Trust Is the Product: Menlo's 2026 Consumer AI Data for Builders

This year, AI users in Menlo Ventures’ consumer survey put accuracy, trustworthiness, and security and privacy ahead of ease of use. Accuracy led at 45%. Ease of use slipped from 38% in 2025 to 32%, behind all three.

That result comes from Menlo’s 2026: The State of Consumer AI, by Shawn Carolan, Amy Wu Martin, and Sam Borja, published September 16. It’s based on a Morning Consult survey of 5,067 US adults, fielded in July 2026. The headline is that adoption barely moved, from 61% to 64% of US adults, while existing users went deeper. The closing section sums it up: “Consumers adopt what’s useful, and they keep what they trust.”

We agree. This post covers the part a report written for founders leaves open. For an AI product, trust isn’t brand sentiment. It comes from engineering decisions: what you can show is accurate, what data you request, and which actions run unattended. Each one can be designed, tested, and measured. We map Menlo’s findings onto those decisions, flag where we read the data more cautiously, and end with a step for this week.

Menlo marks its portfolio companies with an asterisk. Many of the report’s product examples carry one, including Claude, Wispr Flow, Lovable, and Suno. Menlo has backed Anthropic since its Series C and led its Series D, so read the report’s profile of Claude users with that in mind.

Key Takeaways

• In Menlo’s July 2026 survey, AI users ranked accuracy (45%), trustworthiness (40%), and security and privacy (36%) above ease of use (32%).
• Menlo’s three trust questions map to three engineering surfaces: evals, scoped permissions, and action boundaries.
• Give every agent action a risk tier before it ships. Fewer users hand over each successive step, from comparing prices (50%) to completing a purchase (26%).
• Pick first jobs where checking the output is cheap and a mistake is cheap. Expand only when usage data says users accept the work.
• Instrument trust: edit rate, override rate, undo rate, and scope revocations tell you more than sign-ups.

What changed in Menlo’s 2026 data

Consumer AI grew by depth, not reach. Menlo puts US adoption at 64% of adults, three points above 2025. Daily use rose from 19% to 25% of adults. Among AI users, 55% now pay for at least one product, and 41% have tried an AI agent.

For builders, the bigger shift is what users value. Menlo ties it to the stakes of the work. Nobody needs much trust to get a dinner idea. Letting a system read your bank statements or reply to your landlord is different. As the work gets riskier, users stop asking whether the product is easy and start asking whether it’s right.

Holdouts moved the other way. Among people who haven’t used AI in six months, 70% now distrust AI-generated information, up from 58% a year earlier. Asked to choose, 83% of them prefer dealing with a person. Menlo reads this as polarization: users go deeper, while non-users dig in. For a product team, both groups send one message. Being wrong now costs more than being hard to use.

How far to lean on these numbers

The direction is solid. Some specific figures deserve more care than a pitch deck will give them.

• The payer jump isn’t like-for-like. Menlo’s 2025 report derived its paying share, “roughly 3%” in the 2026 footnote’s words, from a market estimate. The 2026 figure of 55% is a survey answer from US AI users. Menlo’s footnote flags the different basis.
• The $40 billion market is an estimate. Menlo’s methodology says spending questions are new in 2026, with no 2025 baseline. Read “spend tripled” as a direction, not a measurement.
• Adoption is the clean comparison. Menlo says the 61.0% to 63.6% figure used identical questions in both years.
• It’s all self-reported. “I used an agent” and “I’d prefer a human” are what people say, not logged behavior.

None of this undercuts the trust finding. It does mean you should use the report to choose what to measure in your product, not to forecast your revenue.

Three trust questions, three engineering surfaces

Menlo frames consumer trust as three questions: “Can they trust what it tells them? Can they trust it with sensitive information? Can they trust it to act on their behalf?” The report treats these as market conditions. We’d treat them as a design checklist, because each one maps to a different part of the system.

