Software Credit Due Diligence: ARR Quality, Churn and Covenants

Software has become the core collateral of the private credit market, and that is precisely why the AI reset matters to lenders. BIS Bulletin 128 estimates that business development companies have lent around $115 billion to software firms, roughly a fifth of all their lending and more than 80% of their fast-growing technology portfolios. The same bulletin finds that generative AI's revenue uncertainty has not yet affected these loans, but that credit spreads have narrowed, reducing the buffers available to absorb losses if it does. S&P Global Ratings likewise notes that concerns about AI disruption of software companies have contributed to heightened volatility across leveraged finance, wider discounts to NAV for listed BDCs and intensified redemptions from non-traded vehicles. The transmission channels run through valuation marks, earnings pressure and refinancing risk rather than immediate defaults.

The underwriting itself is also changing. Of the eight software loans Lincoln International placed in the first four months of 2026, none were annual recurring revenue loans; all were priced as a multiple of EBITDA, which sizes debt against the cash a borrower actually generates. At the height of the market, ARR loans priced at 525 to 550 basis points over benchmark with loan-to-value ratios of 30% to 35%. The borrowers behind the recent EBITDA-based loans each generated at least $100 million of cash operating profit, with median leverage of 5.7x and median loan-to-value of 30%.

• Recurring revenue is no longer accepted as a proxy for debt-service capacity; lenders size against cash generation.
• Retention, concentration and pricing-model durability have moved from commercial diligence topics to credit topics.
• Pro forma and run-rate EBITDA adjustments face sharper scrutiny in software than in other sectors.
• Diligence scope now extends to AI substitution risk, inference cost exposure and pricing-model transition, not only historical financials.

ARR quality is tested before any multiple or leverage assumption is applied, because the label recurring covers revenue streams that behave very differently in a downside case. Contractual ARR, committed in signed agreements for fixed terms, is the strongest form. Implementation, services and one-off revenue are often bundled into a headline ARR figure but do not recur at the same rate. Usage-based and hybrid revenue can scale with customer success, yet it also contracts fastest when customers cut consumption, so it deserves its own sensitivity line in the model.

Concentration, contract duration and the renewal profile determine whether that ARR survives the loan tenor. Advisers report that lenders now scrutinise churn rates and customer concentration closely before underwriting software deals. A renewal cliff concentrated inside the debt maturity, or a top-ten customer group representing a large share of ARR, converts an apparently stable revenue base into a refinancing event. A structured approach to ARR durability starts with the ARR bridge rather than the headline number.

• Decompose headline ARR: contractual subscription, usage-based, services and implementation, and one-off revenue.
• Test customer concentration by revenue and by logo, including the renewal dates of the largest contracts.
• Map the renewal calendar against the loan tenor to identify cliffs inside the holding period.
• Request billing extracts, executed contract terms and cohort-level ARR bridges rather than a single management-reported ARR figure.

Headline ARR growth can hide deterioration, because expansion revenue masks churn in the net figure. The 2026 Aleph and Benchmarkit SaaS and AI Performance Benchmarks, covering full-year 2025 data from 342 companies, put median gross revenue retention at 84%, down 4 points year over year, while median net revenue retention held at 102%. That 18-point spread is how much work expansion is doing to paper over churn. Because the decline hit every quartile, the report reads it as a market-level shift driven by longer sales cycles, ROI scrutiny and buyers evaluating AI-native alternatives, not as isolated execution failures.

Churn then needs decomposing. Logo churn, revenue churn and downsell have different causes and different credit implications, and cohort behaviour varies sharply by segment: the same benchmarks show sales-led companies retaining materially better than product-led ones, with enterprise contracts retaining best of all. A borrower whose NRR rests on expansion into a small set of large accounts carries different risk from one with broad-based retention.

• Reconcile GRR and NRR to the cohort-level ARR bridge so contraction cannot hide inside expansion.
• Separate logo churn from revenue churn and downsell; each implies a different margin and cash-flow path.
• Segment retention by ACV band and go-to-market motion before benchmarking against the 84% median.
• Flag renewal concentration and any customer representing a disproportionate share of contracted ARR inside the debt tenor.

The central AI question for a software lender is whether the borrower's revenue base can shrink independently of customer satisfaction. AI can reduce the number of seats a customer needs, automate the workflows those users performed, shift willingness to pay toward usage or outcomes, and lower the cost of entry for AI-native competitors. Each channel works on a different clock, but all four press on the same covenant base.

The market evidence so far points to transition rather than replacement. industry analysis and Company analysed more than 30 SaaS vendors introducing generative AI capabilities and found that roughly 65% layered an AI usage meter on top of seat-based pricing, while about 35% simply raised per-seat prices; none had fully shifted to usage- or outcome-based pricing. industry analysis also notes the execution burden of that transition: most vendors lack the product telemetry, billing and finance infrastructure to meter AI at scale, and customers' procurement teams are accustomed to buying by headcount, not by value. A framework for evaluating these shifts appears in AI software pricing diligence.

