Both pictures are true: the loans perform, and the revenue behind some of them is less certain than the marks suggest.

4. Why AI Hits Software Borrowers

The BIS names two channels: AI "can both substitute for existing products and lower development costs, reducing barriers to entry and intensifying competition." The first takes customers. The second invites rivals.

• Substitution. If an AI agent can draft the document, build the report or route the ticket, the tool that did that job has less reason to exist. After Anthropic released plugins for Claude Cowork on January 30, 2026, Reuters reported that they "would automate tasks across legal, sales, marketing and data analysis," a list that reads like a map of software categories. On February 3, Private Equity Wire reported BDC shares sliding on software exposure concerns.
• Cheaper development. A small team can now copy a product's core features and charge less, and the incumbent's defenses, price cuts or extra R&D, squeeze the cash that services the debt.
• Seat compression. Many products are priced per user. Assume, for illustration, a customer whose team of ten shrinks to six as AI takes on part of the work: the vendor loses four seats, keeps the customer, and its logo retention still looks perfect.
• Pricing pressure. Customers expect AI features in the price they already pay, and renewals become negotiations against a cheaper rival's quote.

None of this reaches a covenant test until revenue or EBITDA falls. The review looks for it sooner.

5. Which Software Borrowers Are Most and Least Exposed

Most of the exposure sits where AI bites hardest. "About three quarters of BDC software exposure is in horizontal and workflow categories such as productivity, application and automation tools, where AI substitution is most direct," the BIS found. Horizontal tools serve every industry, and so does a general AI assistant.

Lenders see the same split. In Octus's February analysis, "BDCs cite proprietary data, regulated end-market exposure and deeply embedded software platforms as most resilient," while "Single function, tool-oriented content creation software not tied to proprietary data faces greater risk."

• More exposed. Horizontal productivity, application and automation tools; single-function tools such as content creation software without proprietary data; per-seat products for tasks an AI agent now does; anything a customer can replace within a year.
• Less exposed. Products built on data customers can't get elsewhere; software sold into regulated end markets; platforms so embedded in daily operations that replacing them means moving records and retraining staff.

Category tells you where to look first. The borrower's own numbers decide the rest.

6. The Review Framework, Factor by Factor

Ten factors per borrower: four on how exposed the business is, four on whether the pressure shows and how management responds, and two on what the loan and the sponsor can absorb.

Score each factor high, medium or low, with one line of evidence and its source. Missing evidence becomes a question for management.

7. ARR Loans and Covenant Conversion

Recurring revenue loans are where AI risk and loan structure meet. In Morrison Foerster's definition (November 10, 2021), such a loan is "made available and sized on the basis of the revenue stream of a company and the recurring nature of that revenue stream."

Early covenants test liquidity and recurring revenue. At conversion, which Morrison Foerster puts at "commonly up to 3 years after entry into the loan agreement," those tests fall away and a unitranche-style covenant on debt to EBITDA takes over.

AI pressure hits the assumption the loan was sized on: revenue that keeps recurring. Then conversion asks for EBITDA just when the borrower may need to spend more on R&D to keep up.

KBRA's Q2 2026 recurring revenue loan dashboard (September 25, 2026) found no defaults across 24 rated recurring revenue loan ABS deals and 104 borrowers, with debt to ARR of 1.6x, and reported that "Three borrowers transitioned from annual recurring revenue (ARR) to traditional EBITDA-based covenants, compared with four in the prior quarter."

For each ARR borrower, check the conversion date against maturity, how the credit agreement defines recurring revenue, and where the borrower would stand on the EBITDA test if it converted today.

The conversion date is when an ARR loan's assumptions get tested. Know now where each borrower would land.

8. Questions for Management at the Next Quarterly Call

Send these before the call, and ask the same ones every quarter: the change in the answers tells you more than any one answer.

1. Which of the jobs your product does could a customer now hand to an AI assistant or agent?
2. How much revenue is priced per seat, and how have seat counts at your 20 largest customers moved this year?
3. What were gross and net revenue retention this quarter by segment, and what drove the change?
4. Which customers asked for discounts, shorter terms or free AI features at renewal?
5. Which AI-native competitors show up in your sales cycles, and how often do you lose to them?
6. What data does the product hold that a new competitor could not get?
7. What have you shipped with AI in the last 12 months, and do customers pay for it?
8. How has R&D spending moved, and what would you cut first if growth slowed?
9. If the covenants converted to an EBITDA test today, where would you stand?
10. What would the sponsor do if the business needed more equity?

Record each answer with its date. A vague answer is a finding too.

9. Where It Shows Up in Marks and the Credit Memo

The review feeds three places.

• The credit memo. Add an AI exposure section to each software borrower's memo and annual review: the ten scores, the evidence and what would change the view.
• The watch list and risk ratings. Agree the triggers in advance: PIK switched on, net retention below its level at closing, an amendment request, a large customer lost to an AI-native rival.
• The valuation committee. Give it evidence, never a number: the review informs inputs such as the revenue outlook, and the committee's process sets the mark (see AI for private credit valuation).

Run a downside case on the borrowers at the top of the list (see credit portfolio stress testing). What examiners ask about valuation is in the guide to the SEC's 2027 exam priorities for private credit.

The review informs the mark. People set it.

10. Reviewing a Whole Book in 30 Days

Triage first, then go deep where exposure and hold size meet.

1. Week 1, list and sort. Every software borrower with hold size, mark, maturity, PIK status, covenant type and any ARR conversion date, tagged as a horizontal or workflow tool or as vertical software.
2. Week 2, screen. Run the first four factors on every borrower from public information and your own file. Rank by exposure, then by hold size.
3. Week 3, review the top tenth. Take the highest-ranked tenth through all ten factors and send them the management questions.
4. Week 4, write it up. One page per borrower: scores, evidence, open questions and a watch list recommendation, ready for the credit committee.

Keep the one-page limit. It forces the analyst to name the two or three facts that matter.

Repeat the screen every quarter, and rotate the full reviews as the ranking changes.

11. Using AI to Run the Review

Much of this review is reading, and AI tools read fast. Claude or a similar assistant can sort a borrower's product pages, release notes, customer reviews, job posts and listed rivals' earnings calls under the ten factors, with a link for every claim.

• The substitution test. Describe the borrower's core job to a general assistant, using public or dummy data, and see how close it gets. If it gets close in an afternoon, record that under factor one.
• The comparison. Set this quarter's management answers against last quarter's, and pull every ARR definition and conversion date into one table.

Analysts check every claim against its source before it enters a memo. AI never sets a mark or a rating. Borrower documents go only into a business plan the firm controls, under its data rules, never into a personal account; the guide to Claude for private credit covers the setup.

Let AI do the reading. The judgment stays with the people who sign the memo.

12. Where to Start

This week, pull the list of software borrowers. Then:

1. Tag each borrower by category, with hold size, maturity, PIK status and covenant type.
2. Send the ten questions before the next quarterly call.
3. Run the table on the top tenth by exposure and size.
4. Write one page per borrower and take the watch list changes to the credit committee.

For buyout targets rather than loans, the same logic appears in assessing AI disruption risk in a target.