Pre-Empted Intent: Anticipatory Agent Behavior and the Compression of Brand Demand

The closing window

A purchase decision, in the traditional account of consumer behavior, is the product of an extended sequence: a need is recognized, a consideration set is constructed, alternatives are evaluated, and a choice is made. The behavioral economics literature on consideration sets, formalized by Hauser and Wernerfelt in their foundational 1990 Journal of Consumer Research paper, treats this interval as the operative window of brand influence (Hauser & Wernerfelt, 1990). The window is where awareness becomes preference, where preference is tested against alternatives, and where the marketing function historically allocated the majority of its budget. Whatever displaces that interval displaces the function that operates within it.

The behavioral and infrastructural conditions for that displacement are now in place. Generative AI assistants have moved from response-mode systems, which respond to explicit user prompts, to anticipatory-mode systems, which form intent representations before the user articulates them and act upon those representations under delegated authority. OpenAI’s persistent Memory feature, rolled out to all paid users in April 2024 and expanded to a cross-conversation reasoning capability in April 2025, allows ChatGPT to retain and apply context across sessions without explicit recall (OpenAI, 2025). Google’s January 2026 announcement of Gemini Live with persistent personal context extends the same architecture into a real-time conversational surface integrated with calendar, email, and commerce intent signals (Google, 2026). ChatGPT Tasks, launched in early 2025, allows the system to schedule autonomous execution of recurring user objectives without per-execution confirmation (OpenAI, 2025). The shift from request-response to ambient-execution is not a future aspiration. It has shipped.

The shift from request-response to ambient-execution is not a future aspiration. It has shipped.

The technical literature on anticipatory computing, established by Pejovic and Musolesi’s 2015 ACM Computing Surveys paper, defines the field as systems that infer user state, predict near-term intent, and act in advance of explicit instruction (Pejovic & Musolesi, 2015). The anticipatory computing program for the prior decade was hampered by the brittleness of intent inference under sparse signals. Large language models, by their nature, dissolve that brittleness. They ingest unstructured personal context, including conversation history, uploaded files, calendar and email content where authorized, and ambient platform signals, and they produce coherent intent representations from inputs that prior anticipatory systems treated as noise. The field has, in effect, found its substrate.

The behavioral mechanism of pre-emption

The psychological literature establishes a robust prediction: consumers will accept agent-anticipated decisions at high rates provided three conditions are met. The first is perceived personalization, the sense that the agent’s recommendation reflects the user’s actual circumstances rather than a generic default. The second is satisfactory prior outcomes, which, through the automation-bias mechanism formalized in Goddard, Roudsari and Wyatt’s widely cited 2012 JAMIA paper, lower the cognitive effort the user is willing to invest in evaluating the agent’s subsequent choices (Goddard, Roudsari & Wyatt, 2012). The third is the absence of friction in the consummating step, which removes the moment at which the user might revisit the choice. The platform-native checkout protocols deployed by the major AI providers in January 2026 satisfy this condition by construction.

The relevant behavioral science extends beyond the technology adoption literature. Samuelson and Zeckhauser’s 1988 Journal of Risk and Uncertainty paper established status quo bias as a robust regularity: when a default is presented, decision-makers accept it at rates that exceed any rational account of the choice architecture (Samuelson & Zeckhauser, 1988). The principle has been replicated across organ donation, retirement savings, energy provider selection, and online subscription retention, with default-acceptance rates in many domains exceeding 80% (Johnson & Goldstein, 2003). When an anticipatory agent surfaces a single recommendation framed as the natural next step, the choice architecture replicates the default condition that the status quo literature predicts will be accepted without active deliberation.

Iyengar and Lepper’s 2000 Journal of Personality and Social Psychology paper on choice overload, replicated extensively in the two decades since, establishes that the addition of options beyond a moderate range systematically reduces the probability of any purchase being completed and, conditional on completion, reduces satisfaction with the purchase made (Iyengar & Lepper, 2000). A 2015 meta-analysis of 99 observations from 53 conditions found that the choice-overload effect, although heterogeneous in magnitude, is reliably present under predictable conditions: time pressure, complexity, and uncertainty about preference structures (Chernev, Böckenholt & Goodman, 2015). Anticipatory agents act on precisely those conditions. They surface a curated recommendation under the user’s presumed time pressure, in the presence of high specification complexity, and at a moment of uncertain preference, and they do so with the implicit credibility conferred by accumulated automation-bias trust.

The convergence of these mechanisms produces a behavioral equilibrium that is unfavorable to the brand attempting to enter the consideration set after the agent has already constructed it. Hauser and Wernerfelt’s original framework treats consideration set construction as a costly cognitive operation that the consumer undertakes by retrieving brands from memory. When the agent performs this operation on the consumer’s behalf, drawing on training and retrieval rather than the consumer’s own recall, the brand set that reaches the moment of decision is determined by a process to which the brand has no direct access and over which the consumer exercises diminishing oversight.

