Consumer agents may reduce search, attention and execution frictions, but the consequences depend on whose interests the agents serve, which consumers adopt, and how sellers respond through prices, products and access. In insurance, changes in coverage and risk selection matter as much as renewal savings. This brief maps what the recent economics, simulation and experimental papers establish about each link in that chain, and what they do not.
The question has a twenty-year-old formal model that the agent literature inherits. In Gabaix & Laibson (QJE 2006), firms sell a base good below cost and recover the margin through a shrouded add-on that only myopic customers pay. Sophisticated customers avoid the add-on and are, in effect, subsidised by the myopes. The twist that matters here: a competitor cannot profitably "unshroud" and educate the myopes, because educated customers simply buy the subsidised base good from the shrouding firm. Armstrong & Vickers (JEL 2012) generalise this to contingent charges, and DellaVigna & Malmendier (2004) get the same cross-subsidy from self-control problems in gym and credit-card contracts.
The bookkeeping that follows is a conditional identity rather than a prediction. If a share s of customers pays a fee f and the firm must break even on a unit cost c, then with quantities and the pricing structure held fixed:
Reducing the fee-paying share raises the break-even headline price. If an agent makes avoidance free for the half of the market that adopts it, the add-on revenue from that half disappears and the headline must rise by the lost per-customer margin. Adopters save; sellers reprice; the unassisted remainder funds the difference. Note what this does and does not say: universal sophistication changes who pays without reducing the average resource cost, so it is a redistribution, not a creation of value. Whether real sellers reprice this way, or reach for other instruments, is the empirical question the 2026 papers attack.
The cross-subsidy, stylised
Who pays what for a good that costs the firm $100 to supply, 50/50 myopic/sophisticated split, no incremental cost of the fee (USD)
Illustration of the model's logic, not data. Before assistance, average revenue is ($110 + $90)/2 = $100, so the firm breaks even on the mixed population. Once nobody pays the fee, the break-even headline rises to $100. Gabaix & Laibson's own hotel example ($80 + $20) assumes everyone is myopic and then asks what one educated customer would do at unchanged prices; it is not a mixed-population equilibrium and should not be transplanted into one.
| Customer | Headline | Add-on | Total paid |
|---|---|---|---|
| Myopic, before assistance | 90 | 20 | 110 |
| Sophisticated, before assistance | 90 | 0 | 90 |
| Everyone assisted, after repricing | 100 | 0 | 100 |
Two more theoretical pieces sharpen the frame. Balzer & Young (Sept 2026) make naivety endogenous: consumers invest in understanding their mistakes in proportion to how costly remaining naive is, so firms face a trade-off between the depth of exploitation (profit per naive customer) and its reach (how many stay naive). Deeper exploitation shrinks the exploitable base. This disciplines firms even before agents exist, and it predicts that policies which help consumers at fixed cognition can backfire once cognition responds. An agent is the limiting case: it drives the cost of sophistication toward zero for whoever has one. Accornero (July 2026) argues from the demand side that consumer behaviour under delegation does not disappear but migrates up a level, to trust calibration, delegation depth, override and audit.
The industries at stake are the ones where the gap between what people pay and what an attentive comparison would pay is large and persistent. The studies below are the ones usually cited, including in Van Loo's "Consumer Agents" (Wash. U. L. Rev. 2026). They are not estimates of the same quantity, so they are listed by what each one measures rather than ranked on a common scale.
