The use cases for AI assistants are expanding from information retrieval to writing, planning, comparison, transaction preparation, and after-sales support. To distinguish between "asking AI to answer" and "authorizing AI to act," we compared research published by OpenAI, NBER, Visa, Accenture, Adobe, and YouGov between 2025 and 2026 to observe changes in user intent and authorization boundaries. As the survey regions, sample sizes, and classification methodologies vary across these studies, the data is used to compare behavioral structures and development trends, and not as year-over-year results for the same population.
Research Background
Historically, user interaction with AI was primarily understood as Q&A: input a question, receive an explanation, and then the user judges and executes on their own. As AI assistants now begin to support tasks like copy generation, itinerary planning, product comparison, price negotiation, subscription renewal, and purchasing, the research focus has also shifted from "whether users use AI" to "how much execution power users are willing to delegate to AI."
This distinction relates to how brand information enters the AI usage process. Answer-based scenarios focus on whether facts are clear and information can be retrieved; decision-based scenarios require product parameters, prices, reviews, and suitability conditions to be comparable; action-based scenarios further involve inventory, payment, delivery, returns, and human fallback. Understanding the degree of user authorization helps avoid lumping all AI interactions into a single type of search behavior.
AI Interactions in 2025 Were Still Dominated by Asking and Decision Support
In 2025, asking remained the most prevalent intent in consumer conversations on ChatGPT, but messages requesting AI to produce results or complete specific work already accounted for 40%. According to OpenAI's "How people are using ChatGPT" released in September 2025, and NBER's working paper "How People Use ChatGPT" published the same year, 49% of messages were classified as Asking, 40% as Doing, and 11% as Expressing; approximately three-quarters of conversations centered on practical advice, information seeking, and writing.
Asking refers to users seeking information, explanations, advice, or decision support from AI—for example, understanding a policy, comparing options, or asking how to proceed next. In this case, AI provides material for judgment, but the user still decides and acts. Doing is outcome-oriented, including drafting content, organizing plans, writing code, converting formats, or completing other explicit tasks. Expressing mainly includes personal expression, reflection, and non-task-oriented communication.
Doing reaching 40% indicates that AI is no longer just an information gateway. Users are asking it to generate ready-to-use output or take over some of the work that previously required manual processing. However, these tasks often still rely on continuous instructions, human review, and final confirmation, and thus cannot be directly equated with autonomous agents. The analyses by OpenAI and NBER are based on actual conversation classifications, which are closer to the real usage structure at the time than simply asking users if they intended to use a feature in the future.
| Intent Type | Definition | Message Share | Representative Tasks |
|---|---|---|---|
| Asking | Obtaining information, explanations, advice, or decision support | 49% | Information retrieval, option consultation, problem explanation |
| Doing | Requesting AI to produce results or complete explicit work | 40% | Writing, planning, programming, content conversion |
| Expressing | Personal expression, reflection, or non-task communication | 11% | Sharing feelings, discussing viewpoints, open-ended conversation |
Data in the table is from studies by OpenAI and NBER on ChatGPT usage up to mid-2025, publicly released in September 2025.
'Delegation' Intent in 2026 Shows Layered Authorization
Consumer research in 2026 further shows that users are not simply choosing to "allow" or "disallow" AI delegation, but rather forming an authorization ladder based on execution scope, decision-making power, and payment rights. According to Accenture's "Talk to My AI Agent: The New Rules of Brand Value" released in June 2026, a survey of 25,590 consumers across 16 countries found that 74% were willing to let an AI agent handle price negotiation, complaints, renewals, or replenishment strictly according to instructions.
When the agent is given some decision-making latitude, acceptance drops to 32%. This group of consumers is willing to let the agent choose purchase options within pre-set budgets and preferences, but payment is still completed by the user. The proportion who would allow the agent to autonomously initiate and complete purchases is 9%. The decline from 74% to 32% to 9% reflects that willingness to authorize contracts as AI autonomy increases.
