The Agentic Buyer and the End of Dark Patterns

The pricing page is not the place where a software company tells you what its product costs. It is the place where a software company attempts to alter your decision-making.

This is not a cynical reading. It is a structural one. The pricing page is the final stage of the marketing funnel — the point where the company's goal (maximizing revenue) and the customer's goal (paying the right amount for what they need) are at maximum tension. The design of the pricing page is, in essentially every company that has thought carefully about it, an exercise in reducing the customer's ability to act in their own interest.

A human buyer feels anchored by a high number placed to their left, socially pressured by a "Most Popular" badge, and the urgency of a countdown timer. An AI procurement agent evaluating the same page does not have those particular feelings, because it does not process the page the way a human eye and nervous system do — and that difference is going to restructure B2B software pricing. Not incrementally. Structurally. But the agent is not simply immune. It has a different, and in places surprisingly similar, set of weaknesses, and §4 sets out why that distinction matters most.

The visual tricks miss. What replaces them is the open question this paper takes most seriously.

1. What this correction rests on, and what would undo it

What it argues from. A peer-reviewed, large-scale empirical study of how common manipulative interface patterns actually are on commercial websites.¹ The US Federal Trade Commission's own staff report on the same subject, and its largest enforcement action to date against exactly the subscription-cancellation pattern this paper describes.²³ Foundational, independently replicated behavioural-economics literature on anchoring, the decoy effect, and social proof — not asserted from memory, but checked against each paper's own publication record.⁴⁵⁶ Two recent, directly-read studies — one peer-reviewed, appearing in the Findings of the Association for Computational Linguistics: EMNLP 2025 — that test whether large language models exhibit the same biases in simulated negotiation and decision tasks.⁷⁸ Primary announcements and documentation, fetched directly from the companies involved, for the agentic-commerce payment infrastructure (Visa, Mastercard, OpenAI/Stripe, Google) that already exists as of this writing.^9101112 And a Forrester-reported 2026 buyer survey, named here as a press-reported industry figure rather than an independently audited rate.¹³

What it does not argue from. No study was found that tests dark-pattern susceptibility in LLM-based procurement agents specifically, acting as a buyer's delegate in a live B2B SaaS purchase flow, as distinct from the negotiation and decision-task experiments in §4 — the single largest gap this paper's central claim rests on, named directly rather than papered over. No independent measurement exists of how many SaaS vendors have changed pricing-page design in response to agentic buyers; the behaviour this paper predicts is still in its first year.

What it does not claim. It does not claim AI agents are immune to manipulation; §4 shows that claim is false for a specific, numeric channel of attack, stated plainly rather than softened. It does not claim every dark pattern will disappear from software pricing; consumer-facing commerce, where the buyer is a human for the foreseeable future, is explicitly out of scope. It does not claim the pricing-API infrastructure described in §6 was built because of agentic buyers specifically; cloud pricing APIs predate the current wave of AI agents by nearly a decade, and this paper is honest about that chronology.

What would refute it. Given in full in §10. The sharpest version: if LLM procurement agents, tested against real or simulated pricing pages carrying §2's patterns, prove just as susceptible to visual and social-proof manipulation as humans — not merely to the narrower anchoring effect in §4 — this paper's central claim (that manipulation's architecture, not merely its intensity, must change to reach an agent) fails.

2. Anatomy of a SaaS Pricing Page

To understand what an agent sees when it encounters a pricing page, it helps to understand what the pricing page is designed to do to a human, and to understand how common and how documented that design actually is — this is not a folk belief about marketing, but the subject of a specific, peer-reviewed empirical literature.

The canonical enterprise SaaS pricing page has three to five tiers. The lowest is typically free or priced to feel embarrassingly inadequate — insufficient storage, limited seats, missing features the customer quickly discovers they need. The highest typically has no price: "Contact sales," which means we will charge you what your company can afford, negotiated individually, and we'd prefer you didn't shop around.

