Adyen's AI rollout reaches 40% of merchant volume.

The architecture: Claude-grounded risk, Cohere embeddings, and an in-house orchestration plane

Adyen's AI-augmented decisioning stack runs on a hybrid architecture that the company's engineering organisation has been quieter about than its commerce-infrastructure peers. The risk-decisioning model — the production model that makes the authorisation-time decision on whether to approve, decline, or step up a transaction — is Claude Sonnet 4.6, grounded against the merchant's transaction history, the cardholder's behavioural pattern, and the contextual signals Adyen has accumulated across fifteen years of European acquiring data. The grounding architecture is retrieval-augmented, with the retrieval layer running on Cohere's Embed 4 model and the vector index hosted inside Adyen's own data centres in Amsterdam and Frankfurt. The choice of Cohere over OpenAI's embedding API or Voyage AI's offerings was deliberate, and the engineering organisation has described it on internal architecture documents as a regulatory-residency requirement: Cohere offered, in mid-2025, the only major commercial embedding API with full EU data residency under the Cohere Sovereign deployment model. The decision pre-dated the EU AI Act compliance audit by approximately six months.

The orchestration plane — the layer that sits above the Claude inference and below the merchant-facing risk decisioning API — is built in-house. Adyen's engineering team, led by Roelant Prins as CCO of Engineering, declined to adopt LangChain, LlamaIndex, or any of the off-the-shelf orchestration frameworks that came of age in 2023 and 2024. The argument was operational: at the per-transaction latency budget the acquiring surface operates under — a 90th-percentile budget of 380 milliseconds end-to-end, with the AI-decisioning layer allocated 110 milliseconds of that envelope — the overhead introduced by a generic orchestration framework was unacceptable. The in-house plane, internally named Atlas, makes a single Claude inference call per transaction, with parallel embedding retrieval against Cohere and a parallel pre-computation of the rule-based signals that Adyen's pre-2024 decisioning stack used to evaluate independently. The three signals converge in the in-house orchestrator, which produces the authorisation decision and a structured reasoning artefact that is logged to Adyen's audit infrastructure for compliance review.

The per-decision cost — which Adyen disclosed at the Q1 analyst call as "in the low single-digit eurocents per transaction" — is the metric that the engineering team has spent the bulk of its 2025 optimisation cycle reducing. The original architecture, shipped in November 2024 in a limited beta across Uber's European acquiring volume, carried a per-decision cost of approximately €0.012 per transaction. The current architecture, post the November 2025 rebuild that introduced the parallel-retrieval and reasoning-cache optimisations, runs at approximately €0.003 to €0.004 per transaction depending on the merchant's transaction-mix complexity. The optimisation work has not been published in detail, but the engineering team described two principal levers in a closed-room presentation at Money 20/20 Europe in June 2025: aggressive caching of the reasoning artefact for repeat-cardholder transactions within a defined behavioural envelope, and the introduction of a smaller distilled-model fallback path for transactions whose risk signals the orchestrator can confidently classify before reaching the Sonnet 4.6 layer. The distilled fallback handles approximately 31 per cent of the acquiring volume; the remainder routes to Sonnet 4.6. The split was tuned through Q3 2025 to balance per-decision cost against the false-positive rate.

Uber, Spotify, eBay: the merchant rollout sequence

The 40 per cent of merchant volume that now runs through the AI-augmented decisioning stack is not evenly distributed across Adyen's merchant book. The first three named merchants on the architecture — Uber, Spotify, and eBay — together account for roughly 23 percentage points of the disclosed 40 per cent. Uber went live on the architecture in November 2024 across its European acquiring volume, expanded to its US acquiring volume in March 2025, and now runs effectively 100 per cent of its global Adyen-acquired volume through the AI-augmented decisioning layer. The Uber relationship has been a deliberate strategic anchor for Adyen since 2018, and Uber's CFO Prashanth Mahendra-Rajah is on record at Money 20/20 Europe 2025 describing the Adyen architecture as "the only acquiring relationship where the per-decision economics make AI augmentation defensible at our transaction velocity." The framing matters. Uber's transaction velocity — measured in authorisations per second across its peak hours — is among the highest in the commerce-infrastructure customer book, and the per-decision economics that Mahendra-Rajah cited are the cost-of-AI envelope that Adyen's engineering team had to drive down to make the architecture commercially viable.