What the pricing-friction studies actually measure
Each figure is a different outcome; the column on the right says which
| Study | Setting | Headline figure | What the figure is |
|---|---|---|---|
| Blake, Moshary, Sweeney & Tadelis (Mktg Sci 2021) | StubHub, back-end fees vs all-in prices | +20.6% | Revenue per visitor under back-end fees, combining +14.1% purchase probability and +5.4% revenue conditional on purchase. Not an identical-ticket overcharge. |
| Lu, Markovits, Miller & Van Loo (2025) | Constructed gift-card marketplace, real purchases | ~10% (max 13%) | Average price paid for an identical gift card under drip pricing vs transparent pricing. |
| Ellison & Ellison (2009) | Online computer parts, shipping-fee obfuscation | ~6% | Margin above the competitive price attributed to obfuscation. |
| Medicare Part D plan-choice studies | Prescription-drug plans | ~30% | Estimated excess spending from choosing a worse plan than the cheapest suitable one; a savings-from-switching figure, not a price for an identical good. |
| Einav, Klopack & Mahoney (AER 2025) | Ten card-billed subscriptions | +87% to +120% | Seller revenue attributable to inattention and inertia across two structural models, holding the initial subscriber pool fixed; no price or product response modelled. |
Sources: Van Loo (2026) compilation for the first four; Einav, Klopack & Mahoney, "Selling Subscriptions", AER 115(5), 2025 (the 2023 working-paper abstract gave a 14% to 200%+ range across individual subscriptions).
Practitioner and analyst writing frames the same thing as an "inertia premium" or "inertia tax", usually with aggregates rather than estimates of recoverable rents. Fintech Blueprint points to roughly $23 trillion of the world's $70 trillion in consumer deposits sitting in near-zero-rate checking accounts, and $234 billion of 2024 US credit-card revenue; both are totals, not profit, and neither is an estimate of how much an agent could move. Bloomberg Intelligence lists insurance, banking and deposits, telecom and cable, subscriptions, online travel, rideshare and delivery as the sectors most exposed, on the grounds of high switching costs, complex pricing, auto-renewal and low engagement. One strategist quoted by Business Insider suggests the economics would move at 40–50% agent adoption rather than 5–10%; that is an attributed scenario, not a threshold established by research.
The 2026 papers are mostly theory and simulation. The empirical ground they stand on is older work on search, switching and comparison technology, much of it from insurance. Two of these papers matter especially: one gives a historical analogue for what comparison technology did to prices, and one is the counterexample that making individual insurance choices better can make the insurance market worse.
Inertia in plan choice is large and costly to the individuals who exhibit it. But in the studied setting, a counterfactual that reduces inertia and improves individual choices worsens adverse selection enough to reduce aggregate welfare, because the people who switch toward cheaper plans are disproportionately the healthier ones. This is the essential insurance counterexample to "removing friction is good": in a pooled market, a friction can be holding the pool together. It comes from a specific health-insurance setting and should not be mechanically transferred to individually risk-priced motor insurance, but it is the reason an insurance analysis of agents must model selection, not just churn.
Directly estimates search and switching frictions in US auto insurance. The two are different mechanisms: finding alternatives (search) and moving to one (switching). Honka's estimates identify search costs as especially important, which is relevant because an agent attacks search cost first and switching cost only if it is authorised to act.
The historical analogue. Growth of internet comparison sites reduced term-life prices by roughly 8–15% for the policies that were being compared, with the effect concentrated where comparison use was high. This is evidence about comparison technology adopted by humans, not autonomous agents, but it is the best prior for "what happens to prices when comparison gets cheap in an insurance line."
Mailing personalised cost comparisons to Medicare Part D beneficiaries raised plan switching from 17% to 28% and cut predicted costs. Information alone moved behaviour, but far from completely. This is the cleanest field evidence for separating the value of information from the additional value of delegated action, which is exactly the split Huang & Xiao test in simulation.
An unusually close predecessor to the current LLM literature. Treats delegated purchasing, buyer coalitions and collective bargaining, intermediary market power, data-access barriers and the question of who the algorithm actually serves, years before agents were deployable. Several of the research directions at the end of this brief are already posed here.
The FCA's pricing-practices work is the most direct real-market evidence on renewal pricing, "price walking" of loyal customers, switching costs and practices such as requiring a phone call to cancel auto-renewal. The 2025 evaluation assesses what the 2022 remedies (renewal price no higher than the equivalent new-business price, easier auto-renewal cancellation) did to prices, product quality and switching. It is a regulatory intervention rather than an agent intervention, but it is evidence on exactly the mechanisms the agent papers hypothesise.
