Everything About AI
NextFin News - Companies racing to build artificial-intelligence capabilities are doing something that has set off alarm bells in credit markets: they are counting AI cost savings that have not happened yet as if they already have. Borrowers are folding projected AI-driven productivity gains into the adjusted EBITDA figures that determine how much debt they can carry and whether they stay within loan covenants, extending one of leveraged finance's oldest accounting tricks into the newest technology boom.
The timing could hardly be worse. Debt investors are already repricing technology credits on fears that AI will disrupt software business models. Now the same lenders are being asked to underwrite leverage ratios that assume AI will cut costs — a bet that, if it fails, would squeeze covenant headroom across a wide swath of borrowers at the same time.
The Oldest Trick in the Leveraged-Finance Playbook, Now Wearing an AI Label
What is new is the label, not the mechanics. Credit agreements have long allowed "pro forma cost savings" and "run-rate synergies" to be added back to EBITDA, treating projected expense reductions as though they were realized on the first day of the measurement period. Standard covenant language defines these addbacks as "net cost savings and acquisition-related cost synergies projected by the Borrower in good faith to be realized," backed by officer compliance certificates and caps commonly set in the high single digits to mid-teens percent of consolidated EBITDA. The borrower gets today's borrowing capacity against tomorrow's promise.
The practice is not marginal, and the track record is poor. S&P Global Ratings' eighth annual EBITDA addback study, published February 25, 2026, found that addbacks averaged 29% of management-adjusted EBITDA on a median basis at deal inception across the study's history — and that marketing projections have "consistently proven overly optimistic." The consequences show up in the metric lenders actually care about: the study recorded a median leverage miss of 2.3 times in the first year after a deal closes and 2.7 times in the second. Among transactions originated from 2015 through 2020, 95% of companies failed to meet their first-year EBITDA projections, and roughly 55% missed earnings by a third or more over the following two years.
Technology is the sector where the practice is most pronounced. High-technology issuers saw addbacks equal to 66.1% of reported last-twelve-months EBITDA and 35.9% of pro forma adjusted EBITDA on average over the eight-year study window. When AI cost savings enter that same definition, the leverage ratio at origination can look materially lighter than the cash earnings that will actually service the debt. The largest single category of addback across the study's history was "expected synergies and cost savings," comprising 28.2% of all addbacks on an eight-year average — the exact bucket where AI promises now land.
"The reported leverage might be 6x to 7x, but if you look at it on a pure cash basis, in some cases it's closer to 8x to 9x," portfolio-analytics research on CLO collateral has noted.
The gap between marketed and realized earnings is not theoretical. Earlier research by Moody's found that only 45% of one-off cost-saving addbacks were achieved on average, and nearly 20% of issuers achieved none of their projected adjustments. An academic study of non-GAAP EBITDA addbacks estimated that each additional addback raises the probability of a loan becoming 60 days delinquent within three years by 4.2 percentage points, from an unconditional baseline of 1.3%, and lifts default likelihood by 1.6 percentage points. The mechanism is straightforward: more addbacks mean more debt at closing, and when the savings do not arrive, the debt remains while the EBITDA that justified it disappears.
None of this is unique to AI. During the COVID-19 pandemic, borrowers leaned on addbacks for pandemic-related costs; in the buyout boom, synergy addbacks did the same work. Each cycle produces the same pattern: definitions loosen while money is cheap, and tighten only after losses. AI savings are simply the current vocabulary for a familiar argument — pay us for earnings we promise to create.
How an AI Savings Addback Actually Gets Into a Credit Agreement
The mechanics matter because they explain why the risk is hard to see at closing. A borrower proposing an AI-driven cost program will typically present a plan: consolidate servers on fewer cloud instances, automate tiers of customer support, reduce manual back-office processing. Each line carries an estimated annual run-rate saving. The credit agreement then permits a "pro forma cost savings" addback, often capped at a percentage of EBITDA and conditioned on an officer certifying that the actions have been taken or are firmly committed, and that the savings are expected to be realized within a defined window — frequently 12 to 18 months.
That structure creates three distinct failure modes. First, the savings may be real but smaller than projected — a 10% miss rather than a total failure. Second, the savings may arrive later than the covenant measurement date, which is what the cap and certification are supposed to police but often do not in practice. Third, and most damaging for lenders, the savings may be offset elsewhere: AI cost cuts that coincide with AI-driven revenue disruption produce a net EBITDA outcome that no addback definition can rescue. A software company that saves $20 million in support costs while losing $50 million in subscription revenue from displaced products has a negative EBITDA problem that no cost-savings addback fixes.
