ARR Quality for AI SaaS: What Investors Strip Out — and the Signal They Miss

Under a post of mine about build fees versus monthly fees, a founder asked the question this article answers: if competitors catch up in weeks and loyalty is "not even a thing", is ARR still a robust metric — and "what would be the nextgen key indicator?"

My answer: ARR did not stop working. It stopped meaning what the seed benchmark assumes. That benchmark was built for software that cost nothing to serve and took a year to rip out; AI software can cost real money per use and be replaced over a weekend. Same number, different thing underneath.

So the job is not to replace ARR. It is to run three deductions and one test on it before an investor does — and to name the bucket each dollar falls into.

What is ARR quality?

What does "ARR quality" mean for an AI SaaS?

ARR quality is the durability of the revenue behind the number: how much of it comes back next year without a new sale, a new build, or a price cut. For an AI SaaS, the honest measure of that durability is switching cost — what a customer would lose by leaving — because it is the only thing a cheaper competitor cannot copy in a week.

The classic ARR-quality checklist (contract structure, customer composition, net and gross retention) is not wrong — it is late. It describes revenue after the customer has already decided. Bessemer's 2025 survey of ten "Supernova" AI startups — averaging ~$40M ARR in year one and ~$125M in year two — comes with the warning that such adoption often "belies low switching costs" or signals novelty; the same report states plainly: "Retention can be fragile, especially when switching costs are low." Those ten companies also averaged roughly 25% gross margin, often negative.

In the book I call the thing that makes revenue durable the Workflow Wedge: how deeply the product is driven into the customer's work, so that removing it means dismantling the work itself. The test for a wedge is one question: if a competitor offered the same thing 20% cheaper tomorrow, what specifically would my customer lose by leaving? If the honest answer is "not much", you have a sticker, not a wedge — and the ARR on that contract is a sticker too.

What counts as ARR?

What counts as ARR — and what investors strip out first?

Before any quality judgement, there is a definitional one. The SaaS Metrics Standards Board defines ARR as the most recent month's recognized recurring subscription revenue, annualized — and it is explicit about the exclusions: "ARR should not include one-time fees or professional services even if they are recurring." Contracted-but-not-live revenue is out (that is CARR); usage above the contracted minimum is out; and if there is no contracted minimum at all, there is no CARR to count.

That standard is exactly what a seed investor applies to your deck. In my own founder conversations this autumn the shape repeats: a build or setup fee up front, a smaller platform fee after. Say the last twelve months brought in $400K — $310K from builds, setup and paid pilots, $90K from monthly fees. The deck says $400K. The room hears $90K.

Services revenue versus recurring revenue is not a judgement about your business — plenty of good companies start with services. It is a judgement about which column next year depends on, and investors pay for the column that comes back without being won again.

Deduction 1

Deduction 1 — ARR that costs you money to keep

The first AI-specific deduction has no SaaS precedent. In classic software every ARR dollar carried an 80–90% gross margin. In AI, part of your ARR is usage revenue with inference, GPU, vector-database and human-review cost inside it. If the AI Layer Gross Margin on that revenue is negative, you have ARR that costs you money every month you retain it. Growing it makes the business worse. That is the first of the four ways to die at the peak: death by your own growth.

Investors look for this. A $300K usage line at a −15% AI-layer margin is not $300K of ARR to them; it is a liability with a recurring-revenue label. Before you present any ARR, split out the usage-driven part and compute its margin separately (the free AI Gross Margin Calculator does this in eight fields). The margin math itself is covered in the AI Layer Gross Margin guide, and the accounting question — whether production inference belongs in cost of revenue — in Should AI inference cost be COGS or OpEx?

Deduction 2

Deduction 2 — Pilot ARR and LOI ARR

The second deduction is revenue that has not yet decided to exist. AI buying is pilot-heavy: MIT's NANDA research, as reported by Fortune in August 2025, found only about 5% of AI pilot programs in its dataset reaching rapid revenue acceleration (150 interviews, a 350-employee survey, 300 public deployments). Directional, not your conversion rate — but the direction is the point: a paid pilot is a customer evaluating a decision, not one who has made it.

