AI DevelopmentMethodology6 min readPublished September 16, 2026

One survey · 5,067 US adults · July 2026 · every figure with its base

What People Let AI Agents Access: 2026 Survey Numbers

A July 2026 survey of 5,067 US adults: 41% of AI users have tried an agent and 32% have let AI act without a final sign-off. Every figure with its base.

DA
Digital Applied Team
Research and practical guidance
Editorial dateSeptember 16, 2026
SourceMenlo Ventures · Morning Consult

On September 16, 2026 the venture firm Menlo Ventures published its second annual consumer AI report, built on a Morning Consult survey of 5,067 US adults fielded in July 2026. It is the first large public survey we have seen that asks what people let AI agents touch. Of AI users, 41% have tried an agent and 24% use one regularly. In the report's agent section, 36% have given an agent access to their email and 20% to a financial account.

This page is a reference. It tables more than 40 figures from the report, "2026: The State of Consumer AI" by Shawn Carolan, Amy Wu Martin and Sam Borja, and states the base for each one. The bases matter. A share of agent users is a share of about four in ten AI users, who are themselves 64% of adults. Quote a figure with its base or do not quote it.

Key takeaways
  1. 01
    Adoption barely moved; depth did.Use among US adults rose from 61% to 64%. Daily use rose from 19% to 25%, and 55% of AI users now pay for at least one product.
  2. 02
    A third of AI users have skipped the final check.32% of AI users have let AI act on their behalf at least once without a final sign-off. The report counts purchases, email workflows and agent use.
  3. 03
    Email is the most-granted permission.Email 36%, web browser 33%, messaging 31%, cloud storage 29%, calendar 27%, health apps 23%, financial accounts 20%.
  4. 04
    Trust now outranks ease of use.AI users rank accuracy (45%), trustworthiness (40%) and security and privacy (36%) above ease of use, which fell from 38% to 32%.

01The datasetAdoption and daily use

Three groups appear throughout. "US adults" is the full weighted sample. "AI users" are respondents classed as users from the tools they reported using in the past six months. "Payers" are AI users who pay for at least one AI product. The 2025 column comes from the firm's earlier wave of 5,031 adults in April 2025, and is filled only where the report prints a comparison.

Source: Menlo Ventures, 2026: The State of Consumer AI, September 16, 2026; Morning Consult survey of 5,067 US adults, July 2026. A dash means no 2025 figure was published.
MeasureBase20262025
Use AIUS adults64%61%
Use AI every dayUS adults25%19%
Use AI every dayAI users39%31%
Use a general AI assistant every dayUS adults22%
Spend more than an hour a day with AIAI users24%
Spend less than 15 minutes a day with AIAI users29%
Use AI more than a year agoAI users52%
Use AI, household income $100,000 or moreThat income group77%
Use AI, household income under $50,000That income group56%
Have not used AI in the past six monthsUS adults36%39%

The report's footnote gives the adoption figures to one decimal: 61.0% to 63.6%. It also notes that some of the 36% classed as non-users may be meeting AI inside search or email without recognising it. For wider usage data from other sources, see our AI usage statistics reference.

02AgentsAgents and autonomy

The report defines an agent as AI that can reason, make decisions, complete multi-step tasks and take actions in other systems. The agent, spending and sentiment questions are new in 2026, so none has a 2025 baseline. Two footnotes change how to read this table. The 72% awareness figure includes respondents who said they were unsure. The 32% no-sign-off figure includes people who used AI to complete a shopping purchase or an email workflow, as well as those who use agents.

