AI Feature Attribution: Pulling ARR Lift Out of the Noise

AI features get marketed by their headline ARPU and retention numbers and almost never by their attribution. Most "the AI feature drove $X of ARR" claims fall apart under the simplest decomposition: how much of the gap between the AI cohort and the non-AI cohort is the feature, and how much is the user mix that chose it. This walkthrough runs a $1.2M ARR product through the AI Feature Attribution engine, then strips the result down to what is plausibly causal.

1. The $1.2M ARR scenario

The product is a B2B SaaS at $1.2M ARR with 2,400 total users. 38% of users (912) are on an AI-enabled tier; the remaining 62% (1,488) are on the standard tier. Reported ARPU uplift on the AI cohort is 22% vs the non-AI cohort. Monthly infra cost specifically attributable to the AI tier is $4,800 — that covers Claude Sonnet API spend, embedding storage, and the vector database. Annual churn on the AI cohort runs 1.2 percentage points lower than the non-AI cohort.

The blended ARPU is $1,200,000 / 2,400 = $500 per user per year. The non-AI baseline ARPU works out to $461.42. With a 22% uplift, the AI cohort ARPU is $562.94. The engine returns these values exactly because the math is determined by the cohort split and the total. None of the numbers have wiggle room. What has wiggle room is the interpretation.

2. The ARPU delta: $562.94 vs $461.42

The dollar delta per user is $562.94 − $461.42 = $101.52 per year. Multiplied across 912 AI-tier users, the total ARPU lift is $92,586 per year. The engine rounds this to $92,580.29 because of the cohort-split arithmetic; the two figures are the same number within rounding noise.

The decomposition matters because this number gets cited as "the AI feature drives $93k of ARR." The number is technically correct in the sense that the cohort exhibits that gap. It is misleading in the sense that you cannot attribute the entire gap to the feature unless you have evidence that user assignment to the AI tier was random. In a self-serve SaaS with an AI upgrade, assignment is the opposite of random: the users who opted in were already paying more, using more features, and showing more engagement before they saw the AI tier. The churn calculator can sanity-check the cohort difference against baseline behaviour.

3. $92,580 attributable, $34,980 net

Net attributable ARR is gross ARR attributable to the AI feature minus the AI infrastructure cost. The engine returns:

# ai-feature-attribution (computed live from /engines/ai-feature-attribution.js)
Engine input
  total_arr             = 1200000
  ai_specific_users_pct = 38
  ai_users_arpu_uplift_percent= 22
  ai_infra_monthly_cost = 4800
  churn_difference_pct  = -1.2
  total_users           = 2400

Engine output
  aiUsers               = 912
  nonAiUsers            = 1488
  baseArpu              = 461.42
  aiArpu                = 562.94
  arrAttributableToAi   = 92580.29
  annualAiInfraCost     = 57600
  netAiAttributableArr  = 34980.29
  grossMarginPercentAiCohort= 68.78
  grossMarginPercentNonAiCohort= 80
  effectiveChurnReductionPercent= 1.2

The $57,600 annual infra figure is the cost of running the AI feature at the current MAU. It is not the cost of building the feature, which is separately a one-time engineering expense that gets amortised over the feature's useful life. The cost of building rarely appears in attribution math because solo founders treat it as a sunk cost and ask only "is the feature paying its monthly bills." That is the right question if the feature took less than two months to build; the wrong question if it took six months and is still being maintained.

Gross margin on the AI cohort is 68.78%, against 80% on the non-AI cohort. The 11.2-point gap is the per-user share of the $57,600 infra cost ($63.16 per AI user per year). For comparison, the per-user ARPU lift is $101.52. The feature is gross-margin positive on a per-user basis with $38.36 of headroom, before any selection-bias correction.

4. Selection bias eats most of the headline

Self-selection into a paid AI tier is the single largest distortion in attribution math for solo SaaS founders. The users who opted in had three pre-existing traits: higher historical ARPU, lower historical churn, and higher feature engagement. They would have produced a higher ARPU even if the AI tier did not exist, because they are different users. The Stanford HAI 2024 AI Index reports adoption skews heavily toward larger, more engaged accounts^[3], which makes self-selection the dominant explanation for cohort ARPU gaps in early AI feature rollouts.