Can I trust what it tells me? Make accuracy visible

Accuracy topped Menlo’s list, and it’s the one trust property you can test before launch. That means evals scored on your users’ real jobs, run on every model or prompt change. If you don’t have them yet, our guide to designing evals for AI agents starts from real failures and works up.

Pre-release evals aren’t enough, because users see your product, not your test set. So measure trust in production too. Menlo credits user trust in Wispr Flow, a Menlo portfolio company, to “the lowest edit rate in dictation.” That’s Wispr’s own metric, and we found no independent measurement behind the ranking. The idea generalizes well, though. How often do users edit, regenerate, or abandon an answer? That rate is a running accuracy eval scored by your users, and it’s free to collect.

Two design choices help users check the work themselves. Show the sources behind a factual answer, so checking costs a click. And let the system say “I don’t know” when retrieval comes back thin. An abstention costs you one session. A confident wrong answer costs you a user, and in health or money, it can cost them far more.

Can I trust it with my data? Ask for less, later

Agents need access to be useful, and Menlo shows access granted unevenly. Consumers have given agents their email (36%), browsers (33%), messaging apps (31%), cloud storage (29%), and calendars (27%). Health apps (23%) and financial accounts (20%) trail. Everyday tools get connected first.

Privacy is also where specialized products lose. Among people who use only general assistants, just 11% said they weren’t aware of a specialized product that fit. More cited privacy worries about lesser-known tools (27%) or not knowing which tool to trust (21%). The niche app can win on quality and still lose on the consent screen.

Menlo tells founders to “start with the data your core job requires” and widen access later. Two standard practices put that into code. First, request scopes when the feature that needs them is first used. Google’s OAuth documentation calls this incremental authorization and recommends requesting “access to user data in context” (Google Identity). Second, give each tool only the permissions its job needs. OWASP names “excessive permissions” as a root cause of Excessive Agency, one of its top risks for LLM applications.

Then make revocation as easy as the grant, and log both. A spike in revocations after a feature launch is a trust regression, and you want to see it that week.

Can I trust it to act for me? Tier every action

Menlo’s shopping data shows how much authority users hand over. Read as a ladder, it’s our design model rather than a measured sequence.

Fewer users hand over each successive step. Half compare prices with AI, while about a quarter let it buy. Menlo also found that 32% of AI users have let AI act at least once without a final sign-off. That’s more than the 24% who say they use an agent regularly. Many users don’t call a sent email or a finished checkout “using an agent.” So your product may already take actions its users don’t track.

The engineering response is to give every action a risk tier before you give the agent the tool. Sort actions on two axes: can it be undone, and what does a mistake cost? Here’s how that looks for an assistant that handles email, calendar, and shopping.

Action	Can it be undone?	Cost of a mistake	Control
Read the calendar	Nothing to undo	Low, but it’s still private data	Scoped grant, logged
Draft a reply	Yes, the user sends it	Low	Run unattended
Send an email	Only during a delay	Medium	Send delay with undo
Buy within a return window	Yes, with effort	Medium	Preview, then confirm
Pay a bill or move money	No	High	Confirm every time, or don’t ship

Low tiers can run alone. High tiers get a preview and a confirmation. That matches vendor and industry guidance. Anthropic’s computer-use documentation advises asking a human to confirm decisions with “meaningful real-world consequences”, such as financial transactions. OWASP recommends human approval for high-impact actions too.

The undo window does a lot of work in the middle tier. A short delay before an email goes out, or a cancel window before an order is placed, makes many final actions recoverable. Then track how often users override or undo. A rising undo rate on one action shows the agent overreached there, well before support tickets do.

Rank first jobs by cost to verify and cost of error

Choosing a first job is a two-cost decision. One cost is the user’s effort to check the output. The other is the damage when the output is wrong. Ship first where both are low. Climb toward riskier jobs once your telemetry shows users accepting the work below.

You can lower the first cost on purpose. Show diffs instead of rewrites. Show previews instead of taking actions. Attach citations instead of making bare claims. Each of these makes checking faster, so a job can move down a tier without the model getting any better.