This is where the AI variant diverges from plain M&A synergies. Procurement synergies fail because a supplier refuses to give the discount; AI savings fail because the technology changes the demand for the product itself. The addback assumes the cost structure changes while the revenue model stays intact — precisely the assumption the equity market is now questioning in software.
Why the AI Variant Hits Different
AI-driven savings are not the same as merger synergies, and that distinction matters for both sides of the argument. M&A synergy addbacks depend on integration plans that may never be executed; AI savings — server consolidation, automated customer service, reduced back-office headcount — can appear in operating-expense line items within a few quarters and leave an audit trail. Proponents argue this makes them more verifiable than the synergies of past cycles.
But the verification argument cuts both ways. The savings are real only if the AI deployment works as planned, at a moment when the market is actively questioning whether AI returns justify the spending. Capital-expenditure estimates for the five largest U.S. technology companies have risen sharply to roughly $800 billion, according to Man Group's second-half 2026 credit outlook. Morgan Stanley forecasts global AI-related debt issuance reaching about $570 billion in 2026, with $236 billion priced by the end of May — four times the year-earlier pace. Hyperscaler bond issuance alone totaled $220 billion through August 10, up from $12.5 billion in the comparable 2025 period, according to BNP Paribas data.
Into that supply wave, investors are already demanding concessions. Alphabet's August 2026 bond sale was well received but still required a concession of roughly 10 to 15 basis points relative to existing bonds. Salesforce issued A-rated notes at spreads above the U.S. BBB index as newsflow on AI disruption weighed on software business models. Technology now accounts for 4% of the investment-grade credit market, up from 1% a decade ago, and roughly $330 billion of software and technology debt matures through 2028.
The second-order risk is correlation, and it is the part of the story the headline leverage ratio hides. In a covenant-lite market where maintenance covenants are rare, there is no early-warning tripwire when AI savings miss. Incurrence tests are backward-looking and can often be navigated with additional addbacks or basket capacity. The first real signal arrives when a whole cohort's capacity tightens at once — right as asset prices are already falling. That is a different risk from the 2008 model of covenant breaches triggering defaults; it is a 2026 model of no breach, no default, and steadily deteriorating recovery value as the collateral mark drifts lower with no mechanism to intervene early.
Software and services loans have been the weakest corner of the leveraged market in 2026. The Morningstar LSTA U.S. Leveraged Loan Index fell 1.85% from mid-January, while its software and services component — 15.89% of the index — dropped 7.76%. Roughly 14% of software loans were trading below 80, with the sector averaging about 89, and loans from issuers flagged for AI disintermediation risk averaged $86.4, according to Bank of America research. For context, a loan trading at 86 implies the market is pricing a meaningful probability of restructuring losses — not merely a spread-widening event.
That repricing flows straight into collateralized loan obligations. Software and services make up 12.1% of U.S. broadly syndicated loan CLO collateral and 9.11% of the European book. Junior overcollateralization cushions narrowed 82 basis points year over year to 3.49%, and 7.1% of U.S. BSL CLOs had failed their most junior OC test. Median quarterly equity distributions fell to 2.4% in the first quarter of 2026, a full percentage point below the historical average and the lowest payout since 2018, according to Deutsche Bank Research. If AI savings miss and software loans mark down further, cash flows get redirected away from CLO equity to repay debt tranches — a mechanical squeeze that forces selling into a falling market, which marks the collateral down further, which tightens more OC cushions. The loop is self-reinforcing.
"Beyond displacement risks for software companies, there are significant technical risks and valuation pressures that the market is processing," said Andrew Ward, a portfolio manager at Eldridge Capital Management.
Cyclical Looseness Meets a Structural Shift
The right read separates two forces that are easy to blend. The accounting looseness is cyclical: documentation standards expand in cheap-money periods and tighten only after losses, a pattern that has repeated for two decades. The AI cost disruption underneath it is structural — a genuine regime change in how software and services companies generate margins that will not revert on its own.
That separation produces an uncomfortable conclusion. Some AI savings will be real and durable, and companies that achieve them will deleverage faster than skeptics expect. But the EBITDA definitions being written today are priced for a best-case adoption curve. If realization lags, the leverage miss will not be idiosyncratic; it will arrive clustered across borrowers that sold lenders the same story.