Users are typically more willing to delegate tasks that are repetitive, time-consuming, rule-based, and can be withdrawn or reviewed—such as comparing renewal options, compiling complaint materials, or replenishing items according to a list. When it comes to brand selection, budget allocation, and payment, users tend to retain a confirmation step. Therefore, "letting AI do things" does not mean giving up control. The changes seen in 2026 are closer to constrained agency: AI executes, while humans define the boundaries and retain key decisions.
| Authorization Level | Specific Behavior | Control Retained by User | Acceptance Rate |
|---|---|---|---|
| Instruction Following | Handling price negotiation, complaints, renewals, or replenishment | User defines task and method | 74% |
| Decision-Making Within Boundaries | Selecting purchase options within budget and preferences | User sets conditions and completes payment | 32% |
| Autonomous Action | Autonomously initiating and completing purchases | User delegates selection, initiation, and payment | 9% |
Data in the table is from Accenture's June 2026 survey of 25,590 consumers across 16 countries. The rates represent respondents' acceptance of the given authorization scenarios, not indications that these behaviors are already widespread.
Task-Oriented Intent Has Extended Across Multiple Consumer Journey Stages
User authorization for AI tasks is not confined to the payment stage but is distributed across discovery, evaluation, cart & checkout, and after-sales. According to Visa's Agentic Commerce Consumer Research conducted in 2025 in the US, Australia, and New Zealand—covering 1,600 respondents in the US, 1,600 in Australia, and 500 in New Zealand—the average proportions of respondents envisioning AI shopping agents playing a substitute or assistive role in the four consumption stages were 73%, 69%, 62%, and 64%, respectively.
The 73% in the discovery stage involves searching for products, browsing websites, and filtering through social media and advertising information. The 69% in the evaluation stage includes comparing different retailers, reading reviews, organizing candidate products, and identifying price differences. These two stages are closely related to traditional information search, but AI is no longer just returning web pages—it is synthesizing scattered information into decision-ready results.
The proportion for the cart & checkout stage is 62%, which involves adding items to the cart, filling in shipping details, and applying discounts. The after-sales stage is 64%, including tracking logistics, processing returns, and registering warranties. This means task-oriented intent can span the entire purchase cycle: the same user may be willing to let AI search and compare, and also prepare transaction materials, but still retain human confirmation before submitting an order or making a payment.
These proportions are derived from scenario-based surveys and reflect the scope where consumers believe AI can replace or assist in relevant steps; they should not be interpreted as actual transaction penetration rates. What they reveal are task boundaries: user authorization for AI varies with context, risk, and reversibility, rather than remaining constant across the entire consumer journey.
Trust, Transparency, and Human Fallback Affect Delegation Willingness
Whether users move from asking to delegation depends not only on feature availability but also on whether the agent's identity is clear, its actions are explainable, and whether a human can take over when problems arise. Adobe's Asia-Pacific "AI and Digital Trends Report" published in May 2026, based on feedback from 7,000 consumers and business decision-makers, found that 42% of consumers are willing to interact with brand AI agents.
Meanwhile, 35% of consumers have not yet considered using a personal AI agent, and 16% explicitly say they would not accept one. There are also differences across interaction formats: 46% are willing to have their personal agent communicate with a brand's human representative, while 36% accept personal agents directly interacting with brand agents. Scenarios with higher human involvement received greater acceptance, indicating that users still prefer complex issues to be resolved through confirmable, accountable communication channels.
Adobe's research also shows that 26% of respondents value the ability to transfer to a human at any time, and 24% value clear labeling of AI identity. Human fallback handles exceptions, while identity labeling lets users know who they are interacting with and which judgments are made by the system. Together, these two mechanisms reduce uncertainty in authorization and make it easier for users to define accountability.
YouGov's "U.S. Web Search and AI Report 2026" published in July 2026 also identifies trust as a key issue in studying AI search adoption, serving as qualitative corroboration. It is important to distinguish that "willingness to interact" represents acceptance, not sustained usage; functional interest, trial behavior, and long-term authorization are different stages and cannot be combined.
Trend Observations and Industry Implications
AI usage intent has not made a wholesale leap from "asking" to "delegation" but is instead gradually expanding authorization on low-risk, reversible, and clearly defined tasks.