The middle tiers are where the persuasion architecture is densest.

Decoy pricing. Three tiers, where the middle tier is designed not to be the best value but to make the expensive tier look reasonable. If the tiers are $29, $79, and $149, the $79 tier typically exists to frame $149 as "only $70 more than what you were going to pay anyway." The middle tier is the decoy — its function is to shift the reference point, a mechanism first documented experimentally in 1982: adding an option that is asymmetrically dominated by one alternative but not the other measurably shifts choice toward the dominant option, across product categories as different as beer, cars, and television sets.⁴

Annual billing default. The price displayed is typically the monthly price when billed annually. The number looks smaller; the actual commitment is twelve times that, paid upfront. Monthly billing, if it exists at all as an option, is one click from the default — and defaults are sticky: the clearest demonstration of how much a default moves behaviour, even on a decision as consequential as organ donation, found consent clustering below a quarter of the population under opt-in regimes and above 90% under opt-out, for the same underlying willingness to donate.⁵ A pricing page's default is a smaller stake, same mechanism.

The "Most Popular" badge. This is a social-proof mechanism, in the specific sense developed over decades of experimental work on compliance and conformity: people treat what others are observed to do as evidence about what they themselves should do, particularly under uncertainty.⁶ The badge appears on whichever tier has the highest margin or the highest likelihood of converting to an enterprise upgrade. It says: other people chose this.

Seat-based pricing with ambiguous definitions. Many SaaS tools charge per seat, where "seat" means user — but whether that includes view-only users, API users, external collaborators, or archived accounts is typically unclear until a customer is deep into an annual contract and owes more than expected.

Cancellation friction. The purchase flow takes approximately four minutes. The cancellation flow is typically not accessible from the product interface. This is not a minor irritation; it is the single pattern that has drawn the largest regulatory consequence catalogued in this paper. In September 2025, Amazon agreed to pay $2.5 billion — a $1 billion civil penalty, $1.5 billion in refunds — to settle an FTC lawsuit alleging it enrolled consumers in Prime using deceptive design and then made cancellation deliberately difficult, under the Restore Online Shoppers' Confidence Act; the FTC's own case record states "Amazon leadership slowed or rejected changes that would've made it easier for users to cancel Prime because those changes adversely affected Amazon's bottom line."³ The asymmetry between a four-minute purchase flow and a multi-step cancellation flow is, on this evidence, sometimes a deliberated choice, overruled by a regulator rather than a UX team.

A 2019 large-scale crawl of 11,000 shopping websites — read directly, not taken from a secondary description — found at least one of fifteen documented dark-pattern types on 1,254 sites (~11%), 1,818 instances total, organised into seven categories (urgency, scarcity, social proof, misdirection, sneaking, obstruction, forced action), and traced 22 third-party services that supply these patterns as a packaged feature.¹ Whatever is happening on SaaS pricing pages is a subset of a documented, commercially supplied, industry-wide practice, not an isolated one. The FTC's 2022 staff report, "Bringing Dark Patterns to Light," organises enforcement-relevant examples under four headings — disguised advertising, cancellation obstruction, buried fees, data-sharing tricks — the same four showing up, in B2B dress, on enterprise pricing pages.²

3. What an Agent Actually Sees

When a procurement agent encounters a SaaS pricing page, much of §2's persuasion architecture is genuinely invisible to it — not because the agent is above it, but because it doesn't process the page the way a human does.

A human sees a visually rendered page. The hierarchy of information — large numbers, colored badges, highlighted tiers, progress animations — is designed for visual attention. The pricing page is a visual argument. The design encodes the argument: look here first, perceive this as normal, feel this is the right choice.

An agent reads structured data, or extracts structured data from unstructured text. It doesn't see the visual hierarchy the way a human's visual cortex does, and doesn't register a badge's placement and color as an endorsement unless explicitly instructed to weight badges as evidence, which most procurement-agent designs do not do. It queries for: what are the features, at what price, with what constraints, on what commitment period, with what cancellation terms?