The private-credit channel amplifies this. The Financial Stability Board's May 2026 report on private-credit vulnerabilities warned that growing use of EBITDA adjustments may understate headline leverage, with "true" leverage in private credit closer to 7x debt-to-EBITDA. Private credit sits at the center of the middle-market software lending that is most exposed to AI disruption, and its valuations are marked by managers rather than cleared by a market. When a private-credit fund marks a software loan at 95 while the syndicated equivalent trades at 86, the addback-driven leverage story looks far more comfortable on the fund's books than it would at a market price. The FSB's concern is precisely this gap between modeled EBITDA and cash EBITDA, between net asset value and the price at which the loan could actually be sold.
Regulators and standard-setters are already circling. IOSCO's good-practice guidance urges that pro forma adjustments for future synergies and cost savings be made on a reasonable basis, subject to credible time horizons and caps, and ideally independent review. The gap between that guidance and what appears in actual credit agreements is where the risk lives — guidance is not contract language, and a borrower-friendly definition negotiated in a competitive deal will almost always win over a regulator's suggestion.
The Case That the Alarm Is Overdone
The strongest argument against the bears is scale and verifiability. JPMorgan Private Bank estimates high-yield AI-related debt at about $62 billion — roughly 4% of the high-yield market — and leveraged-loan AI exposure at about 1%. On that measure, the AI addback story is a niche within a niche, not a systemic fault line. And unlike the synergy promises of the 2010s, AI savings can be traced to specific cost lines within quarters, giving lenders more to verify than a slide deck.
That case is coherent but incomplete. It addresses the stock of existing AI debt, not the flow of new documentation being written into the $570 billion AI financing wave. It also assumes lenders will enforce caps and certifications — yet the S&P record shows addbacks have persisted and grown despite two decades of documented misses. The guardrails exist on paper; whether they bite is a question of lender discipline, not contract language. And the scale argument ignores the concentration channel: it does not take much correlated stress in the 12.1% of CLO collateral sitting in software and services to hit junior tranches and equity distributions.
There is also a selection-bias problem in the verifiability argument. The AI savings easiest to verify — cloud consolidation, headcount reduction — are also the ones most likely already reflected in a company's run-rate before the deal. The savings that move the needle on a leverage ratio tend to be the forward-looking ones: automation that has not yet been deployed, support deflection that has not yet occurred. Those are the least verifiable by definition, and they are the ones doing the heaviest lifting in the EBITDA bridge.
What to Watch
The falsifying signal is concrete: S&P Global Ratings publishes its addback study annually. If the 2027 edition shows technology-sector AI addback realization above 70% and the median leverage miss narrowing below 1.0x in year one, the current alarm is overdone. If instead realization stays near the historical 45% and leverage misses hold above 2.0x, the AI savings addback will join the list of credit-market practices that had to be learned the expensive way.
Near term, watch three markers. First, whether new credit agreements cap AI-related savings addbacks below the historical high-single-digit to mid-teens range — a sign lenders are tightening rather than competing for volume. Second, whether software-loan prices stabilize above 90, which would indicate the AI-disruption repricing is exhausting itself. Third, whether CLO junior OC failure rates continue climbing past the current 7.1%, which would signal forced selling pressure rather than orderly mark-to-market adjustment.
The forward view splits cleanly by horizon. Short term, the pain is in mark-to-market and CLO equity distributions — a liquidity and valuation story, not yet a default story. Medium term, it is in refinancing: borrowers that sold AI savings to get today's deals done will face lenders asking for proof when roughly $330 billion of software and technology debt rolls due through 2028. That is when addback definitions get renegotiated, and when the difference between realized and projected savings becomes a pricing event rather than a footnote.
Long term, the structural question is whether AI genuinely rewires software margins. If it does, the companies that achieve durable savings will look back on today's skepticism as the bottom of the cycle. If it does not, today's EBITDA definitions will be remembered as the mechanism that let the industry borrow against a future that never arrived. The credit market is not betting on which outcome occurs; it is financing both, and charging the same price for each.
The AI buildout is colliding with the credit market's oldest accounting trick: selling lenders on earnings that have not happened yet. This time the earnings are labeled "AI savings," but the test is unchanged — and the last eight years of addback data suggest most of them will fail it.
Explore more exclusive insights at nextfin.ai.
