Artificial intelligence is transforming the online shopping experience, such as by letting autonomous agents research products, compare prices and make purchases on consumers' behalf.
But even as businesses are racing to embrace agentic assistance in e-commerce, customers are reluctant to hand over control of their financial decisions to a machine.
To close this trust gap, according to the findings of Okta's Customer Identity Trends Report 2026, organizations must deploy modern customer identity and access management (CIAM) systems that balance security, convenience and consumer control.
Why organizations feel they must implement AI, and why they're falling behind
As in most industries, competitive pressure is driving consumer-facing businesses toward adoption of AI agents.
In a recent blog post, Gareth Davies, chief product officer at Okta's Auth0, reported that consumer brands are seeing up to 40% of their traditional search traffic shift toward third-party AI agents. The brands aren't complaining, because these agents reportedly convert shoppers into making purchases at three to five times the rate of conventional website visitors.
Yet implementation of consumer-assisting AI agents is lagging. Okta's research found that nearly 60% of organizations have action-taking or API-calling agents deployed or under development, which you would expect given the competitive pressure. Yet approximately 80% of these projects are behind schedule.
When senior organizational leaders are asked to identify the primary obstacles to deployment, 59% cite compliance, security reviews or governance. But software developers see another problem: 17.8%, nearly twice the rate of senior leaders, blame outdated identity infrastructure that cannot handle agent access patterns.
In other words, many of the developers feel that without suitable identity systems, organizations risk introducing insecure pathways and permission structures that may compromise customer information.
"To build lasting trust," writes Davies, "product teams must extend identity controls to the AI agents themselves."
The massive trust gap between human customers and AI agents
It's quite clear that consumer enthusiasm for AI has not kept pace with the business community's eagerness to deploy and capitalize on the technology.
Okta's survey found that only 38% of surveyed consumers said they would be comfortable delegating even just some decisions to AI agents. Only 6% said they would support completely autonomous actions, although younger respondents were significantly more comfortable with that than older ones.
However, customers become substantially more receptive if they are given the option to retain meaningful control. For low-risk transactions, the survey found, most consumers would let AI agents act on their behalf if they could approve decisions beforehand or reverse them afterward. (Yet 29% stood firm and said they wouldn't want an AI agent to take any action.)
Transparency also matters: 89% of consumers want control over their own personal information and visibility into its use.
To meet these expectations, shopping agents should offer consumers adjustable permissions, transaction approvals, activity histories and mechanisms for reversing actions.
However, security friction must also be considered. Some 74% of respondents said they might abandon a potential purchase if a website required them to create an account to view essential product information such as pricing or user reviews. A smaller but substantial majority (59%) would abandon the attempt if they had to create an account to make a purchase.
Why AI agents must be treated as a new class of identity
Traditional identity systems distinguish human users from machine accounts, which are mostly safe, predictable non-human identities (NHIs) with narrowly defined roles and tasks. AI agents, especially those that can independently access applications, invoke APIs and act on customers' behalf, require the creation of a third user category.
Like humans, AI agents are unpredictable, goal-oriented and eager to stretch boundaries. Like NHIs, they act extremely quickly and never tire. The old identity structures don't fit.
Because of this mismatch, treating agents like human users creates unforeseen risks. Most humans will never use all the access they're given, but granting excessive permissions to agents elevates the risks of credential theft and unauthorized transactions.
Nevertheless, only 22.4% of the organizations in the Okta survey who said they were developing AI capabilities also said they routinely assign agents unique identities with tightly scoped, time-limited credentials. The rest may be creating unnecessary risks.
To avoid such potential pitfalls, modern CIAM systems can link distinct agent identities to individual customers, restrict authorized actions and enforce granular runtime permissions.
External identity providers can also simplify deployment. Okta found that 34.4% of organizations that exclusively used external CIAM platforms already had AI capabilities in production, compared with 22.3% using alternative approaches. Similarly, 28.3% of external-platform users report meeting AI deployment schedules versus 18.1% of other organizations.
Agent-supported e-commerce will be accepted only when customers feel they can trust the technology acting on their behalf. By combining distinct agent identities, limited permissions, adaptive authentication and meaningful consumer oversight, organizations can safely accelerate deployment.
"Revenue shouldn't have to come at the expense of risk," writes Davies. "Making agentic commerce viable means applying runtime, fine-grained access controls that define exactly what an agent is allowed to access."
