Britain’s lenders have now raised the same concern twice. The Bank of England’s July report said banks had less appetite to lend to firms at risk of revenue erosion from artificial intelligence. Its September report repeated the warning.
That repetition makes AI exposure a real credit question. It does not yet show how much a borrower pays for it.
The clearest public evidence that concern has changed a decision comes from outside Britain. Goldman Sachs Private Credit said it passed on its first software deal because of AI concerns in October 2023. That is one attributable denial. Goldman did not identify the borrower, proposed facility, spread, leverage, covenants or the size of a wider rejected pool.
The distinction matters to bank credit officers, software-company finance chiefs and private-credit investors. A warning can change questions in a credit memo before it produces a measurable premium in a term sheet.
A selective market is still a competitive market
The Bank’s two Agents’ summaries are qualitative intelligence from business contacts, not a statistical survey of credit decisions. July said credit conditions were broadly unchanged and borrowing appetite was weaker than supply. It also said firms exposed to AI-driven revenue erosion had to rely on second-tier funding, alongside smaller companies and borrowers in construction and hospitality with weaker lender demand.
September retained the AI warning while describing an increasingly competitive market. Credit supply continued to outstrip demand, the gap broadened modestly across company sizes and distressed borrower levels remained low. Tier-one lenders were competing mainly on price for viable borrowers. The same update said acquisition finance was readily available and competition was encouraging some early refinancing.
Those statements can coexist. Banks may compete hard for businesses that clear their underwriting tests while steering a narrower group towards faster or more risk-tolerant funders. What the release does not show is how many borrowers were placed in the AI-risk group, how lenders defined revenue erosion, or whether those borrowers paid more after moving to a second-tier provider.
The Bank’s July Financial Stability Report takes the evidence one step further, but no further than the decision process. It says some financial firms already assess AI’s potential impact on businesses as part of loan-level decisions. It does not name the firms or sectors and gives no approval rate, spread, maturity, covenant, advance rate or denial count.
This is evidence of a new underwriting variable, not evidence of a new market price.
One rejection clears a low evidence bar
Goldman’s disclosure supplies the missing behavioural example. The manager said it introduced an internal framework for AI disruption risk in early 2025. For new software investments since January of that year, it reported an average Rule of 40 score of 55.7%, entry loan-to-value of 33.7% and EBITDA of $192 million.
Those figures describe the companies Goldman accepted. They do not establish that the framework caused tighter leverage, stronger covenants or wider spreads. There is no before-and-after series, no score for rejected borrowers and no allocation of the one reported denial between revenue risk and other weaknesses.
Goldman also reported clear differentiation in traded debt. In March 2026, broadly syndicated senior-secured loans for the best-performing software names were down only 0.625%, while names perceived to face higher AI disruption risk traded at least 15% lower. That is a secondary-market price signal. It is not an origination spread or a bank loan term, and it should not be presented as one.
Private credit supplies the counterweight
The broadest loan-level test points away from an AI premium. A Bank for International Settlements study examined business development companies, US vehicles that originate about one fifth of direct loans. It counted about $115 billion of lending to software firms, roughly one fifth of BDC lending and more than 80% of their technology portfolios.
Using 152,116 spread observations involving 208 BDCs and 2,302 borrowers from 2015 through the end of 2025, the authors found no systematic spread difference between software and other technology loans after controlling for timing, lender, investment type, tenor, accrual status and principal size. Newly issued spreads had fallen faster than those on existing loans. Fewer than 1% of software loans were behind on payments.
The study stops before the most recent Bank warning and covers US private credit, not British banks. It therefore cannot disprove selective tightening in the UK. It does show why the two Bank reports cannot support a claim that lenders already charge a general AI premium.
More recent fund disclosures do not close the gap. Blackstone Private Credit Fund said selected software investments were marked down in the second quarter because of wider spreads and company-specific fundamentals. It reported about 510 basis points of spread and 39% loan-to-value for private debt investments in new portfolio companies during the quarter, but did not isolate software or AI-exposed deals. Blue Owl Technology Income Corp said market dislocation was beginning to create opportunities at improved terms, while reporting a 35% loan-to-value for its existing technology portfolio and a $400 million asset sale at 99.6% of par. It did not quantify new terms by deal or AI-risk band.
The term sheet is the next disclosure test
A bank credit officer now has reason to test whether a borrower’s revenue depends on narrow software features, easily replicated workflows or weak switching costs. A software-company finance chief should be ready to show retention, pricing power, proprietary data and the cost of adapting the product. A private-credit investor should ask managers for acceptance rates and terms by AI-risk band, not settle for portfolio averages.
The useful next data point is not a third warning. It is a matched comparison: similar borrowers, separated by assessed AI exposure, with their approvals, spreads, leverage, maturities and covenant packages disclosed over time.
For now, Britain’s banks have identified the question twice. One global private-credit manager has documented a rejection. Public evidence still does not put a price on the risk.
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.