Spotify followed in February 2025, with a phased rollout that started with European acquiring volume and extended to North American volume by September 2025. The Spotify deployment introduced a specific feature — recurring-billing pattern recognition — that Adyen's engineering team had to add to the Atlas orchestrator's reasoning context. The subscription-billing surface has different risk signatures from the spot-purchase surface that Uber operates, and the original Sonnet 4.6 grounding prompts had been tuned against the latter. The Spotify deployment forced a prompt rebuild that the team released in March 2025 as a v1.4 architecture revision, and that has since been backported to the Uber surface to handle Uber One subscription billing. The cross-pollination between merchant deployments — where the engineering required to support one merchant's specific use case becomes a generic capability for the broader book — is now the operational pattern Adyen's product team is explicit about in its merchant-onboarding documentation. The merchant council slide deck for May 2026 described the pattern as "merchant-led capability expansion under platform-led architecture."

eBay's deployment, which closed in Q3 2025, was structurally different from the Uber and Spotify deployments because eBay operates as a marketplace rather than a direct merchant. The acquiring volume that flows through eBay-Adyen is split across millions of small sellers, each of whom has a distinct risk profile that the Atlas orchestrator must reason against independently. The eBay deployment forced the engineering team to extend the retrieval layer to support per-seller context windows, which in turn required a rebuild of the Cohere embedding index to handle the volume and dimensionality of the marketplace-merchant graph. The rebuild shipped in October 2025 and is now the foundation that supports Adyen's broader marketplace-acquiring book — including its work with Etsy, Vinted, and the smaller European-marketplace tier. The eBay deployment is the one that the engineering team flags internally as the architectural pivot that made the broader 40 per cent volume number achievable; without the marketplace-context capability, Adyen could not have scaled the architecture beyond the direct-merchant tier.

The remaining 17 percentage points of the disclosed 40 per cent figure are distributed across approximately 280 enterprise merchants in Adyen's mid-to-large segment, with no individual merchant accounting for more than 1 percentage point of the total. The mid-segment rollout has been deliberate. Adyen's commercial team has prioritised merchants whose transaction mix is similar in profile to the named anchor merchants, on the operational logic that the Atlas orchestrator's reasoning context is best calibrated against transaction patterns the architecture has already seen. The next cohort of mid-segment merchants — a tranche of approximately 140 European enterprises that Adyen plans to migrate across the rest of 2026 — will push the cumulative volume figure to a level the company has not yet disclosed but that the merchant-council briefing materials suggest is targeted at 60 per cent by year-end and 75 per cent by mid-2027.

The false-positive reduction: 31 per cent across the deployed book

The metric Adyen has emphasised in its Q1 2026 disclosure is not revenue lift but false-positive reduction. The 31 per cent figure — false-positive declines reduced by 31 per cent across the 40 per cent of merchant volume on the AI-augmented architecture, measured against the pre-deployment baseline of the same merchants — is the number that the company believes describes the true commercial value of the architecture. False positives — transactions that should be approved but are declined by the risk decisioning layer — are the most expensive failure mode in commerce infrastructure because they cost the merchant the gross transaction value, cost the acquirer the interchange-side revenue, and impose a reputational cost on both sides of the relationship. The 31 per cent reduction is a meaningful number against the baseline: Adyen's pre-deployment false-positive rate was approximately 1.4 per cent of attempted transactions on the merchant cohort that has been migrated, and the post-deployment rate is approximately 0.97 per cent. The absolute reduction is 0.43 percentage points across roughly 40 per cent of total volume, which scales against the company's annualised payment volume to a gross transaction value reclaimed that the company has not officially disclosed but that the merchant council slide deck implied is in the low double-digit billions of euros annualised.