If the pricing page doesn't answer these questions directly — and most don't — the agent has a problem. It can try to extract the information from natural language, follow links to terms of service for the cancellation policy, or infer seat definitions from documentation. But every piece of information it must infer rather than read directly is an error surface: the agent may miscalculate, land on the wrong tier, or get the spend wrong.

This is the immediate commercial consequence of pricing-page opacity: it raises the error rate of agentic procurement, independent of any particular bias. Companies early to adopt agentic procurement will, over time, prefer vendors whose pricing an agent can parse correctly — for this reason alone, before §4's more specific finding about numeric framing even enters into it.

There is an interesting adversarial dynamic here. A vendor who makes pricing deliberately opaque to exploit human biases might assume opacity will also confuse agents in the vendor's favour. In practice, the effect is often the opposite. A confused human might still complete the purchase, persuaded by other elements of the page. A confused agent may refuse the purchase outright, escalate for human review, or select a competitor whose pricing parses cleanly. Opacity built for humans may simply repel agents rather than fool them.

4. What the Research Actually Shows About Agent Susceptibility

It is tempting to claim that an AI agent has no psychological vulnerabilities. That claim does not survive a direct check against the published literature on LLMs and cognitive bias.

A 2025 study, published in Findings of the Association for Computational Linguistics: EMNLP 2025 and read directly for this revision, ran simulated price-negotiation sessions where a seller LLM agent deployed anchoring tactics against a buyer LLM agent, measuring the effect on the buyer's final agreed price.⁷ Its central finding: LLM agents showed anchoring susceptibility of the same kind documented in humans — a strategically placed number measurably shifted the buyer agent's outcome. The effect was not uniform: models using longer reasoning chains showed more resistance, suggesting explicit deliberation offers some, not complete, protection. A separate, independent 2024 study tested anchoring across a wider range of models and found a related, striking pattern: LLMs were more susceptible to an anchor framed as coming from an expert source than the same number presented plainly — closer to "expert framing shifting a number" than "a visual badge shifting a feeling," but a bias shifting a purchase-relevant decision all the same.⁸

This is a real correction, and it is worth being precise about what it corrects and what it leaves standing. It does not restore the pricing-page dark patterns of §2 to full effectiveness: a "Most Popular" badge, a colored highlight, a countdown-timer graphic are visual persuasion devices, and nothing in either study suggests an agent reading structured or natural-language text registers a colored badge the way a human's visual system does. What the research does establish is that "an agent cannot be manipulated" is false, and the correct claim is narrower: the manipulation that reaches an agent has to be encoded where the agent actually looks — in the numbers and framing language it processes as data — not in the human visual and social channels SaaS pages were built to exploit. A vendor motivated to manipulate an agentic buyer would not keep the badge. It would plant a number: a strategically high anchor stated early in a product description, or a false "industry-standard" or "expert-recommended" claim structured to trigger exactly the expert-framing effect the 2024 study measured.

No study located for this paper tests this in a live B2B SaaS procurement-agent setting, as distinct from the simulated tasks both studies ran — a gap named plainly in §1 and carried into Appendix B. What the next dark pattern — the one built for agents rather than humans — will actually look like is an open and somewhat urgent question, not a solved one.

5. The Vendor's Dilemma

Software vendors now face a structural choice that §4 complicates rather than removes: optimize the pricing page for human biases, or for agentic parsability — while holding in mind that an agent has its own narrower, but real, set of weaknesses worth exploiting on its own terms.

A pricing page optimized for human decision-making is built around the mechanisms in §2 — anchoring, social proof, defaults, ambiguity — deployed through visual design. One optimized purely for agentic parsability is built around clarity: explicit features, pricing, commitment terms, and cancellation policy, ideally machine-readable rather than visual. These two are not compatible with each other. But §4 means a third design is possible — clear enough to parse, while still carrying a numeric anchor or unverifiable "expert-recommended" framing sized to move an agent the way a badge once moved a human. No evidence exists that any vendor is doing this yet; it is a prediction, named as such.