Source: Menlo Ventures, 2026: The State of Consumer AI, September 16, 2026; Morning Consult survey of 5,067 US adults, July 2026. A dash means no 2025 figure was published.
MeasureBase20262025
Aware of AI agentsAI users72%
Have tried an agentAI users41%
Use an agent regularlyAI users24%
Have let AI act without a final sign-off, at least onceAI users32%
Pay for at least one AI productAgent users92%
Use AI every dayAgent users63%
Use an agentAI payers65%
Use an agentNon-payers13%
Pay $100 or more a monthPaying regular agent users21%
Pay $100 or more a monthPaying non-agent users5%
Use CodexAgent users35%
Use Claude Code or CoworkAgent users32%
Use PerplexityAgent users24%
Let AI complete a purchaseShoppers who use AI26%
Made a final purchase decision on AI's recommendationShoppers who use AI37%

The product rows are shares of agent users and respondents could name more than one. The report adds that three newer agents each reach between 10% and 16%. Agent users are a heavy-use group: 92% pay for AI against 55% of AI users overall, and 63% use AI daily against 39%.

03PermissionsWhat agents can reach

These seven figures are the reason this page exists. One caution on the base: the report presents them in its section on agent users and words them as access that "consumers have given agents". The text does not restate the denominator, and the report says its chart titles carry the base. We show them as presented and recommend citing them as shares of agent users only after checking the chart.

Access granted to AI agents, by account type

Menlo Ventures, 2026: The State of Consumer AI; Morning Consult survey, July 2026. Presented in the report's agent-user section.
Email
36%
Web browser
33%
Messaging apps
31%
Cloud storage
29%
Calendar
27%
Health apps
23%
Financial accounts
20%

The order is worth noticing. Email, the browser and messaging are the accounts through which most other accounts can be reset or reached, and they are granted most often. Health and financial accounts, which people think of as sensitive, come last. The gap between first and last is 16 points. People appear to rate the sensitivity of an account by its contents, not by what it can unlock.

04Money and trustPayment and trust

Who pays is a "select all" question, so those three rows sum to more than 100%. The report's text says 48% of payers cover costs "entirely themselves" while its footnote says 48% pay for at least one tool themselves. We use the footnote wording. The report also gives the typical payer's spend as $20 to $49 a month.

Source: Menlo Ventures, 2026: The State of Consumer AI, September 16, 2026; Morning Consult survey of 5,067 US adults, July 2026. A dash means no 2025 figure was published.
MeasureBase20262025
Pay for at least one AI productAI users55%
Spend more on AI than a year agoPayers46%
Spend less on AI than a year agoPayers12%
Spend $100 or more a month (60% of all spend)Payers14%
Pay for at least one tool themselvesPayers, select all48%
A family member or friend paysPayers, select all34%
An employer or school paysPayers, select allabout 20%
Rank accuracy as a selection factorAI users45%
Rank trustworthiness as a selection factorAI users40%
Rank security and privacy as a selection factorAI users36%
Rank ease of use as a selection factorAI users32%38%
Less likely to engage with content known to be AI-generatedUS adults36%
More interested in content known to be AI-generatedUS adults14%
Distrust AI-generated informationNon-users70%58%
Have privacy concernsNon-users76%
Would rather deal with a personNon-users83%
Prefer a human for a medical diagnosis (19% prefer AI)All respondents63%
Prefer a human for legal advice (21% prefer AI)All respondents59%
Say AI has improved their life (10% say worse)All respondents38%
Left out on purpose

The report's $40 billion global consumer spend figure, its two billion global users and its five predictions for 2027 are the firm's estimates and forecasts, not survey results. They are not in these tables. If you cite them, label them as Menlo Ventures estimates.

05ReadingWhat a product should ask for

Two figures together set the design problem. A third of AI users have already let AI finish something without a final check, and the accounts granted most often are the ones with the widest reach. At the same time, security and privacy now rank above ease of use when people choose a product. Users will grant access, and they are watching what you do with it.

Our reading for anyone building a consumer-facing agent:

  • Ask first for what the first task needs. A calendar agent asks for the calendar. Email and browser access come when a task requires them, with the reason shown.
  • Never ask for financial or health access at install. Only 20% report granting a financial account. A request at the door reads as overreach.
  • Make approval the default for actions that leave the account. Sending, paying, deleting and publishing get a confirmation until the user turns it off for that action type.
  • Show what was done. An activity log is the answer to the accuracy and trust rankings above.