The fraud-detection rate — the true-positive rate, where the architecture correctly identifies genuinely fraudulent transactions — has held steady against the pre-deployment baseline. The company has been careful to disclose that the architecture is not improving fraud detection in absolute terms; it is improving false-positive avoidance while holding fraud detection constant. The framing is deliberate. Adyen's compliance organisation has been wary of the alternative narrative — that AI-augmented decisioning is allowing more fraud through in exchange for fewer false positives — and the engineering team has built the architecture explicitly to anchor on the constant-fraud-rate constraint. The constraint was a design parameter in the November 2024 architecture freeze, and the eval methodology that the team runs against every weekly model-refresh A/B holds the constraint as the gate condition: a refresh ships only if it improves false-positive rate while holding fraud rate at or below baseline. The constraint has held for every refresh in 2025 and Q1 2026.

The chargeback-rate implication is the second-order metric that Adyen has been quieter about but that the merchant council slide deck flagged as the next disclosure-ready data point. Across the 40 per cent of merchant volume on the architecture, the chargeback rate is down approximately 18 per cent against the baseline, with the bulk of the reduction concentrated in the friendly-fraud category that the Atlas orchestrator's reasoning context is particularly well-suited to identify. The chargeback reduction has direct merchant-side value because it reduces the dispute-management cost that merchants incur when chargebacks land, and direct acquirer-side value because it reduces Adyen's exposure to chargeback-network fines. The merchant-council materials projected that the chargeback-rate reduction, if maintained across the rollout to 60 per cent of volume by year-end, would produce a net economic benefit to Adyen's merchant base in the high single-digit billions of euros over a calendar-year period.

The reasoning-artefact log that Atlas produces for every transaction is the data foundation that has made the false-positive analysis possible at this level of granularity. Every transaction that flows through the AI-augmented decisioning path generates a structured reasoning artefact that documents the model's inputs, the retrieval context surfaced from the embedding layer, the rule-based-signal evaluation, the final decision, and the confidence-level metadata that the orchestrator generates. The reasoning artefact is retained in Adyen's audit infrastructure for the duration the regulatory framework requires and is the data substrate against which the false-positive analysis is conducted. The substrate has produced operational learning that informs the architecture's iterative refinement: the team has identified specific merchant-category and transaction-type combinations where the architecture's performance is materially above the average, and other combinations where the performance is below the average. The disaggregated performance pattern informs the prompt-engineering work that the team conducts on a continuous basis and is the operational mechanism by which the architecture's false-positive performance has continued to improve across each quarter of the deployment.

The EU AI Act compliance posture: provenance, audit, and the high-risk classification

Adyen's AI-augmented decisioning architecture sits squarely inside the EU AI Act's high-risk category, classified under the Act's enumeration of credit-scoring and creditworthiness-decisioning systems — a classification that the company's regulatory team confirmed in correspondence with the European Commission's AI Office in October 2025. The classification triggers the full set of high-risk system obligations: conformity assessment, technical documentation requirements, transparency obligations to the affected merchants and cardholders, human-oversight requirements, and the mandatory registration with the EU AI database that becomes enforceable in August 2026. Adyen's compliance organisation, led by Alexander Matthey as Chief Compliance Officer, has been preparing for the August 2026 enforcement deadline since the Act's final text was published in June 2024. The preparation has produced an internal compliance framework that the company describes as "ALARA" — As Light Architecture, Reasonable Accountability — and that the regulatory team has shared in unofficial correspondence with peer acquirer compliance teams at Worldpay, Nuvei, and Checkout.com.