The forensic case here is 11x. Per TechCrunch's reporting in March 2025, the company sold 12-month contracts with an opt-out after three months and reported contracted ARR (CARR) to its board. A former employee told TechCrunch the company might cite $14M in ARR when contracts that survived the three-month trial totalled about $3M, and described churn in one line: "We were losing 70-80% of customers that came through the door." The company disputed the framing. Whatever the individual numbers, the mechanism is general: when the metric counts the contract and the customer counts the opt-out, ARR is a time-bomb metric — healthy until someone splits it.

Deduction 3

Deduction 3 — Zombie ARR: invisible until renewal

The third deduction is the one no dashboard shows. Zombie ARR is revenue where the logo is present and the activity is gone — held in place by a contract until the renewal date, when it leaves all at once. A founder put it to me this way: the invoices keep going out and the margin keeps booking, but nobody on the customer side is still using the product — and nothing in the metric will move until the contract ends, nine months from now.

The data says this is the AI-native base case, not the exception. Kyle Poyar's ChartMogul analysis of 3,500 software companies with at least $250K ARR found a 2025 median gross revenue retention of 40% for AI-native companies (median NRR 48%), against a median NRR of 82% for B2B SaaS — and the AI-native GRR figure had improved from 27% in January to 40% in September as the early tourists left. His phrase for it: "AI has a tourist problem." And Software Equity Group reports that 28% of SaaS CEOs do not track gross retention at all — for them the zombie share is unknown by construction.

The test

The Switching-Cost Test: five questions to run before the investor does

What survives the three deductions is real recurring revenue; now score its durability. This is the Workflow Wedge question broken into five parts, each scored 0 (no), 1 (partly), 2 (yes) per customer or segment, for a total of 0–10. The bands are my practitioner judgement, not a published benchmark.

1. 1Loss on exit. If the customer left tomorrow, would they lose accumulated data, context, or a team protocol that lives in your product? Not "they'd miss the feature" — what would they have to rebuild?
2. 2Who paid for embedding. Did the customer invest their own money or hours in implementation — a build fee, an integration, training? A paid setup is evidence of switching cost; it is not revenue, but it is a signal, and this is where it counts.
3. 3Share of workflow. How many of the customer's working steps pass through your product, as opposed to one task they could route elsewhere? Research on online services found that "systems usage measures and systems quality are associated with reduced switching" (Chen & Hitt, 2002) — depth of use is a measurable predictor, not a hope.
4. 4Parity shock. If the foundation-model vendor shipped your core feature for free next quarter, would the customer stay for the workflow, or was the model the product?
5. 5Renewal behaviour. Has usage grown from month three to month twelve, and did the first renewals happen at full price without a rescue call?

Setup fees

Why a setup fee is a signal — but not revenue

Back to that founder's point: setup fees do signal that customers "won't step away tomorrow", and the switching-cost literature says why. Burnham, Frels and Mahajan split switching cost into procedural (time and effort), financial (money already sunk) and relational (the relationship itself); a paid implementation creates all three at once, and economists since Klemperer count start-up and implementation costs as switching costs by definition.

So the build fee is evidence of a wedge — and still not ARR. The forensic twist most founders miss: accounting already agrees with the investor. Under ASC 606-10-55-51 (IFRS 15 B49), a non-refundable setup fee that transfers no distinct service to the customer is "an advance payment for future goods or services" — recognized over the service period, not at invoice. Your auditor spreads it; your investor strips it; only your deck counts it up front. Price the build fee as what buys the wedge and the platform fee as what the wedge protects — the two-part tariff from the pricing-models guide — and report them as two different things.