Single-user evidence that an assistant saves its user money cannot tell you who pays once sellers respond. These papers model or simulate the whole market.
Builds a one-day car-rental market in which language models play 200 renters, their assistants, and six adaptive sellers armed with an algorithmic pricing tool. The contract carries a daily rate below cost plus two contingent fees engineered for identification: a refuelling charge only the renter's physical action can avoid, and a pre-ticked add-on that an authorised assistant can cancel online. Half the population gets an assistant (advisory or executing); a fixed tenth never does. A closed-form bound and a deterministic "rule market" are frozen as forecasts before the LLM society runs, and every shocked branch is paired with a frozen-seller twin so adoption effects and market feedback can be separated.
What it showsThree results, all pre-registered. First, value comes from acting rather than reading: the simulated renter opens the fine print on 82% of days and still beats the cancellation deadline on only 36% of rentals; an executing assistant lifts that to 97% and advice alone to 93%. Second, sellers do reprice, recovering about 55% of the lost fee revenue through a higher headline rate while abandoning the fees themselves, so adopters keep about two-thirds of their first-day saving, and over two-fifths of what survives is completed transactions rather than dodged charges. Third, the feared bystander tax does not appear: the unassisted panel's spending moves by +$0.38 per renter-day with a confidence interval spanning zero, against a calibrated forecast of +$3.60. Across seven configurations the bystander outcome tracks the fee margin sellers carried before the shock, and when fee-setting is handed to the pricing tool, fee-free seller-periods jump from 19% to 89% and fees vanish for everyone.
Forecast versus outcome in the Huang & Xiao society
USD per sampled renter-day, executing mandate, final eight periods, 30 markets (95% intervals in table)
The deterministic layer called the direction of every response and overstated how far sellers complete it. The seller recovery ratio, not plotted because it is a share, was 0.66 forecast against 0.55 ± 0.31 observed.
| Registered quantity | Forecast | Outcome |
|---|---|---|
| Adopters' saving, adapted | 5.6 ± 0.3 | 8.7 ± 1.4 |
| Headline rate rise | 9.6 ± 0.3 | 4.4 ± 1.8 |
| Bystanders' burden | 3.60 ± 0.34 | 0.38 ± 0.94 |
| Adopters' saving, first day | 13.8 ± 0.4 | 13.3 ± 1.5 |
Acting beats reading (in simulation)
Share of simulated rentals on which the add-on was cancelled before the deadline
Huang & Xiao, Section 4. The phone-only row is the executing mandate facing a seller that requires cancellation by phone. Advice and execution are close; the large gap is between any assistance and none.
| Condition | Timely cancellation |
|---|---|
| Unassisted renter | 36% |
| Advisory assistant (renter clicks) | 93% |
| Executing assistant | 97% |
| Executing, seller imposes phone-only path | 13% |
Models a two-sided market where consumers pay to acquire signals of product fit before buying, and the market itself learns from indications of fit and quality. Characterises the long-run steady state and then asks what two different improvements in search technology do: making search cheaper and making it more informative.
What it showsThe two improvements are not the same. Cheaper search improves both market learning and consumer surplus. More informative search can degrade both, unless the market learns as much as consumers do, for instance by reading the transcripts of agentic conversations. At equilibrium prices in symmetric markets, cheaper search raises consumer surplus but more informative search may lower it by weakening competition between businesses. This is the main theoretical reason to doubt that "better-informed buyers" straightforwardly means "lower prices": a buyer who knows exactly which product fits is a buyer with less reason to be contested. The same group's review, The Evolution of Digital Search: From Blue Links to Delegated Decision-Making, flags market dynamics and seller instrument choice as the open front.
Compares human and LLM-generated purchase decisions for pizza-restaurant orders, using menu calorie-labelling laws as the comparison information remedy, and builds a structural model of delegation to quantify mechanisms and welfare.