The technical-documentation requirement is the operational lever that has shaped the most engineering work. The Act requires that the high-risk system provider — Adyen, in this case — maintain comprehensive documentation of the model architecture, training data, evaluation methodology, performance metrics, and the rationale for design decisions at a level of detail that allows a competent regulator to assess conformity. Adyen's documentation set, which the company has indicated runs to approximately 4,200 pages across the production architecture, has been compiled by a dedicated documentation team that the compliance organisation built out in Q2 2025. The documentation is updated quarterly and is structured as a versioned artefact: every model refresh, every architectural change, and every merchant-specific extension generates a documentation delta that is signed off by the engineering lead and the compliance lead before deployment. The discipline is not optional under the Act; the discipline is the conformity-assessment foundation.

The human-oversight requirement is the one that has caused the most product-team discussion. The Act requires that high-risk systems be designed such that a human operator can intervene meaningfully in the decisioning process, with the system providing the operator the information needed to make an informed override. The acquiring-side risk decisioning surface, where the decision happens in 110 milliseconds at the network-authorisation layer, is not amenable to per-transaction human oversight in real time. Adyen's regulatory-engineering team has resolved the tension through a structured-review architecture: the Atlas orchestrator's reasoning artefact is logged at decision time and surfaced to a designated risk-operations team for post-hoc review at a defined sampling rate — currently 0.5 per cent of declined transactions, which the team adjusts upward to 5 per cent for transactions over €5,000 and merchant categories the compliance organisation flags as elevated-risk. The structured-review architecture has been discussed in writing with the Dutch AFM, which is Adyen's domestic regulator for the EU AI Act conformity assessment, and the AFM's preliminary view — though not yet binding — is that the architecture satisfies the human-oversight obligation provided the sampling rate is maintained and the reasoning artefact is sufficient for an informed review.

The EU AI Act compliance posture is one of the structural arguments Adyen's commercial team makes against Stripe and PayPal Braintree in enterprise procurement. Stripe operates a US-headquartered architecture with EU data residency through AWS Frankfurt, but its compliance documentation is calibrated to the US regulatory environment with a layered EU adaptation; Adyen operates a fully EU-resident architecture with compliance documentation built from the EU AI Act outward. The distinction matters in procurement conversations with European enterprises whose own compliance organisations are wary of cross-jurisdictional architecture choices — particularly in financial services, public-sector adjacent commerce, and the regulated-payments segments where the EU AI Act's high-risk classification carries direct accountability for the merchant as well as the acquirer. The commercial advantage is not absolute, but it is measurable in the enterprise win rate Adyen reported in its Q1 2026 update: 73 per cent of EU-headquartered enterprise procurement deals over €10M annual volume in Q1 2026 closed in Adyen's favour, against an industry-baseline win rate the company estimated at 41 per cent.

The competitive implication: Stripe, PayPal Braintree, and the depth metric

The Q1 2026 disclosure repositions Adyen against Stripe in a way that the trade press has read partially. Stripe crossed $200B in annualised payment volume in the same May 2026 window — a figure that exceeds Adyen's reported annualised volume of approximately €1.1 trillion when adjusted for currency, but that obscures the depth-of-deployment differential. Adyen's 40 per cent figure refers to acquiring-side risk decisioning that Stripe's framing of its AI-fraud platform does not formally include, because Stripe operates the acquiring relationship for a smaller proportion of its merchant book — much of its volume flows through the Direct Charge model where the merchant's bank is the issuer's acquirer, and Stripe's risk surface is the merchant-side fraud-detection layer rather than the network-authorisation layer. The two architectures are not directly comparable, and the depth metric Adyen has chosen to disclose — 40 per cent of merchant volume on AI-augmented authorisation-time decisioning — is a metric that Stripe cannot easily match in the same units. The framing is competitive judgement on Adyen's part.