The visible signal in deployment

The behavioral signal is no longer purely speculative. Several of the major platforms now disclose operational telemetry on their consumer-facing AI assistants, and the publicly disclosed deployment record across the leading retailers and payment processors describes a category whose volume has grown rapidly rather than gradually. Specific figures vary across reporting periods and vendors, and a research-grade synthesis of the data requires source-level verification that is beyond the scope of this paper. The qualitative pattern is consistent across the available reports: agent-initiated and agent-influenced transactions are growing at rates that exceed those of any adjacent commerce category, and the growth is concentrated among consumers who delegate repeat decisions after a small number of satisfactory prior interactions.

Three categories of deployment make the pattern legible. The first is the consumer-facing shopping assistant integrated into a retail brand or payments platform, of which Klarna’s AI Assistant is the most operationally instrumented public example. The second is the anticipatory recommendation surface launched by a marketplace operator, of which Etsy’s Gift Mode is the clearest case in which an LLM constructs a candidate gift set from a thirty-second user prompt. The third is the replenishment-and-recommendation layer embedded in a major retailer’s digital experience, of which Walmart’s Sparky and Amazon’s Rufus are the most visible deployments. The implication that follows from the spread of these surfaces is that anticipatory purchasing is escaping the high-frequency low-consideration corner of consumer behavior into adjacent categories where the marketing function has historically expected sustained brand-influence windows.

The anticipatory commerce architecture extends into B2B contexts where procurement software has begun to surface pre-formed vendor selections to human approvers. The behavioral pattern identified in the consumer literature on automation bias and choice architecture appears to extend into organizational decision contexts: the relevant agent evaluates vendors against structured criteria, proposes a recommendation, and the human approver decides whether to accept, modify, or reject it. The fraction of such proposals accepted without modification is the operative variable, and the structural conditions under which it rises (perceived personalization, satisfactory prior outcomes, frictionless consummation) replicate inside procurement workflows for the same reasons they replicate in consumer shopping.

The fraction of agent-proposed selections accepted without modification is the operative variable, and emerging telemetry suggests it is high.

The technical infrastructure of anticipation

The technical preconditions for pre-emption are persistent memory, ambient signal collection, and proactive scheduling. Each has been shipped in production by the major AI platforms in the eighteen months preceding this paper. Persistent memory, the capacity to retain stated preferences, recurring contexts, and inferred goals across sessions and apply them without explicit recall, is now standard across the leading consumer assistants, and the 2025 rollouts dissolved the prior boundary between in-session and cross-session context (OpenAI, 2025). Ambient signal collection, the integration of calendar, email, location, and platform-specific telemetry into the assistant’s working memory, is now active under user authorization across Gemini Live and the parallel deployments inside Meta’s messaging surfaces. Proactive scheduling, the capacity for an agent to initiate action on behalf of the user without per-execution confirmation, was first introduced as a consumer feature with ChatGPT Tasks in early 2025 and now appears across productivity, shopping, and replenishment surfaces. The relevant observation is that the three preconditions, treated as separate research programs through the prior decade, are now operating together inside production systems available to several hundred million users.

The retrieval-augmented generation architecture underlying these systems shapes the brand-visibility implications. Lewis et al.’s foundational 2020 paper on retrieval-augmented generation established that RAG systems combine parametric knowledge, encoded in the model weights through training, with non-parametric knowledge retrieved at inference time from external sources (Lewis et al., 2020). The brand presence required for citation in an anticipatory agent’s recommendation is therefore a function of two distinct conditions: inclusion in the model’s training corpus with sufficient coherence to surface as a salient candidate, and presence in the retrieval-time sources the agent consults under user authorization. Brands satisfying only one of these conditions are surfaced inconsistently, while brands satisfying neither are absent from the consideration set the agent constructs.

The compression of the brand-influence surface

The traditional marketing funnel, codified in St. Elmo Lewis’s AIDA framework in 1898 and developed across a century of marketing science, assumes a sequential consumer journey in which awareness precedes consideration, consideration precedes evaluation, and evaluation precedes conversion. The expectation that the consumer will spend cognitive effort at each stage is the foundation on which media planning, brand positioning, and creative strategy have been built. Anticipatory agents collapse this sequence into a single moment of agent evaluation. Awareness becomes inclusion in the agent’s knowledge representation. Consideration becomes retrieval at the moment of inferred need. Evaluation becomes the agent’s comparison of structured signals across the candidate set. Conversion becomes the consummating step, which the frictionless agent-checkout architecture executes without further human deliberation.