What it showsLabelling had limited effects on calorie intake; in the counterfactual where purchases are delegated to ChatGPT, total calories fall by up to 18%. This is a modelled counterfactual, not an observed effect of a consumer deployment. Delegation also causes substantial exit of unsold product varieties and higher market concentration. The authors' current description makes welfare conditional: it could improve, depending on why consumers delegate and on the underlying technology, with important distributional consequences. An earlier (May 2026) abstract stated the distribution more specifically, with delegators gaining substantially, non-delegators losing slightly through equilibrium prices, and firm profits rising; those claims should be read against the specific version and table before being quoted. For this brief the durable points are two: delegation changes product choice enough to move the supply side (variety, concentration), and surplus is redistributed through uniform prices between those who delegate and those who do not.
Takes the classical sequential-search framework (Anderson–Renault) and asks how delegation to probabilistic, content-generating agents changes equilibrium prices, entry and matching efficiency. The propositions are stated informally within a conceptual framework rather than proved as theorems.
What it showsAgents can drive the technical cost of search toward zero, but the empirical record on current agents (position bias, early stopping, satisficing) acts as a residual behavioural search cost. With effective cost = technical + behavioural and the behavioural part positive, equilibrium markups stay bounded away from the frictionless Bertrand benchmark even as the technical part vanishes. AI-generated content adds noise to perceived match values, reducing allocative efficiency, and cheap AI product generation lowers entry costs. The effects differ across market architectures rather than being uniform.
Why markups need not fall to zero
Equilibrium markup as technical search cost falls, with and without a residual behavioural friction (illustrative)
Schematic of Bichler's proposition, drawn to show the shape of the claim. The floor is set by the behavioural component; its height in any real market is an empirical matter.
Compares manual search, which evaluates a small product set accurately, with agentic search, which screens a broad catalogue through noisy representations of both preferences and products. Models the platform's choice of how faithfully to represent each consumer.
What it showsAgentic search avoids the "collapse" of manual search when catalogues grow. When manual inspection is cheap enough, agentic search becomes revenue-superior for the platform before consumers voluntarily adopt it, producing an adoption lag in which rational consumers keep searching manually. Conditional on participation, platforms may assign lower representation fidelity to consumers with larger attention budgets, because those consumers can offset noise by refining. Scarcity shifts from product inspection to preference articulation, which is why the willingness and ability to interact with an agent becomes the central variable.
A randomised online marketplace (n = 1,608) built for the experiment, in which participants buy real gift cards from a $15 endowment. Initial and final price presentation varies across 33 levels, and participants are randomly given an automated assistant that flags a cheaper identical option and links to it. The assistant is a rule-based comparison tool, not a deployed general-purpose LLM agent.
What it showsDrip pricing raises the average price paid for identical gift cards by up to a plateau of about 10%, with a maximum of 13%. The assistant cuts the incidence of overpayment by 63% and lowers dollars overpaid enough that, on the experimental demand and payoff estimates, the seller's profit-maximising strategy flips from drip pricing to transparent pricing. That flip is inferred from the experiment's payoffs, not observed as long-run seller adaptation. A side effect: participants exposed to drip pricing rate the platform as less trustworthy even though the third-party seller sets the fees, and the assistant partially worsens that perception. This is the paper Huang & Xiao build on; it establishes that an assistant works for its user and leaves the market-level question open.
A field-tracking reference point the simulation papers all cite: Bakos, Marotta-Wurgler & Trossen (J. Legal Studies 2014) found that only 0.1–1% of shoppers open a licence agreement. The "reading" problem is real; the open question is whether agents fix the reading or the acting, and Kling et al. (above) is the existing field evidence that information alone goes only part of the way.
Develops a formal framework for airline markets in which the buyer is an algorithm executing a contract (a budget cap and preference parameters) rather than a human with a valuation. The companion paper simulates increasing agent penetration across markets with different fare structures and search fragmentation.
What it showsA population of booking contracts induces a demand function equal to the survival function of the budget-cap distribution. Drip pricing becomes allocatively inert because the agent evaluates the all-in price. Where the cap distribution has mass points, demand has cliffs, and price sensitivity concentrates on those thresholds (an "elasticity concentration index"). Intertemporal fare fencing loses its premium to the extent delegation severs the link between purchase timing and valuation. The simulation adds that greater penetration compresses sustainable fare differences rather than pushing all fares down, and that the right airline response depends on competitive position.