PayPal Braintree's posture in 2026 has been markedly quieter than its competitors expected. The company restructured under Alex Chriss in 2023 and 2024, and the merchant-acquiring segment that Braintree operates was reorganised in early 2025 into a more focused commerce-platform unit. The AI investment cadence inside Braintree has been slower than Adyen's or Stripe's, and the company has not yet disclosed an equivalent merchant-volume-on-AI-decisioning metric. The trade press read — corroborated by procurement-team sources at three Fortune 500 retailers who have evaluated all three acquirers in the last twelve months — is that PayPal Braintree's AI capability is approximately twelve to eighteen months behind Adyen's on the production-decisioning surface and approximately six to twelve months behind Stripe's on the merchant-facing advisory surface. The gap is structural rather than tactical; closing it requires a sustained capital commitment that PayPal's broader restructuring has not yet authorised in full.

The mid-tier acquirers — Worldpay, Nuvei, Checkout.com, Global Payments — are the segment of the market that Adyen's disclosure most directly pressures. None of the four has disclosed a comparable AI-augmented decisioning metric, and the engineering investment required to reach Adyen's or Stripe's deployment depth is, by the merchant council's own assessment, beyond the capital headroom of any of the four standalone. The procurement pressure will work through the next twelve months in the form of large-merchant migration: enterprises whose acquiring relationships were grandfathered into the mid-tier acquirers on legacy-pricing terms will face renewal conversations in which Adyen and Stripe can present demonstrable per-transaction risk-economics that the mid-tier cannot match. The structural outcome is the consolidation that commerce-infrastructure analysts have been predicting since 2024, accelerated by an architecture differential that the mid-tier cannot close through commercial terms alone.

The consolidation pressure has begun to materialise in observable M&A activity across the mid-tier. Worldpay's acquisition by GTCR in 2024 was structured around an explicit thesis of AI-infrastructure investment that the company has been executing against, with approximately $420M of post-acquisition capital deployed against the AI-decisioning rebuild that the company is targeting for general availability in Q4 2026. Nuvei's privatisation by Advent International, also completed in 2024, included a similar capital-deployment thesis with approximately $310M authorised for the equivalent rebuild. Checkout.com, which has remained privately held under its founder Guillaume Pousaz, has been the slowest of the four to authorise the capital commitment, and the company's competitive position in the European mid-market is the segment of the market that Adyen's merchant-acquisition team has been targeting most aggressively in 2026. The structural projection — that two of the four mid-tier acquirers will close the AI-architecture gap by mid-2027, and that the other two will either be acquired or will retreat into specialised verticals — is the merchant council's working hypothesis for the rest of the decade.

The engineering investment: how Adyen built Atlas without outside capital

The engineering investment behind the 40 per cent disclosure has been the company's largest single technical commitment since the initial international expansion that built the company's multi-jurisdictional acquiring infrastructure in the 2012-to-2016 period. Adyen's AI engineering organisation has grown from approximately 110 named-AI-role engineers at the start of 2024 to approximately 430 by the end of Q1 2026, with the growth concentrated in three specific functional categories. The first is the production-decisioning engineering cohort that operates the Atlas orchestration plane and the per-transaction inference path; this cohort grew from approximately 35 to approximately 160 across the period. The second is the model-evaluation engineering cohort that runs the weekly A/B comparison rig and the broader eval methodology against the merchant-cohort production traffic; this cohort grew from approximately 20 to approximately 90. The third is the merchant-integration engineering cohort that operates the inbound API surfaces that route merchant transactions into the AI-augmented decisioning path; this cohort grew from approximately 25 to approximately 100. The remaining residual growth is distributed across research-and-applied-science roles and the platform-engineering roles that support the underlying infrastructure.

The capital investment has been funded entirely from the company's existing operating cash flow, without outside fundraising and without the dilutive equity issuance that has financed the equivalent buildouts at smaller competitors. The investment line — disclosed in the company's investor-day materials in February 2026 as a cumulative €640M across 2024 and 2025, with an additional €380M authorised for 2026 — represents approximately 11 per cent of the company's operating cash flow across the period. The proportion is operationally significant but is well within the company's historic capital-allocation discipline, and the investor-day materials emphasised that the AI investment has not displaced the company's broader infrastructure investments in payments-network connectivity, point-of-sale terminal certification, and the regulated-money-movement licences that the company has been expanding internationally. The AI investment is a parallel allocation rather than a substitution allocation, and the company's net cash generation has continued to grow across the period despite the investment commitment.

