We are building consumer-facing agent workflows. Text extraction, language models and agentic interfaces that let a person complete an application or a transaction in a different way than the old form-and-button web has trained them to expect.
The technology works. We have been less sure the consumer is ready.
The 2026 data suggests we have been right to be cautious. Riskified's Agentic Commerce Pulse for the first quarter of 2026 found that a 55 percent majority of consumers were not comfortable with AI agents making purchases on their behalf, a quarter after 70 percent had said they were at least somewhat comfortable. That is not slow progress. That is a reversal. Bain's research found that 72 percent of US consumers have used AI in some form, but only 24 percent are comfortable using it to make a purchase. Thales' 2026 Digital Trust Index, surveying more than 15,000 people, found 77 percent are concerned about AI agents acting on their behalf online.
The pattern across all of these is the same. Consumers will use AI for recommendation, discovery and summary. They are deeply uncomfortable letting AI complete a transaction or take a consequential action in their name. The trust cliff sits at the moment of agent action.
So here is the design pattern we have arrived at, almost without realizing it, after enough consumer-facing engagements. We build dual-path.
The AI does its work in one lane. Documents come in, structured data gets extracted, fields get pre-populated, and the workflow advances quietly behind the scenes. In parallel, the consumer is given a fully visible form view of their own information, an edit path to correct anything the agent got wrong, and a confirmation step before anything consequential is submitted. The two paths converge at the moment the consumer commits.
The AI does the work. The consumer keeps the control. The trust cliff is respected because the consumer never crosses it.
This is not what the agentic AI keynotes describe. The vendor demos show frictionless one-shot completion with the agent acting on its own. That demo is technically possible. The consumer data suggests it is not yet what the consumer will accept. The dual-path pattern is the version that survives the trust gap, because it does not ask the consumer to cross it.
The cost of building dual-path is real. The interfaces take longer to design. The conversion funnel has one more confirmation step than the vendor would model. The system is also the one that ships and gets used, because it accepts the consumer as they are, not as the keynote says they will be.
For builders working in this space, take the dual-path question seriously before you ship. The trust gap is the actual product problem. The model is not the bottleneck. The interface that earns the consumer's trust at the moment of action is.
We are still building the agents. We are also building the bail-out point next to them. We will revisit when the cliff flattens.