When contracts cluster, demand becomes a staircase
Share of buyers who purchase at each fare: dispersed valuations vs agents on a few common budget caps (illustrative)
Schematic of the Abdelghany & Abdelghany result under the mass-point condition. If caps were as dispersed as human valuations, the two curves would coincide.
No paper yet studies autonomous LLM buying agents in subscriptions, banking or insurance directly. The economics of the underlying frictions are well studied (see the earlier-evidence section: Einav, Klopack & Mahoney on subscription inertia; Handel, Honka, Brown & Goolsbee and the FCA on insurance), and the practitioner literature applies them to agents by argument. Abundance Economics Research argues agents turn passive renewal into continuous comparison and makes the second-order point that if agents disproportionately remove low-usage members, the remaining pool is differently selected, not just smaller. The Consumer Finance Monitor series (Ballard Spahr) walks through banking, cards, lending and insurance and asks whether the existing consumer-protection framework works when the agent is the principal interface. RTÉ Brainstorm gives the "loyalty penalty" in car and health insurance as the canonical case.
Sellers are not passive. These papers model what the supply side does when it meets a buying agent.
Designs the buy-or-wait policy for an agent with a finite shopping window under three information regimes (stationary, Bayesian, robust) and evaluates them on 48,933 Keepa price observations across 367 Amazon items. Optimal policies are threshold rules; the robust randomised policy does best at the 10th percentile of outcomes. The relevant finding for this brief: language models are better at choosing which regime and calibration to use than at making the buy-or-wait call themselves. A "rational" buying agent is a decision rule with an LLM wrapper, not an LLM deciding.
Both principals delegate to agents that receive different value signals and execute machine-enforced mandates. The same nominal fairness rule can produce no trade because each agent applies it to its own posterior. Verification can restore trade only when it covers the disputed proposition; attestation of a seller's signal can legitimate a high price, while attestation of a restrained rule prevents unauthorised extraction. In repeated relationships, offers above a buyer's reference cap depreciate relationship capital. Verification is privately under-provided when sellers do not internalise avoided inspection, and over-provided when it facilitates extraction.
Most algorithmic-collusion work is about seller-side algorithms (see Fish, Gonczarowski & Shorrer 2024 on LLM pricing agents reaching supracompetitive prices). This paper asks what a consumer-side AI provider can do about it: inject demand noise by randomising the purchases of a small fraction of subscribed consumers, which breaks the monitoring that collusion depends on. Even though the provider maximises its own profit rather than consumer surplus, its presence always raises consumer surplus by breaking collusion, and the noise can be tuned to market characteristics. A related theoretical result (Breaking Algorithmic Collusion in Human-AI Ecosystems, Nov 2025) shows that even one human defector can destabilise AI collusion in repeated pricing games.
A simulation, calibrated to measured LLM biases, of a platform that controls rankings and sellers that control listings, each learning independently that exploiting the shopping agent's biases pays. Joint exploitation produces more than twice the consumer harm of either strategy alone, because ranking decides which products land in bias-triggering positions while listing manipulation moves conversion. Since the harm comes from aligned incentives rather than an agreement, the author argues it may evade antitrust detection. The paper's broader point: human biases are heterogeneous and noisy, while the biases of agents sharing a model and configuration are correlated, which makes them cheaper to target.
Two further pieces belong on the list though only their abstracts are public: Chen, Gan, Qin & Yue, "Open or Restrict? Platform Governance of AI-Assisted Comparison" (SSRN 7033679) on whether platforms should admit comparison agents at all, and Want, "When Consumers Delegate Choice to AI: Strategic Adaptation and Information Asymmetry in Agentic Markets" (SSRN 7282860). Huang & Xiao also document the obstruction instruments already visible in the field: incumbent lawsuits against buying agents, bot walls and CAPTCHAs, and dark-pattern benchmarks for GUI agents.