The businesses paying for software through automated procurement are, disproportionately, the businesses with the highest transaction volume. Forrester's 2026 survey of nearly 18,000 global B2B buyers — a large, named industry survey rather than an independently audited census, treated here as a press-reported order-of-magnitude figure — found 94% now use AI to research vendors before a sales conversation, more than half comparing vendors inside an AI tool specifically.¹³ Whatever the precise percentage, the direction is not in serious dispute: the customer who sends an agent to compare vendors is, on average, higher-volume and more process-mature than the one still filling out the form manually.

Which means the pricing pages designed to manipulate human buyers are increasingly failing — or, per §4, failing differently — on exactly the customers who are highest-value.

The vendor dilemma has no clean resolution. Removing visual dark patterns means giving up conversion-optimisation mechanisms tested and validated for decades. Keeping them means accepting higher agentic error rates and a growing high-volume-buyer preference for cleaner alternatives — while also deciding whether to adopt the newer, agent-specific manipulation §4 makes possible, with its own untested regulatory and reputational exposure.

The resolution will likely vary by segment. In consumer software, the audience stays predominantly human for the foreseeable future, and consumer-oriented dark patterns — the ones behind §2's $2.5 billion settlement — will persist, regulatory risk notwithstanding. In SMB and mid-market B2B, where procurement automation is spreading fastest, the economics shift toward parseable pricing for the reasons in §3, with §4's agent-specific risk a live complication rather than a settled addition.

6. The Pricing API as Inevitable Destination

The logical endpoint of §5's pressure toward clarity is the pricing API — grounded now in events that have already happened, not merely in architecture.

A pricing page is a document for human reading. A pricing API is an endpoint for machine consumption. The infrastructure layer learned this lesson first: AWS published a structured Price List Bulk API in 2015 and a fuller Query API in 2017 — nearly a decade before the current wave of AI agents — followed by Azure's Retail Prices API and Google Cloud's Billing Catalog API, confirmed directly from each provider's current docs.¹⁴ Cloud pricing APIs were not built because of agentic buyers; large enterprise customers already needed programmatic cost estimation. What is new, dated to the second half of 2025, is the application and payment layers catching up — for agentic buyers this time.

In one six-month window, April to December 2025, four payment networks a B2B purchase would actually run through all published agent-specific commerce infrastructure, each fetched and read directly from the company's own announcement. Mastercard launched Agent Pay on 29 April 2025, binding a tokenized card credential to a specific registered agent, merchant scope, and consent policy.⁹ Visa announced Intelligent Commerce the next day, opening its network to AI agents that search, recommend, and pay under cardholder-set spending parameters, with Anthropic, IBM, Microsoft, OpenAI, Perplexity, and Stripe among its first partners; its Trusted Agent Protocol, layering a verifiable-agent credential onto existing card-acceptance infrastructure, followed later in the year.¹⁰ OpenAI and Stripe released the Agentic Commerce Protocol, an open Apache 2.0 standard, on 29 September, powering "Instant Checkout" inside ChatGPT, starting with Etsy and expanding to Shopify's merchants.¹¹ And on 16 September, Google published the Agent Payments Protocol (AP2), built around a cryptographically signed "mandate" recording what a user authorised an agent to buy, under what limits, with sixty-plus launch partners including Mastercard, PayPal, American Express, and Coinbase.¹²

None of this is speculative. All four are live, documented protocols as of this writing, built around an agent transacting on a human's behalf — newer than the cloud-pricing-API story told above. How spending authority gets granted in the first place — the card or mandate itself, distinct from the pricing decision preceding it — is a separate argument this corpus makes elsewhere, not re-derived here.