Substitution risk differs by segment. Arcmont's chief investment officer draws the fault-line between entrenched systems of record such as ERP, accounting and HR software, which carry very low churn because replacement is costly and disruptive, and point solutions layered on top, which historically churned more and are now the most exposed to AI-native challengers. Financeable assets are still clearing: Cegid's €1.1 billion private credit loan led by Arcmont and Ares closed in June 2026 after the syndicated market pulled away. Some lenders comfort themselves with the view that AI will erode margins over a six-to-seven-year horizon rather than two to three, long enough to make a five-year loan defensible.

Adjusted EBITDA is an underwriting input, not a fact, and in software it deserves particular scrutiny. Private credit deals have historically included EBITDA adjustments averaging 20% to 30% of total EBITDA, and lenders now view pro forma and run-rate adjustments in software, whether from price increases, newly signed contracts or planned headcount reductions, with sharper scepticism. Stock-based compensation, capitalized development, restructuring charges and acquisition add-backs each deserve to be tested against what recurs in cash terms.

AI adds a new cost line to the same analysis. ICONIQ data reported by SaaS Mag puts average AI product gross margin at 52%, with inference alone consuming roughly 23% of AI product revenue at scaling-stage companies. Public SaaS filers have begun disclosing inference-related costs generally between 4% and 9% of revenue. For a leveraged borrower, an AI roadmap can therefore compress gross margin even as revenue grows, which is why inference economics now belongs in credit diligence.

• Isolate AI and inference costs from generic cloud spend; they scale with usage, not with headcount.
• Reconcile adjusted EBITDA to operating cash flow, adjusting for working capital and deferred revenue dynamics.
• Deduct R&D capex and committed AI infrastructure spend to reach the free cash flow actually available for debt service.

Debt-service analysis starts from the structure, not the story. Recent software loans have printed at a median 5.7x leverage and 30% loan-to-value, with lenders testing interest coverage and minimum liquidity against the borrower's own cash generation rather than against projected ARR. Covenant design then determines how much warning a lender gets: leverage and coverage tests, the EBITDA definition embedded in them, and the choice between incurrence and maintenance testing all shape whether deterioration surfaces early or late. In covenant-lite structures, early-warning signals come from information covenants, reporting cadence and the borrower's own liquidity runway rather than from a breach.

The useful discipline is to run the downside as a cascade, quantified as a sensitivity rather than a single base case. Each stage feeds the next, and the output that matters is the covenant headroom at the end of the chain, a logic set out in covenant risk analysis for software borrowers.

• ARR shock: model a gross-revenue-retention scenario below the 84% market median, split between logo loss and contraction.
• Churn pass-through: translate the ARR shock into renewal outcomes by cohort, weighting concentrated contracts by their renewal dates.
• Margin compression: apply seat compression and inference cost assumptions to gross margin before operating leverage.
• Cash conversion: run the shocked EBITDA through working capital, deferred revenue and R&D capex to reach free cash flow.
• Covenant headroom: test the resulting cash generation against leverage, coverage and minimum liquidity covenants across the loan tenor.

The framework below consolidates the preceding sections into a single lender-facing view. It is a diligence map, not a credit opinion: it identifies what to test and what evidence supports each test, leaving the judgement to the investment committee.

Across every dimension, the standard is document-level evidence: executed contracts, billing extracts, cohort reports and board packs, not management summaries. Advisers describe lenders going through software situations with a fine-tooth comb, and the difference between a supported claim and an unsupported one is exactly what that scrutiny finds. The same document-first discipline applies to the frameworks in software moat diligence and AI impact due diligence.

Plausity is an AI-native due diligence and deal intelligence workspace that helps M&A advisory firms, VC and PE funds, corporate development teams and investment-banking teams structure evidence, findings and questions across a data room. Plausity supports evidence extraction, source grounding, findings management and IC preparation — it does not replace human analysts, advisers or investment professionals, does not provide legal, tax, audit, regulatory or investment advice, and does not make autonomous investment decisions. All findings require human review. Built for today's investment and deal teams. Trusted by >200 firms.

To explore the underlying capabilities, see the Plausity AI analysis engine, findings and risk intelligence and evidence gap detection product pages, plus the IC memo and AI Q&A Assistant product pages. For team-level workflows, see how VC and PE funds and M&A advisory firms use Plausity across live deals, and how AI Impact due diligence, value creation, Tech DD and Commercial DD workstreams support the analysis.

1. bis.org
2. spglobal.com
3. finance.yahoo.com
4. getaleph.com