Cross-study comparison reveals a continuous but non-linear structure. ChatGPT conversation data from 2025 shows that asking and decision support still occupy a significant position, while execution-oriented interaction has already reached a considerable scale. Consumer research from 2026 further breaks down "doing" into instruction-following, decision-making within boundaries, and autonomous action, making varying degrees of control observable.
Therefore, users are not uniformly shifting from search to agents, but rather adjusting the level of authorization across different tasks. Information retrieval, option suggestions, product comparison, transaction preparation, and after-sales processing are gradually forming a continuous AI usage process, but the acceptance conditions differ at each step. The more explicit the task and the easier the results are to verify, the more willing users are to delegate to AI; when it comes to payments, identity, rights, or irreversible decisions, human confirmation still holds a prominent place.
For brands, the scope of AI visibility observation also needs to extend accordingly. Early questions focused on whether the brand appears in answers and whether its information is cited; as scenarios move into comparison and action, it is also necessary to assess whether product parameters can be distinguished by machines, whether pricing and service conditions are consistent, whether inventory and after-sales information are verifiable, and whether agents can recognize authorization boundaries.
GEO audits, content strategy, and data analysis therefore need to differentiate between answer-based, decision-based, and action-based queries. Using only a single mention rate is inadequate for explaining the availability of brand information in product comparison, transaction preparation, or after-sales tasks. Different query types should be examined separately for factual completeness, comparison dimensions, evidence sources, and action conditions.
Research Implications
Brands can start by dividing the likely AI queries from their target users into three tiers: information acquisition, solution comparison, and task execution. The information acquisition tier focuses on whether basic facts are clear; the solution comparison tier focuses on whether product differences, applicability conditions, and evidence can be processed side-by-side; the task execution tier requires supplementary dynamic information such as pricing, inventory, processes, delivery, returns, and human support.
Next, brands can check the availability of their information in answer, recommendation, and action scenarios separately, and record differences in how different AI platforms present the same type of query. These observations can be linked to the GEO audit, content strategy & distribution, and data analysis components of the AIPO framework to identify information gaps and guide future updates, but they do not imply that specific platforms will necessarily cite, recommend, or execute brand information.
As AI usage extends from information queries to comparison and task execution, brands need to assess the completeness of their information across answer, recommendation, and action scenarios. You can discuss monitoring scope with the aipogeo team, combining GEO audits, content performance, and trend data to understand current visibility.
Related Questions
In 2025, were users primarily asking ChatGPT questions or completing tasks?
According to studies published by OpenAI and NBER in 2025, Asking accounted for 49%, Doing for 40%, and Expressing for 11%. Asking still holds a higher proportion, but outcome-oriented tasks already constitute a significant portion.
What are the differences between the three types of AI usage intent: Asking, Doing, and Expressing?
Asking involves obtaining information, explanations, and advice; Doing involves requesting AI to generate results or complete tasks; Expressing focuses on personal expression and open-ended communication. The main difference lies in whether the user wants judgment material, usable output, or conversational engagement.
Which daily tasks are users willing to delegate to AI agents?
Users are more willing to delegate tasks that are rule-based, repetitive, time-consuming, and can be reviewed—such as price negotiation, complaint compilation, renewals, replenishment, product comparison, and logistics tracking. Authorization typically decreases for tasks involving payments or irreversible decisions.
Why do users accept AI executing tasks but are unwilling to let AI make purchases autonomously?
When following instructions, the task scope and accountability boundaries are relatively clear; autonomous purchasing simultaneously involves choice, budget, and payment authority. In Accenture's 2026 survey, acceptance dropped from 74% for instruction-following to 9% for autonomous purchasing.
How will AI shopping agents affect product search, comparison, and after-sales processes?
AI shopping agents can participate in information filtering, retailer comparison, cart preparation, discount application, logistics tracking, and returns. Human confirmation points may still be retained at different stages, so assisted use should not be directly interpreted as fully autonomous transactions.
How should brands distinguish between answer-based, decision-based, and action-based AI queries?
Answer-based queries focus on facts and explanations; decision-based queries focus on differences, evidence, and applicability conditions; action-based queries involve pricing, inventory, processes, and after-sales rules. Tiered monitoring helps identify where brand information gaps may exist.