For how coding agents handle the same question, see our census of coding-agent permission defaults and the comparison of default-deny and default-allow permission models. We design permission flows for client agents through our AI transformation practice.

06MethodMethodology

Methodology

Every figure on this page comes from one published report. We collected nothing ourselves and we did not see the survey instrument or the raw data.

Who ran it
Menlo Ventures commissioned the survey and Morning Consult fielded it. Menlo Ventures is a venture capital firm that invests in AI companies, including some named in its report.
Sample and dates
5,067 US adults, July 2026. Comparisons use the firm's 2025 wave of 5,031 US adults from April 2025. The report says 37 of this year's 41 activities were asked identically in 2025.
Weighting
The report describes the sample as nationally representative, with weights for education, gender, age, race and ethnicity, and region.
Base per section
Adoption rows are US adults unless marked. Agent awareness, trial and no-sign-off rows are AI users. Payment rows are payers. The seven access figures are shown as the report presents them, in its agent-user section, without a restated base.
As-of date
The report was read on September 17, 2026. This page is dated to the report's publication day; the collection date is stated here and not elsewhere.
Known limitations
One survey, one country, self-reported behaviour, commissioned by an investor in the sector. Respondents may not recognise AI inside other products, and may not agree on what counts as an agent. The 2025 wave ran in spring and the 2026 wave in July, which the report says dampens growth on school and work activities.

07Next stepPeople grant the widest keys first

Put it into practice

List every permission your agent requests and when it asks

Write down each permission your product asks for, the screen where it asks, and the first task that needs it. Move anything not needed for the first task out of onboarding. Then check which actions run without confirmation today. The survey says a third of AI users have accepted that. It also says they now choose products on trust before convenience, so the log and the confirmation are features.

Digital Applied

Design agent permissions people are willing to grant.

We map what an agent needs against what it asks for, build staged permission flows and approval defaults, and log every action it takes.

Staged permissionsApproval defaultsActivity logs
Your next project

Start with the permission inventory

  • List each permission and its first use
  • Move non-essential asks out of onboarding
  • Confirm before send, pay, delete, publish
Questions and answers

Using these figures

Cite the original: Menlo Ventures, 2026: The State of Consumer AI, survey of 5,067 US adults conducted with Morning Consult, July 2026. Always include the base, for example '41% of US AI users have tried an AI agent', not '41% of Americans'.
Digital Applied newsletter

Deep dives on AI, marketing and development.

Practical guides and fresh insights by email. No recycled takes.

Related dispatches

Continue reading

AI Development

Replaying an AI Agent Run: What Must Stay the Same?

Record the inputs, tools and environment behind an AI agent run. Separate trace playback from fresh execution with a practical reproducibility worksheet.

September 12, 2026 · 6 minRead
AI Development

After AI Context Compaction, Which Instructions Survive?

Check whether an AI agent follows the right instructions after context compaction. Use behavioral probes for task scope, permissions, evidence and progress.

September 12, 2026 · 6 minRead
AI Development

Parallel AI Agents: Which Resource Limits Still Apply?

Map the shared limits behind parallel AI agents. Check API quotas, worker capacity, file access and review queues before increasing simultaneous work.

September 12, 2026 · 6 minRead
AI Development

Changing an AI Agent’s Priorities During a Long Task

Change an AI agent’s priorities while work is running. Choose what pauses, continues or queues, and verify pending actions before switching direction safely.

September 11, 2026 · 6 minRead
AI Development

AI Agent Shopping: Who Is Responsible After the Agent Pays

Map approval, delivery, cancellation, returns and refunds after an AI-assisted purchase, with the Muse-specific Link protection limits and evidence.

September 8, 2026 · 6 minRead
AI Development

Who Owns Unfinished Work When an AI Agent Hands It Off?

Keep agent handoffs accountable with an explicit recipient, acceptance check and remaining-work record. Separate sending a task from accepting ownership.

September 5, 2026 · 4 minRead