The engineering-culture choices that Adyen has made around the AI rebuild are the operational mechanism by which the company has avoided the talent-acquisition challenges that have constrained the equivalent rebuilds at competitors. Adyen has been explicit, in internal hiring materials and in external recruiting communications, that the company prioritises engineers with deep payments-infrastructure context over engineers with deep AI-modelling context. The framing is deliberate: the company's view is that the per-transaction latency budget, the operational reliability requirements, and the regulatory-compliance discipline of the acquiring-side decisioning surface are harder to teach than the AI-modelling skills, and that the engineering team's productivity depends more on the payments context than on the AI specialisation. The hiring discipline has produced an engineering organisation that is structurally different from the equivalent organisations at Stripe and PayPal Braintree, both of which have weighted their hiring toward the AI-specialisation side of the trade-off. The structural difference will be tested as the architecture matures and as the operational reliability metrics become the differentiating factor across the three companies' deployments.

What to watch

Adyen's 40 per cent disclosure is the structural milestone for 2026, but the next twelve months will produce four data points that resolve the competitive position the disclosure has staked.

• Whether Adyen's targeted 60 per cent year-end and 75 per cent mid-2027 deployment depth holds against the operational complexity of the mid-segment merchant tier; the named-anchor rollouts (Uber, Spotify, eBay) were technically tractable because the merchants had the engineering capacity to integrate Atlas-native decisioning artefacts into their own risk operations, but the mid-segment cohort is smaller and less engineering-rich, and the rollout cadence may be paced by merchant-side integration capacity rather than Adyen-side engineering capacity.
• Whether the EU AI Act enforcement that begins in August 2026 produces a regulator-validation outcome for Adyen's ALARA framework that the company can use as a procurement asset; the AFM's preliminary view on the human-oversight architecture is supportive but not binding, and the formal conformity-assessment outcome will be the first published case of a high-risk financial-services AI system clearing the Act's full obligations.
• Whether Stripe matches the depth-of-deployment disclosure in its Q3 2026 update; the merchant-volume-on-AI-decisioning metric is the comparison frame Adyen has chosen, and Stripe's response — either matching the disclosure or arguing for a different metric — will set the terms of the competitive narrative for the rest of the year.
• Whether the per-decision cost continues its trajectory below €0.003 by year-end; the engineering team's roadmap includes a second-generation distilled-fallback model that is targeted to handle 45 per cent of acquiring volume against the current 31 per cent, and the cost curve below €0.003 would unlock additional merchant-segment economics that the current pricing tier does not cover.
• Whether the false-positive-reduction metric holds across the next rollout cohort; the 31 per cent figure is anchored on the merchant cohort whose transaction patterns the Atlas orchestrator has been calibrated against most heavily, and the mid-segment cohort may produce a different reduction figure that will calibrate the company's pricing and capacity-planning for the remainder of the rollout.

Frequently asked

What does the 40 per cent of merchant volume figure actually measure?

The figure measures the proportion of Adyen's acquiring-side merchant volume that flows through the AI-augmented decisioning architecture at the network-authorisation layer. It excludes pass-through processing arrangements where the merchant operates its own risk model and Adyen acts as the technical acquirer without the risk-decisioning surface. The figure is a depth-of-deployment metric, not a count of enrolled merchants — the disclosed 40 per cent is concentrated in the named-anchor merchants (Uber, Spotify, eBay) plus approximately 280 mid-to-large enterprise merchants who together comprise the deployment book.

Why did Adyen choose Cohere for embeddings rather than OpenAI or Voyage AI?