ARR multiple

What ARR quality does to your SaaS ARR multiple

Quality shows up in the multiple, but not as a clean table. Software Equity Group's data for 2024: companies with net revenue retention above 120% achieved a median 11.7× multiple, more than double the industry median of 5.6× — a revenue multiple from SEG's Annual SaaS Report, not a private ARR multiple for your round. No primary source I could verify publishes a "GRR band → private multiple" discount; anyone quoting one is estimating. (For orientation: Aventis Advisors' 2,172 software M&A deals, 2015–early 2025, show a median 3.0× revenue, first quartile 1.5×.)

In practice the room applies its multiple to the Wedge ARR, discounts the Sticker ARR, deletes the Zombie ARR — and marks you down for having left the last two in the number.

Five times a number you can defend beats thirteen times a number that splits under one question.

Worked example

A worked example

A composite of founder conversations from the past two months — figures illustrative, rounded, and anonymised. A deck shows $1.5M ARR. Split it:

• $420KAnnualised build and setup fees. One-time. Out — but logged as evidence of switching cost for the customers who paid it.
• $180KFive paid pilots on three-month opt-outs. Pipeline, not ARR. Out, with a conversion rate attached.
• $300KUsage revenue on the AI feature, running at a negative AI-layer margin after inference and review cost. Deduction 1: flagged as margin-negative until routing or pricing fixes it.
• $600KLive platform subscriptions — run the test. $350K scores 8–10 (customers co-funded implementation, their data and approval flows live in the product, usage up since month three): Wedge ARR. $200K scores 4–7 (single-feature accounts a model update could replace): Sticker ARR. $50K is three logos with usage down 70% and renewals next quarter: Zombie ARR.

The deck said $1.5M; the room underwrites $350K, with a credible path on $200K more. Neither number is a lie; only one survives a question.

Prescription

What to do this quarter

Diagnosis without a prescription is just a scare, so:

1. Split the columns in your own reporting before anyone else does: once (builds, setup, pilots) versus every month (platform, subscription, contracted usage).
2. Compute the AI-layer margin on every usage line and move margin-negative ARR out of the quality number until it is fixed.
3. Score each account on the five questions; put the Wedge / Sticker / Zombie split and the renewal timeline in the board deck. Investors trust a named weakness more than a hidden strength.
4. Build the wedge deliberately. The book covers this in two places: chapter 3 on the Workflow Wedge — how to turn a sticker into something that takes root — and chapter 8 on time-bomb metrics, including the honest-accounting fixes for ARR, LTV and payback. The AI Survival Canvas puts both on one page.

ARR is still the metric. The next-generation indicator that founder asked for is not a replacement for it; it is the share of ARR that would survive the 20%-cheaper test. Report that share and the room stops splitting your number — you already did.

FAQ

How this was gathered. Benchmarks come from named primary sources — a VC report, a standards body, an M&A advisor's data, a 3,500-company retention dataset, original journalism and the accounting standards themselves — each opened and verified individually before use. The Switching-Cost Test bands and the Wedge / Sticker / Zombie split are first-hand fractional-CFO practice, not third-party data; the worked example is an anonymised composite of founder conversations in September–October 2026. Frameworks here are teaching tools for your own analysis, not financial forecasts or advice.

Sources. Bessemer Venture Partners, The State of AI 2025 · SaaS Metrics Standards Board, Annual Recurring Revenue and Contracted ARR standards · Kyle Poyar / ChartMogul, The AI Churn Wave (Dec 2025) · Software Equity Group, SaaS Multiples 2025 (Apr 2025) · Aventis Advisors, Tech Company Valuation Multiples (Aug 2025) · TechCrunch, 24 Mar 2025 (11x) · Fortune, 18 Aug 2025, on MIT NANDA The GenAI Divide · FASB ASC 606-10-55-50 to 55-53; IFRS 15 B48–B51; FASB–IASB TRG Agenda Paper 6 (Oct 2014) · Burnham, Frels & Mahajan, Journal of the Academy of Marketing Science (2003) · Chen & Hitt, Information Systems Research (2002) · Klemperer (1995); Farrell & Klemperer (2007).