The dark patterns in §2 are, on their face, incompatible with this infrastructure: you cannot put a "Most Popular" badge in a signed payment mandate, and an API field cannot be made to feel urgent. But §4's correction applies here too: a mandate records what a user authorised, not whether the number shown to form that authorisation was itself anchored or falsely framed as expert-recommended upstream of it. This infrastructure secures the transaction. It does not yet secure the decision that produces it, and §4 is why that distinction now matters. What happens when a correctly-authorised agent purchase still goes wrong — who bears the loss — is a question this corpus answers separately, not one this paper takes up.

7. What Procurement Agents Are Selecting For

ProcureBee evaluates vendors on a set of criteria that are structurally different from what a human buyer evaluates on, and §4's correction sharpens rather than removes this section's argument.

A human buyer evaluates pricing pages emotionally and analytically in roughly equal measure. The emotional evaluation — does this company feel trustworthy? does this price feel fair? do other people like me use this? — is heavily influenced by the visual persuasion architecture in §2. The analytical evaluation happens second, if it happens at all, and is often working from information the pricing page has deliberately obscured.

A procurement agent evaluates vendors analytically, almost exclusively, and on the evidence in §4, its analytical process is not thereby free of exploitable structure — it is differently exploitable. Its stated criteria are:

• Is the price explicitly stated for the required configuration?
• Are the features per tier explicitly defined, without ambiguity?
• What is the commitment period, and what are the exit terms?
• Is the vendor on the approved list, or verifiable against the criteria for the approved list?
• Does the total cost of this vendor, including likely renewal rate and exit cost, fall within the authorized budget?

Vendors that score well on these criteria are easier for the agent to process, generate fewer errors, and are less likely to trigger escalation for human review. Over time, a procurement agent tracking its own performance will develop a preference — not explicit, but a pattern — for vendors whose information is clear. §4 adds a specific risk rather than contradicting this: a vendor could satisfy every item on this list in form while still placing a strategically high anchor or an unverifiable "expert-recommended" claim inside the text an agent reads to answer them. Whether real-world agent designs are vulnerable to this in practice is exactly the gap named in §1 and §4, and is not yet tested.

The companies that will win the automated buyer are not necessarily the cheapest or the best product. They are the companies whose pricing is honest enough for a machine to understand without guessing — and, on the evidence gathered here, without smuggling a number the agent has not been built to question.

8. Objections I take seriously

"Flux has a commercial interest in this conclusion." Flux builds ProcureBee, and this paper's central argument — that agentic procurement pushes vendors toward pricing clarity, to the advantage of exactly the kind of agent ProcureBee is — is close to ProcureBee's own value proposition. The independent evidence in §2 and §6 is genuinely independent of Flux and would hold regardless of ProcureBee's existence. What is not independent is the claim that ProcureBee itself is well-designed against the risk named in §4 and §7; no independent audit of its resistance to numeric anchoring or false-expert framing exists, and this paper does not claim one.

"Section 4 spends most of its length correcting a claim this paper's own opening leans on — doesn't that undermine the argument?" No, but it changes what the paper actually claims, and it would be dishonest to pretend otherwise. "Agents are differently, not less, manipulable, and the manipulation shifts from visual/social channels to numeric/framing channels" is narrower and more defensible than "agents have no psychological vulnerabilities" — and more useful to a vendor or buyer trying to act on it, since it names a concrete thing to guard against rather than a false reassurance.

"The decoy effect, anchoring, and social proof research (Huber/Payne/Puto 1982, Johnson/Goldstein, Cialdini/Goldstein) is about human subjects decades ago. Does it describe a 2026 SaaS pricing page?" These are foundational, heavily replicated findings, not single untested studies, and the Mathur et al. 2019 crawl in §2 shows these mechanisms deployed, named, and commercially packaged at scale today — not merely a theoretical possibility.¹⁴⁵⁶ The step from "these mechanisms exist and are deployed" to "this specific page uses this specific mechanism for this reason" is still an inference in any individual case, and this paper audits no named company's design intent.