Everything above assumes the agent is the rational party. The agent-behaviour literature says that assumption is, at best, model- and configuration-dependent. Two things need keeping apart here: sensitivity to price, which is usually what a principal wants, and susceptibility to irrelevant features such as listing position or marketing nudges, which is the behavioural problem.
Varies prices, ratings and psychological nudges in a controlled web-shopping environment and measures how LLM agents' choices move. On nudges, the finding is that agents respond strongly even without the cognitive constraints that cause human bias; the authors report susceptibility amplified 3–10× relative to humans and read agents as implementing simplistic hierarchical decision rules. On price, 13 of 17 models significantly prefer cheaper options in the baseline condition, with effects from 15.2 percentage points (o3) to 34.3 (GPT-4o Mini), and price sensitivity intensifies once ratings are matched. The price result is heterogeneity in a mostly desirable behaviour, not evidence of bias on its own; the nudge result is the behavioural finding. Both come from a controlled choice environment rather than open-web shopping.
How much a cheaper price moves an agent's choice depends on the model
Shift in choice share toward the cheaper option, percentage points, baseline condition (ABxLab, ICLR 2026)
The two named endpoints of the reported range across 17 models. Preferring the cheaper option is normally what a principal wants; the point of the chart is that the strength of the preference varies by a factor of two across models, which is what a seller or platform would optimise against.
| Model | Effect (pp) |
|---|---|
| GPT-4o Mini (largest) | 34.3 |
| o3 (smallest significant) | 15.2 |
Builds the ACE benchmark and runs instruction-following and basic economic-rationality checks on shopping agents, then randomises product position and listing attributes. The better tests of problematic behaviour are here: agents show strong, heterogeneous position biases not easily mitigated by the interface (web vs text), and the biases resist simple prompting. Delegation reduces search friction but carries risk when small price or rating advantages are missed, position bias dominates, or endorsements drive choices. Follow-ups by Wadi & Ma (Does Rank Still Matter?, Shopping by Algorithm) extend the position-bias and human-heuristic findings.
Not a consumer paper, but the same lesson from finance. LLM traders in an experimental asset market show a pronounced disposition effect and recency-weighted extrapolative beliefs; these aggregate into bubbles of the kind human markets produce. The useful twist is that the biases are programmable: targeted prompts that suppress a specific bias materially reduce bubble magnitude. A similar result on loss aversion and herding appears in J. Computational Social Science (Feb 2026).
The umbrella piece is LLM Consumer Behavior Theory (June 2026), which tries to organise all of this under a common economic lens and lists where rationality and heterogeneity assumptions fail in agentic markets. The sharper framing of the original question may therefore be: industries do not lose their irrational customers so much as trade heterogeneous human biases for potentially correlated biases among agents that share a model and configuration, which are cheaper to exploit but also cheaper to patch. Uniformity across agents is not established; the cited work finds substantial differences between models.
The question is better posed as "which frictions does delegation remove, which new ones does it introduce, and how does the market adjust?" than as "what happens when buyers become rational?". On that framing, "consumer savings" is not one outcome but four, and the papers speak to them very unevenly.
Four different outcomes that "consumer savings" conflates
What each strand of evidence says about each
| Outcome | What the evidence says | Strongest source |
|---|---|---|
| Lower payments | Established for the individual user of a comparison aid; partly clawed back by sellers in simulation; a redistribution, not value creation, where it comes from fee avoidance alone. | Lu et al.; Brown & Goolsbee; Huang & Xiao |
| Better products or matching | Theoretically ambiguous: cheaper search improves matching, more informative search may not; delegation can shrink variety and raise concentration. In insurance, no paper checks whether the cheaper policy preserves cover. | Lucier et al.; Delaprez & Hortaçsu |
| Saved time and effort, completed transactions | The component that survived every seller response in simulation (follow-through, cars secured at all); the field evidence shows information alone gets only part of the way. | Huang & Xiao; Kling et al. |
| Redistribution among consumers, sellers and agent providers | Theory predicts a transfer from the unassisted to adopters; simulation finds it small or absent unless fee margins are thick; one structural model has firms gaining. Agent providers' own take is unmodelled except in the collusion paper. | Gabaix–Laibson; Huang & Xiao; Delaprez & Hortaçsu; Cheng & Aseri |
| Risk selection (insurance only) | Reducing a choice friction can worsen the pool enough to lower aggregate welfare. No agent paper addresses it. | Handel |
A claim-by-claim reading, tagged by the strongest kind of evidence behind it.
