Limitation: This content is identified as Perspective / Viewpoint. Evaluation of this content might be partial, and should be treated as indicative.
AI agents are beginning to research, compare and buy on people's behalf. When an agent visits a business, it reads a small part of what the business publishes and moves on; if the fact that would decide this customer's choice cannot be found or verified quickly, the business is dropped, even when it has the right product. This paper introduces AIVO Assist: a business's own expert that visiting AI agents can consult. The agent passes on what its customer needs; AIVO Assist judges the fit, answers with the deciding facts and their evidence, and says plainly when nothing fits. The paper's central argument is that such an expert cannot be built from product knowledge alone. Product detail answers "what is it?". An agent deciding for a customer needs decision context: why this option, for this customer, against these alternatives, and with what proof. That context can only be derived from measurement of how AI systems actually decide: which facts decide outcomes (AI Win-Rate), which facts the models hold but do not use (the Linkage Gap), which facts agents cannot find or verify (legibility), what evidence a claim needs (brand.context), how the models characterise a business's reputation and the limits they attach, and what causes a measured outcome, including beliefs left out of date by events (AIVO Prism). Measurement after deployment then shows whether the expert changed decisions, and traces answers to sales. The paper sets out the concept, the measurement loop, the principles the expert must follow, early evidence from controlled tests with AI agents, and a numbered list of the concepts it discloses. Concept working paper. Describes what each part of the system does and the principles it follows; implementation is not described. Early results are from AIVO, Inc.'s own test agents.
such an expert [a seller-side expert that visiting AI agents consult] cannot be built from product knowledge alone
no catalogue-built expert was tested against the proposed system
That context [decision context] can only be derived from measurement of how AI systems actually decide
tests of one measurement-led system do not establish exclusivity over alternative context-derivation approaches
On realistic websites, where the deciding facts were behind a booking engine or in vague copy: right recommendations rose from 47% (79 of 167) without AIVO Assist to 93% (157 of 168) with it.
the comparison supports the reported direction within a controlled fictional market using proprietary agents
Derived from the full evaluation — not a separate score.
Strengths
The proposal joins customer-fit judgments, evidence-bearing answers, abstention, and outcome measurement into an operational loop rather than treating the interface as catalogue search alone.
↳ Sections 3–5; Table 1
The expert made no difference when a clean page already supplied every deciding fact. The paper also reports an earlier refusal failure and a corrected attribution fault, helping identify where the proposed mechanism adds value.
↳ Section 6, Table 3
The system description specifies three fit outcomes, sourced statements, honest refusal, information parity, and precedence for independent evidence. These concrete rules make the intended behavior and safeguards understandable.
↳ Sections 3 and 5
Limitations
The tests compare the proposed expert with no expert, not with a catalogue-built expert or alternative context-building system. They therefore do not establish that decision context can only come from measurement.
↳ Abstract; Section 2.1; Section 6, Table 3; Section 9
The principal result comes solely from the company's own agents without released sessions, prompts, adjudication records, or code. Missing allocation and analysis details also prevent a full assessment of test execution.
↳ Scope of disclosure; Section 6
The catalogue-assistant comparison relies on asserted limitations, while much of the supporting programme is drawn from the company's own working papers. The limitations discussion does not explain how the absent catalogue-only comparison constrains the central thesis.
↳ Section 2.1; Section 4.3, Table 2; Section 6; References 1–9
Sections 3–5 define a coherent seller-side expert and connect its answers to a measurement and attribution loop. Table 3 supplies a relevant controlled demonstration and an informative null result when pages already contain the deciding facts. However, the comparison neither tests a catalogue-built expert nor substantively engages alternative approaches, leaving the Abstract's necessity claim unestablished. Missing source-selection and experimental methods further constrain execution quality, while proprietary agents and unreleased records limit independent verification.
Nabu’s assessment, alongside the field’s view.
Are you an author of this paper?
Limited2.7
Confidence mediumThe paper combines fit judgments, evidence sourcing, abstention, measurement, and attribution into an identifiable seller-side framework with a controlled demonstration. The demonstration does not establish the central necessity claim because it lacks a catalogue-built expert comparator.
The argument is traceable and reports informative null and adverse results, but its central comparison does not substantively engage alternative approaches. Missing source-selection and test-protocol methods further limit assessment of execution, while the proprietary single-source result remains independently unverifiable.
The progression, defined terms, and before-and-after tables make the proposed system readily followable. Categorical necessity language in the Abstract and conclusion nevertheless conveys more than the controlled evidence establishes.
The paper connects the proposal to agent-commerce developments and meaningfully acknowledges external-validity limits. Positioning remains concentrated in the authors' own programme and omits the decisive limitation that no catalogue-only expert was compared.
Caveats4 of 4 checks
The reported numbers and statistics are consistent with the stated results.
Corporate involvement and use of proprietary test agents are disclosed. The principal controlled result remains an unverified single-source finding without released underlying sessions, prompts, adjudication records, or code.
Conflict of interest: declared · Preregistration: not stated · Data availability: not stated · Code availability: not stated
16 references in manuscript 9 of 9 checkable references found in an index 7 have no canonical index record — counted, but not index-checkable 1 reference confirmed by manual review
References not verified automatically are resolved manually within 24 hours.
No retraction or expression of concern found in Crossref (includes Retraction Watch).
Sources: Crossref (incl. Retraction Watch) ✓
Where this paper’s evidence sits on the path from initial observation to real-world use.
The controlled fictional-market tests establish the Demonstrated readiness stop through a working proof of concept. Implementation is withheld, and performance with consumer agents, real businesses, competitors, and sales remains unestablished.
AI-generated, human-governed. Something look off? Contact us to request a review.
Keep reading
Other papers Nabu has evaluated blind, starting with the same field.
Kubicek, Caciolai, Peng, et al. · 2026 · Computer and information sciences
At 10M tokens, one attention layer of D-LSA prefiles 22.7× and decodes 4.6× faster than sparse attention in GLM-5.3-FLASH, and 72.4× and 42.5× faster than Full Attention in QWEN-3.8-27B.
SupportedLiu, Zhao, Xie, et al. · 2026 · Computer and information sciences
CLAIMNOV catches 97% of disguised papers at a 1.5% false-alarm rate
SupportedSon, Han, Koh, et al. · 2026 · Computer and information sciences
Decomposition delivers: peak KV working set of 14.3 MiB per query against 35.5 and 35.3 MiB for single-pass and retrieval-augmented baselines.
Supported