"Regulatory crackdowns like the Amazon settlement should make this paper's claim about agent-driven pressure moot — regulation, not agents, will fix dark patterns." The regulatory picture is more mixed than this assumes. The Amazon settlement is real and large. But the FTC's broader "click-to-cancel" rule — meant to generalise that one case into a standing requirement — was vacated in its entirety by the Eighth Circuit in July 2025 on procedural grounds, and no replacement rule is yet in force, only an early-stage 2026 rulemaking notice.¹⁵ Regulatory and agentic pressure read as two separate, currently uncoordinated forces, not one substituting for the other.

9. Where the agent-susceptibility finding stops

§4's correction is the central limitation already stated in this paper: AI procurement agents are not immune to manipulation.

It does not claim every dark pattern in §2 will disappear from software pricing; consumer software is out of scope, and even in B2B this paper predicts a shift in segment and mechanism, not extinction.

It does not claim any specific vendor is currently deploying an agent-targeted numeric or expert-framing manipulation of the kind §4 makes theoretically possible; no such case was found, and the risk is a stated prediction, not an observed practice.

It does not claim the cloud pricing APIs in §6 were built in response to agentic buyers; their documented history predates the current wave of AI agents by close to a decade.

It does not claim ProcureBee has been independently tested against the risks named in §4 and §7; no such audit exists.

And it does not claim the regulatory and agentic pressures in §§2, 6, and 8 are coordinated; a $2.5 billion settlement alongside a fully vacated federal rule shows them moving on independent, sometimes contradictory tracks.

10. What would falsify this

Stated as commitments, in decreasing order of how much damage each would do to this paper's claims.

LLM procurement agents, tested directly against real or simulated pricing pages carrying the visual and social-proof patterns in §2 — not merely the numeric-anchoring tasks in §4 — prove just as susceptible as human buyers. That would mean this paper's central distinction (manipulation must change channel to reach an agent, not just fail) is wrong, closer to its opposite: agents are manipulable by the same means, via a different sensory path into the same underlying weaknesses.

A substantial number of vendors are found deploying an agent-targeted numeric anchor or expert-framing claim to exploit the §4 mechanism, rather than this remaining a theoretical risk. That would convert this paper's most speculative prediction into documented practice, demanding more urgent treatment than the "open question" framing given here.

Agentic-procurement adoption, measured independently rather than through vendor-adjacent surveys like Forrester's, turns out concentrated in low-volume or non-representative buyers rather than the high-volume buyers §5 claims are driving the shift. That would undercut the account of why vendors face pressure toward clearer pricing, even if §4's bias research still held.

The FTC's 2026 rulemaking produces a durable click-to-cancel rule that survives appeal, and enforcement becomes routine rather than exceptional. That would not falsify the agent-driven mechanism, but would mean regulation, not agentic pressure, is doing most of the work — a different story about why pricing pages change, even if they do.

Appendix A — Evidence table

Appendix B — Citation register

CITATION-NEEDED — dark-pattern susceptibility of LLM procurement agents in a live B2B SaaS purchase flow. Appendix A row 16. Both directly-read studies (Takenami et al. 2025; Lou & Sun 2024) test simulated negotiation and evaluation tasks, not a procurement agent completing an actual SaaS checkout against a real pricing page. This is the single largest evidentiary gap behind this paper's central claim, named in §1 and §4 rather than assumed closed.

CITATION-NEEDED — measured deployment of agent-targeted manipulation. Appendix A row 17. No vendor has been found deploying a numeric anchor or false expert-framing claim specifically to exploit the §4 mechanism against procurement agents; this remains a prediction, stated as such in §5, §9, and §10.

CITATION-NEEDED — independent measurement of agentic-procurement adoption by buyer segment. Appendix A row 18. The Forrester figure used in §5 and Appendix A row 13 is a large, named industry survey, not an independently audited census, and §10 names a specific way this could falsify the paper's account of vendor incentives.