The compression has measurable consequences for the timing of brand investment. Edelman and Singer’s 2015 Harvard Business Review analysis of consumer decision journeys found that, in high-consideration categories, brand-influence opportunities were distributed across an extended exploration phase that lasted weeks or months (Edelman & Singer, 2015). The McKinsey Consumer Decision Journey research, which the Edelman paper extended, identified the “loyalty loop” in which satisfied prior buyers re-entered the consideration set preferentially through brand recall (McKinsey, 2009). Anticipatory agent behavior preserves the loyalty loop in form but inverts its mechanism: the satisfied buyer’s preference is encoded in the agent’s memory rather than the consumer’s, and the loop closes through the agent’s default selection rather than the consumer’s active recall. Brands not yet present in the loop are correspondingly disadvantaged because the recall mechanism on which they would historically depend has been mediated by an entity to which they have no direct relationship.

The structural change is best described as a compression of the journey into a single moment of agent evaluation. The implication for brand investment is that the surface on which influence operates shifts from sustained presence in the consumer’s media environment to verifiable presence in the agent’s candidate construction process. The two surfaces overlap in some respects, since both reward consistent multi-channel signal, but they diverge in the determinants of inclusion. Media buying optimized for human attention does not produce the structured corroboration patterns that retrieval systems weight at inference time, and the inverse holds.

What replaces the funnel

The mental model that replaces awareness, consideration, and conversion under anticipatory agent behavior is not a sequence at all. It is a state. Brand influence under compression is the property of being present in the agent’s working representation of the consumer’s plausible needs at the moment those needs are inferred. The state has internal structure: an entity-level identity sufficiently coherent that the agent’s training corpus and retrieval sources surface the brand as a salient candidate, a record of satisfactory consumer outcomes accumulated in the agent’s memory through prior delegated transactions, and a multi-channel signal density that the agent’s confidence calibration treats as evidence the brand can be recommended without elevated risk.

The traditional funnel was a flow. The new shape is a position.

The state is durable in ways the funnel was not. A consumer who moves through the awareness-to-conversion sequence completes a journey; the journey ends, and the brand’s influence on subsequent journeys depends on the consumer’s subsequent recall. A brand that occupies the agent’s working representation continues to occupy it across many consumers and many transactions, and the agent’s confidence in the brand grows monotonically with each satisfactory outcome it observes. The traditional funnel was a flow. The new shape is a position. Holding it pays compounding returns; losing it requires a structurally more expensive re-entry than the position required to establish.

The implication for media strategy is direct. Investments that produce attention without producing the conditions for agent memory are increasingly unproductive. Investments that produce structured corroboration, entity-level coherence, and a verifiable record of satisfactory outcomes acquire compounding value, since they install the brand in a state the agent will continue to draw upon long after the investment has been made. The shift is not that media is faster. The shift is that the artifact of media is structurally different: a position in agent memory rather than a residue in human memory.

Conclusion: what this paper does not resolve

The argument advanced here is that anticipatory agent behavior compresses the deliberation window in which brand influence has historically operated, that the compression replaces a sequenced funnel with a held position in agent memory, and that the mechanism is supported by a behavioral literature spanning consideration set theory, status quo bias, choice overload, and automation bias. Three questions remain open, and Signyl Research expects to take them up in subsequent work.

The first is the rate at which the equilibrium forms. The behavioral mechanisms are well-characterized at the individual level, but the rate at which they compound across a population to produce a stable agent-mediated default is an empirical question the available longitudinal data do not yet answer. A population-scale measurement framework will require either platform cooperation or a panel-based instrument neither of which presently exists at the appropriate scope.

The second is whether agent memory remains stable across model upgrades. Anticipatory agents derive accuracy from the persistence of their representations across interactions. When the underlying model is replaced, which happens at non-trivial frequency at this stage of the technology, what survives in the agent’s effective memory is partly a product of the platform’s data engineering choices. The strategic implication of memory volatility for brand investment depends on a property of the system currently set by platform operators rather than by market participants. A serious account of brand strategy under anticipatory agents requires either a verifiable commitment from platforms regarding memory persistence or an independent method for measuring it across model transitions.

The third is what brand and marketing organizations should do during the transition itself. Existing budgets are committed to channels optimized for human attention; the infrastructure required for agent memory inclusion is not yet costed against the measurement frameworks that justify the human-attention investments; and the brands acting first in the absence of those frameworks bear the cost of constructing them. The reallocation decision is therefore a decision under measurement uncertainty, and the mechanisms by which firms make such decisions, including the role of incumbent measurement vendors, the formation of new attribution standards, and the political economy of the marketing team within the broader firm, are the next-order subjects of study.

The compression itself is empirically observable. The path through it is not yet charted.