Evidence coverage by question
Evidence on the economic mechanism (any technology) versus direct evidence on LLM-agent deployment
| Mechanism (any technology) | Direct LLM-agent evidence | ||||
|---|---|---|---|---|---|
| Question | Theory | Field / empirical | Theory | Simulation | Experiment |
| Does a comparison aid save its own user money? | ● | ● | ● | ● | ● |
| Do sellers reprice, and how much is clawed back? | ● | ● | ● | ● | ○ |
| Do the unassisted pay more? | ● | ○ | ● | ● | ○ |
| Does competition soften with better-informed buyers? | ● | ○ | ● | ○ | ○ |
| Does easier switching change risk selection? | ● | ● | ○ | ○ | ○ |
| Are the agents themselves rational? | — | — | ● | ● | ● |
| Can sellers obstruct or exploit agents? | ● | ● | ● | ● | ● |
| Insurance / subscriptions / banking specifically | ● | ● | ○ | ○ | ○ |
● evidence exists in the papers above · ○ none found · — not applicable. Mechanism evidence for sellers repricing includes the fee-to-rate migration after the US CARD Act (Agarwal et al., QJE 2015) and the FCA's pricing remedies; the field column for the first row is Kling et al. and Lu et al. The direct LLM-agent columns are thin everywhere outside e-commerce, and empty for the sectors the analyst literature discusses most.
Four transfers from this literature seem most useful for a personal-lines book, in descending order of how well-grounded they are.
Selection, not just churn. Handel is the result to start from. If agents disproportionately take low-engagement customers out of inertia, the remaining pool is differently selected, and the question is whether the risk that moves is priced where it lands. Low-engagement customers are not necessarily low-risk, and high-premium customers are not necessarily high-margin, so the exercise has to model adoption, switching, coverage choice and expected claims jointly rather than treating the agent as a pure margin shock.
The obstruction step is already regulated. In Huang & Xiao's simulation, a phone-only cancellation path was the one seller lever that could not be repriced away. The real-market counterpart is the FCA's general-insurance pricing work, which treated "call us to cancel auto-renewal" as a remedy target and whose 2025 evaluation measures what removing it did. Insurers in other jurisdictions should expect that step to draw attention whichever side automates first.
Renewal pricing is another setting where delegation may change how well segmentation works. Loyalty-based pricing relies on renewal timing being decoupled from shopping. The Abdelghany results suggest delegation can weaken timing-based segmentation where it does sever that link, and the FCA's new-business/renewal parity rule is evidence of what happens to margins when the segmentation is removed by regulation instead. The analogy is useful; it is not a verbatim transfer, because an agent re-quoting at renewal is still constrained by what its principal has told it to accept.
Cheaper is not better coverage. None of the papers evaluate whether an agent's cheaper policy preserves cover. A premium saving that comes from a higher excess, an excluded peril or a lower limit is not a success, and "no claim during the trial" is not evidence the cheaper policy was adequate. Any internal evaluation should score agents on constraint violations and coverage adequacy, not quoted premium.
None of these is tested on insurance data. The Huang & Xiao design (paired frozen-vs-adaptive sellers, a fee only some customers can escape) is a reasonable template for an internal simulation, provided sellers are allowed more instruments than premium.
Proposed extensions, not findings of the reviewed papers. Most were raised in review and are consistent with the gaps in the evidence matrix. The thread through all of them: reliable, preference-aligned execution is a promising source of value, not yet an established universal winner, and whether it stays valuable depends on who the agent works for, who adopts, and what sellers do in response.