Following comparison sites and aggregators, conversational interfaces based on Large Language Models (LLMs) are becoming a primary entry point for financial decisions. The risk for banks is specific: to continue designing and backing products, but to lose control over the interface, the recommendation, and ultimately, the subscription. This is the “dark kitchen” scenario applied to finance, producing without being seen. This movement is anything but speculative. Already in 2024, 23% of consumers were using generative AI for financial tasks<\/strong> at least once a month,(1) and agents could become the preferred channel for banking interactions within the next three to five years. The question is therefore no longer whether these interfaces will position themselves between banks and their customers, but how financial institutions intend to remain visible within them.<\/p>\n Historically, banks controlled the end-to-end journey through their own channels: branches, websites, and apps. Digitalization had initiated an initial disintermediation by making comparison easier and reducing switching costs, without threatening the centrality of brand ecosystems.<\/p>\n Conversational interfaces are shifting the very nature of this dynamic. Customers no longer search for a specific product; instead, they express a desired outcome (optimizing savings, financing a project) and delegate the research, selection, and execution to an agent. These agents now know how to plan, memorize, orchestrate, and execute multi-step financial journeys autonomously. Above all, agentic payment infrastructure is already rolling out: OpenAI and Stripe launched the Agentic Commerce Protocol in late 2025<\/strong> to enable purchases directly within ChatGPT;(2) Visa, Amazon, and Google have since launched their own agent-initiated transaction rails. In-interface subscription is already a live capability in production in certain countries.<\/p>\n These interfaces are no longer limited to transaction history. They capture intent, projects, and trade-offs, serving as an “emotional CRM” that banks have never possessed. As these flows intensify, they are establishing themselves as the central aggregator of the relationship, serving simultaneously as the point of understanding the need and directing the user toward the offer.<\/p>\n The risk, therefore, is shifting toward a “dark kitchen” model: a growing decoupling of production and distribution. Banks continue to design products and carry the associated risk and compliance, but they lose the point of contact. Four consequences stem from this:<\/p>\n Loss of visibility.<\/strong> Customers no longer interact with their bank and may, over time, ignore its role in the value chain. The brand becomes a background supplier.<\/p>\n Commoditization.<\/strong> Exposed to constant comparison and reformatted by the interface, offers become interchangeable. The filter is brutal: recommendation models generally mention only one to three brands per query.(3) Existing or disappearing comes down to a minuscule space.<\/p>\n Weakened relationships.<\/strong> More fragmented and transactional, the relationship loses its depth. Customer loyalty and cross-selling become more complicated, while volatility spikes: in an agent-driven environment, the switching cost tends toward zero.<\/p>\n Pressure on margins.<\/strong> Price transparency and the interface\u2019s recommendation power intensify competition. And the bias does not work in favor of established players: today, LLMs tend to recommend neobanks and digital-native lenders rather than traditional institutions. For a legacy player, the risk is not just being disintermediated, but being systematically sidelined by the machine.<\/p>\n This scenario is by no means far off. As adoption spreads and agents gain autonomy, it becomes urgent to initiate the transformations that will keep banks within the agentic value chain today.<\/p>\n Faced with this evolution, three stances are possible: wait and see; accept the new interface layer and retreat into a product-provider role; or fight to retain the customer relationship through proprietary platforms. The first two lead straight to the “dark kitchen.” Only the third preserves a position in the value chain, and it requires a transformation of architectures, journeys, and relational models, structured around three battles:<\/p>\n To exist.<\/strong> As conversational interfaces become entry points, the stakes are no longer about proprietary channels but about the discoverability of the offer within these environments: this is Generative Engine Optimization (GEO)<\/strong>. In LLM responses, visibility relies heavily on third-party sources (press, reviews, analyses) rather than the brand’s own assets. Existing requires expanding one’s informational footprint far beyond one’s website. The parallel with 2000s SEO is direct: companies that invested early in search optimization built lasting structural advantages. The same window is closing today.<\/p>\n To be operable.<\/strong> In a world of automated recommendations, the interface favors offers that are the simplest to understand and execute. The advantage goes to “agent-ready” products<\/strong>: interoperable, exposed via open architectures, and designed for frictionless subscription. This requires clear, reliable, “machine-readable” services. Beyond visibility, the ability to be distributed without friction becomes a decisive factor.<\/p>\n To be chosen.<\/strong> When offers become comparable and technically integrated, differentiation shifts toward the brand as perceived by the machine. LLMs overrepresent the most cited entities, following a “big brand bias” logic: for a purchase recommendation, certain models prove to be up to 18 times more likely to suggest an already dominant brand<\/strong>.(4) A feedback loop sets in: the model suggests, the user clicks, and dominance is reinforced. Players absent from editorial corpora and review platforms are mechanically sidelined. Building an active presence in these environments and integrating one’s journeys directly into them becomes the decisive lever for retaining the customer relationship.<\/p>\n Agentic commerce is not just evolving the customer relationship; it is shifting its fundamental balances. The center of gravity is sliding toward those who control access, selection, and the prioritization of offers. Banks capable of reintegrating themselves into this market architecture will strengthen their role. The others will become invisible producers, kept away from the point of contact and the place where value is created. The choice is being made now. In three to five years, it will already be made.<\/p>\n [1] McKinsey, \u201cHow Gen AI Agents Threaten Retail Banks\u2019 Customer Relationships,\u201d May 6, 2026. Adoption data: 23% of consumers reported using generative AI for financial tasks at least once a month (2024).<\/em><\/p>\n [2] OpenAI & Stripe, Agentic Commerce Protocol and \u201cInstant Checkout,\u201d Sept. 2025. Competing rails: Visa Intelligent Commerce (Apr. 2026), Visa-OpenAI partnership (June 2026), Amazon \u201cBuy for Me\u201d-AWS Agentic Shopping Assistant, Google Universal Commerce Protocol.<\/em><\/p>\n [3] Estimate from generative optimization practitioners (Searchable, C. Donnelly, 2026); mainstream recommendation models generally cite only one to three brands per query.<\/em><\/p>\n [4] Measurement of “entity bias” reported by L.-M. Lorin (Apr. 2026): depending on the models, the propensity to recommend an already dominant brand reaches up to approximately 18 times (smartphones). Indicative order of magnitude, to be cross-referenced with the primary study.<\/em><\/p>\n<\/div><\/div><\/div><\/div><\/div>The “Dark Kitchen” Risk Applied to the Banking Sector<\/h2>\n
Gaining Control: To Exist, To Be Operable, To Be Chosen<\/h2>\n
\nNotes:<\/h2>\n
![](\"https:\/\/www.artefact.com\/\/wp-content\/uploads\/2026\/06\/images.png\")
<\/a><\/span><\/div>