Three corrections produce a defensible attribution number:

• Match on pre-treatment ARPU. Compare AI-tier users only against non-AI users who had the same ARPU in the 90 days before the AI tier launched. The matched-comparison ARPU gap is typically 40 to 60% smaller than the raw cohort gap. Applied here, the $101.52 per-user delta drops to $40-$60, and net ARR drops to $14,000-$20,000.
• Run a holdout group. Block AI feature access for a random 10% of eligible users for one quarter. The ARPU gap between the holdout and the treatment group is closer to the causal effect. Most solo SaaS will not do this because it feels like leaving money on the table; it is the only clean way to know the answer.
• Use a regression discontinuity on rollout date. If the AI tier launched on a specific date, compare ARPU trajectory of users who signed up the week before launch vs the week after. The pre-launch cohort never had AI access; the post-launch cohort did. The trajectory difference is causal under reasonable assumptions.

For a solo founder who cannot run any of these cleanly, the working assumption should be that the causal ARPU lift is 40 to 60% of the headline. The $92,580 attributable ARR figure is the optimistic ceiling; $37,000 to $55,000 is the realistic band; $14,000 to $20,000 of net ARR after infra is the conservative anchor.

5. The 1.2-point churn lift, separately

Churn reduction is more credible than ARPU uplift because retention is harder to self-select into. A user does not opt into the AI tier to lower their churn; they opt in for the feature and then either keep using it or do not. The 1.2-percentage-point churn reduction is the engine's `effectiveChurnReductionPercent` output, which holds up under cohort decomposition better than ARPU.

In dollar terms, a 1.2-point annual churn reduction on a cohort of 912 users at $562.94 ARPU is roughly 11 users retained per year (912 × 0.012), worth about $6,192 of ARR. That is a separate number from the $92,580 ARPU attribution and should not be double-counted. The engine reports them separately for that reason. ChartMogul's 2024 retention data places median net revenue retention for B2B SaaS around 100%, with top quartile near 110%^[1]; a 1.2-point churn improvement nudges NDR upward by roughly the same amount.

6. Sensitivity bands on every input

The engine's output bends on three inputs: AI cohort share, ARPU uplift percent, and infra cost. Sensitivity to each:

• Cohort share 38% → 28%: AI users drop from 912 to 672, attributable ARR drops from $92,580 to $68,221. Net ARR after infra ($57,600) drops to $10,621. The feature is barely positive.
• ARPU uplift 22% → 12%: ARPU delta drops from $101.52 to $55.37, attributable ARR drops from $92,580 to $50,499. Net ARR drops to negative $7,101. The feature loses money on infra.
• Infra cost $4,800 → $7,200/month: Annual infra rises from $57,600 to $86,400. Net ARR drops from $34,980 to $6,180. Bessemer's 2024 cloud benchmarks note model costs are the fastest-rising line item in 2024-2026^[2]; this scenario is plausible within 12 months. Anthropic's published pricing^[4] remained stable through 2025, but the volume-driven absolute cost trends up.

The single most important sensitivity is ARPU uplift. A 10-point error there (22% vs 12%) flips the feature from net-positive to net-negative. That is the input most exposed to selection bias. Founders who run the attribution engine quarterly and watch this single number catch problems early; founders who rely on the raw cohort gap discover the problem when the next round of cost shocks pushes the feature underwater.

7. What this number is for

Attribution numbers are decision inputs, not marketing claims. Three decisions they should inform, and one they should not:

• Should we keep building the AI feature? If net attributable ARR (after selection-bias correction) clears 2x the annual infra cost, yes. In this scenario, the conservative band of $14k to $20k of net ARR against $57,600 of infra is below 1x. The honest answer is "maybe not, until ARPU uplift is verified causally."
• Should we raise the AI-tier price? If the cohort gap survives a matched-comparison correction, yes — there is willingness to pay you are not capturing. If the gap collapses, the price is already at or above causal value.
• Should we accelerate AI-feature investment? Only if the gross margin gap (11.2 points here) is shrinking as scale grows. If the gap widens, the feature has bad unit economics and more investment makes the problem larger.
• What it is not for: investor pitches. The $92,580 number reads great in a deck and falls apart in due diligence the moment an analyst asks for the holdout group. Use the conservative net-net figure ($14k-$20k) when claims will be verified, and the gross figure only with a clear "before selection-bias correction" footnote.

One operational rule. Run attribution quarterly, not monthly. Monthly cohort gaps are too noisy to act on; annual gaps lag too far behind product changes. A four-quarter rolling decomposition catches feature drift, infra cost creep, and the slow churn improvement that usually accumulates after the first six months of a feature being live. See the methodology page for the full derivation^[5].