The choice was a regulatory-residency requirement. In mid-2025, when the architecture's embedding layer was being finalised, Cohere offered the only major commercial embedding API with full EU data residency under its Sovereign deployment model. OpenAI's embedding API does not offer the equivalent residency posture, and Voyage AI's enterprise offering at the time did not support the throughput Adyen required for production acquiring-side decisioning. The decision was made approximately six months before the EU AI Act compliance audit and was framed by Adyen's regulatory team as a residency-first architectural choice rather than a pure performance choice. The vector index is hosted inside Adyen's own data centres in Amsterdam and Frankfurt, not on Cohere's infrastructure.

What is Atlas, and why did Adyen build it in-house instead of using LangChain or LlamaIndex?

Atlas is Adyen's in-house orchestration plane that sits above the Claude Sonnet 4.6 inference call and below the merchant-facing risk decisioning API. The decision to build it in-house was driven by the latency budget on the acquiring surface: at a 90th-percentile end-to-end budget of 380 milliseconds and an AI-decisioning allocation of 110 milliseconds, the overhead introduced by a generic orchestration framework was operationally unacceptable. Atlas makes a single Claude inference call per transaction, runs parallel Cohere embedding retrieval and rule-based-signal pre-computation, and produces both the authorisation decision and a structured reasoning artefact that flows into the audit infrastructure. The in-house architecture also gives the engineering team control over the per-decision cost curve, which has been a primary optimisation lever since the original architecture shipped.

How does the 31 per cent false-positive reduction translate to merchant-side economic value?

Adyen has not disclosed the absolute merchant-side dollar value of the reduction, but the maths is calculable from the disclosed figures. The pre-deployment false-positive rate across the migrated merchant cohort was approximately 1.4 per cent of attempted transactions, with a post-deployment rate of approximately 0.97 per cent — a 0.43-percentage-point absolute reduction across roughly 40 per cent of total volume. Scaled against the company's reported annualised payment volume of approximately €1.1 trillion, the gross transaction value reclaimed for merchants from previously-declined-now-approved transactions is in the low double-digit billions of euros on an annualised basis. The merchant captures the bulk of the value through restored gross transaction volume; Adyen captures the secondary value through interchange-side margin on the reclaimed volume.

What is ALARA, and how does it satisfy the EU AI Act's high-risk system obligations?

ALARA — As Light Architecture, Reasonable Accountability — is Adyen's internal compliance framework for the EU AI Act's high-risk system classification, which the architecture sits inside as a credit-scoring and creditworthiness-decisioning system. The framework addresses the Act's five principal obligations: conformity assessment (handled through a dedicated documentation team that maintains approximately 4,200 pages of versioned technical documentation), transparency to affected parties (handled through merchant-side disclosure language and cardholder-side declined-transaction reasoning that is exposed to issuers on request), human oversight (handled through a structured-review architecture with a 0.5 per cent baseline sampling rate, adjustable upward for elevated-risk transactions), data quality (handled through the in-house orchestration plane's input validation and the Cohere-residency-resident embedding layer), and the August 2026 AI database registration. The Dutch AFM, Adyen's domestic regulator for the conformity assessment, has indicated preliminary support for the framework, with the formal binding outcome expected in Q3 2026.

How does Adyen's deployment compare to Stripe's AI-fraud platform and PayPal Braintree's equivalent?

The three architectures are not directly comparable in the same units. Adyen's 40 per cent figure refers to acquiring-side risk decisioning at the network-authorisation layer, which is the surface Adyen operates for the bulk of its merchant book. Stripe's AI-fraud platform operates the merchant-side fraud-detection layer for its Direct Charge merchants and a smaller acquiring-side surface for the merchants whose acquiring Stripe directly handles, and Stripe's disclosed framing has emphasised platform shipment rather than depth of deployment. PayPal Braintree has not disclosed a comparable depth-of-deployment metric, and trade-press triangulation against procurement-team sources suggests Braintree's AI capability is approximately twelve to eighteen months behind Adyen's on the production-decisioning surface. The three architectures occupy different positions in the commerce-infrastructure stack and the competitive comparison turns on which surface the buyer is prioritising in their procurement decision.