Menlo’s healthcare data fits this frame. AI use for navigating healthcare rose from 16% to 23% of the people doing it. Menlo says the growth came at the edges: symptom research, appointment prep, and reading results. Those are cheap to check and cheap to get wrong. The expensive judgment stayed human. In Menlo’s survey, 63% would rather have a person make a diagnosis, while 19% would pick AI.

The frame also splits low adoption into two problems. Menlo asks founders whether people avoid a product because it isn’t good enough, or because they don’t want AI involved at all. The first is an eval problem: measure the gap and close it. The second is a scoping problem. Start with a smaller, safer slice of the same job.

Measure depth, not sign-ups

Menlo argues that engagement depth says more about product-market fit than sign-ups. It suggests sessions per week, minutes per session, and use cases per user. We’d add the trust telemetry from the diagram above: edit rate, override and undo rate, scope grants and revocations, and handoffs to a human. Those show why depth is growing or stalling.

Read the money data with one caution. Half of paying users use AI daily, against 26% of those who don’t pay. Agent use splits more sharply: 65% of payers, against 13% of non-payers. That’s a correlation from one survey. It doesn’t show that adding agent features creates payers. People who already rely on AI may simply both pay and try agents. Run your own experiment before you build a pricing roadmap on it.

Depth also has a cost side. The 14% of payers who spend $100 or more a month account for 60% of consumer spend. Menlo says these big spenders are more than 2.5 times as likely as the average American to be regular agent users. Our inference is that they’re also your most expensive users to serve. Menlo reports that plans are tiered, with usage billing once heavy users outrun their allowance. Already, 15% of payers mix a subscription with usage charges. Before you price a heavy tier, know your cost per successful task. Our post on right-sizing AI spend covers how to measure it.

Where we’d push back

Proactivity isn’t a shortcut around consent. Menlo praises Instinct, a proactive assistant, because a new user “doesn’t have to actively decide if they trust it.” That works for suggestions. For actions, skipping the decision is the risk itself. A proactive agent should propose, explain, and wait for approval on anything above the lowest tier. It earns autonomy per action type, not per user. Menlo’s other proactive example, Town, takes that line. Its homepage says it “never sends emails or calendar invites without your approval.” That’s the default we’d copy.

Design the handoff to a human. Menlo asked about nine situations, and people picked a human expert over AI in every one. Even for learning a new skill, the split was 44% for a person and 28% for AI. A product that can pass a case to a person cleanly, with context attached, earns more trust than one that can’t. Build the handoff path as carefully as the happy path.

More autonomy isn’t always more value. If an action is rare, irreversible, and hard to verify, automating it may never pay back. A good draft plus a one-click send is often the better product. We make the same argument about agent architecture in complexity is easy, reliability takes design.

Common questions

Does Menlo’s data apply to B2B products?

Partly. The survey covers US consumers, but the trust mechanics transfer. Menlo notes that a work email in a consumer sign-up can lead to enterprise growth. So one person’s trust decisions can become the first step of a company deal.

What’s the first trust metric to add?

Edit or regenerate rate on your main output. It needs no new UI and it tracks perceived accuracy daily. Add override and undo rate as soon as your product takes actions.

Should every agent action need a confirmation?

No. Confirming everything trains users to click through. Tier actions by reversibility and cost of error, run the low tiers alone, and save confirmation for steps that are costly or hard to undo.

Where to start this week

Menlo says the winners will treat trust as something “earned in low-stakes work, verified cheaply, and extended deliberately.” We agree, and each of those three verbs is work an engineering team can own.

Building an agent and unsure where its boundaries belong? Moonlight AI helps teams design the evals, permissions, and action controls that make it trustworthy.

Your step for this week:

1. List every action your AI product can take, from reading a file to sending a message.
2. Tag each one with two labels: reversible or not, and the cost of a mistake.
3. Find the irreversible, high-cost actions that run without a preview. Add a confirmation or an undo window to the worst one.
4. Start logging edits, overrides, and undos, so next month you’ll know whether users trust the change.