Okta Says 80% of Agentic Commerce Projects Are Stalled — and Buyers Want a Say Before an Agent Acts

Okta calls 80% of agentic commerce projects stalled in a new post — and its own survey shows buyers will not hand over the keys. The company behind Auth0 surveyed 5,250 shoppers and 910 developers and leaders, and just 6% of shoppers are fine with an AI deciding and acting fully by itself. Six in ten would never allow an agent to make a purchase over $100.

Okta just put a number on the AI shopping stall

Gareth Davies runs product for Auth0, Okta's customer identity arm. In a post dated November 4, 2026, he argued that four in five AI efforts in customer journeys are late. The post's headline goes further and calls those efforts stalled.

The numbers come from a study Okta calls its 2026 Customer Identity Trends Report. One poll, run by Statista, reached 5,250 shoppers in seven markets: the US and Canada, the UK and Germany, plus Japan, Australia and New Zealand. Rep Data fielded the other, with 354 senior leaders and 556 developers among the 910 people surveyed. Fieldwork ran in July and August 2026.

That 80% rests on one question. Only 20.3% of the 910 developers and leaders said their company was shipping AI as fast as it planned. Okta's summary turns that into 80% of transformation efforts behind schedule.

Turning to demand. Just 6% of consumers would let an AI make the final call and operate fully without oversight. Meanwhile, 52% described the speed at which AI capabilities appear in apps and websites as too fast, compared with 35% who considered it about right.

Put those two together and you have the shape of the problem for anyone selling online. The builders are behind (20.3% at target pace). The buyers are interested, but they want a say first (6% want full autonomy). That is my read of the two numbers, not the report's.

The detail buried in the middle is the one to remember

Okta tested 11 actions and gave shoppers four answers to pick from: the AI acts alone, they can undo it, they approve it first, or they never allow it. Not one action drew unconditional trust above 9%. Once approval or undo was on the table, a majority said yes for nine of the 11.

The two exceptions were the tasks closest to money. Six in ten shoppers said they would never let an agent move money between accounts or buy anything over $100. Even under $25, 48% said they would never allow it.

Approval beat undo every time. On each of the 11, more shoppers chose to approve an action up front than chose to undo it later. Davies summed it up in his post: shoppers "want AI assistance, but they demand control" — meaning they want a say over what an agent does.

The report's 38% figure is worth reading carefully. In the body of the report, 38% said they would be at ease letting AI make some calls within two years. The summary describes that same 38% as people ready to delegate to AI agents today — different timing, same number.

That is the whole ballgame for a dealer or a contractor. Your customer is not refusing the technology. They are refusing to hand over the keys.

Where the 40% number came from, and why I would slow down on it

Here is the line that got passed around. Davies' post says up to 40% of a brand's search traffic is moving to third-party AI agents, and that those agents convert three to five times better than a person browsing a storefront.

That number is not in the survey. It came from Oktane, Okta's annual conference, which ran September 22 to 24, 2026. The post names no brands, no time period and no method behind the figure.

Okta's September 2 product announcement used a different number. It cited Bain, which projected that agents could drive up to a quarter of US e-commerce within four years.

Other data points do not line up with that. A Microsoft Clarity study from November 2025 put signups from AI referrals at 1.66%, against 0.15% from search. Walmart said in March 2026 that ChatGPT purchases converted at a rate three times lower than click-out links. On November 3, 2026, Shopify said AI-search orders grew 3.3 times over the past year, from a base it called small and gave no figure.

One more thing worth knowing before you quote the 40% to your partner. Okta sells Auth0, an outside customer identity platform, and the report's strongest finding favors that kind of vendor: companies that use only an outside platform do better on AI than everyone else. The report does not publish a margin of error, a weighting method or the size of each group.

My read: the direction is real, the timeline is not. Treat the 40% as a number shared on stage at a conference, not one you build a budget on.

The other half of the problem: nobody can tell an agent from a bot

Davies argues security teams must flip an instinct they have held for years. He writes, "Now, we have to let the good ones in."

That sounds simple. It is not. In October, Arkose Labs said 97.7% of the agent sessions it judged legitimate had no verified identity. Web Bot Auth is a signing standard that Mastercard's Agent Pay and Visa's Trusted Agent Protocol both took up through Cloudflare in October 2025, and it is spreading slowly. In Resemble AI's own test, six bot tools caught only 3.1% to 14.8% of disguised agents.

For an owner, this is the problem you already have with your forms, just louder. About a third of organizations say bots inflate engagement numbers or abuse promotions. ADEX data from October 27, 2026 found bots posing as humans in 64% of the AI-related ad fraud cases it flagged.

So the traffic is getting weirder on both ends. Some of it is a bot you do not want. Some of it is an agent shopping for a real person. You cannot tell which is which.

What this means if you sell HVAC, roofs or cars

Everything from here is my read, not the report's.

Picture a home services company running ads for a furnace replacement. A homeowner taps the ad and lands on a page that wants name, email, phone, address and project details before anyone will give a ballpark. Okta's report says a sign-up form shown before prices turns away three in four shoppers, by their own account, while the same ask at checkout drives away fewer. The wall is in the wrong place. Move the ask to the end, after you have given them something.

Picture a car dealer with a chat assistant that books test drives. If it books without asking, you are asking a shopper to accept the exact thing only 6% say they want. If it offers two slots and lets them tap one, you are offering the approval step most shoppers picked. Walmart already made that call. When it rolled out Agent Profiles on October 21, 2026, approval became the default for each purchase, with optional caps like $50, $100 or $200.

Picture either business running a quote form on a Saturday. Roughly one in three organizations reports bots puffing up engagement metrics or abusing promos. If you count raw form fills as leads, your best-looking week may not be people.

And notice what shoppers say would make them comfortable. The top answer was setting permission levels, at 36%, followed by alerts that let them approve actions and the option to reach a human, both at 34%. Another 29% picked the choice that they do not want AI acting for them — a bigger share than wanted a live view of what the AI is doing. Not one measure won a majority, which tells me there is no single fix. There is a stack of small ones.

How to use this week

1. Time your own price request. Open your site on your phone and see how long it takes to reach a price or a ballpark. If a form stands in front of it, that is fix number one. Three in four shoppers say a form ahead of the price would stop them.

1. Count the fields on your quote form. The report contrasts long sign-up forms against one-tap passwordless, passkey or social login. Cut every field you do not actually count. If you run a customer portal, the report says the clear favorite is signing in with an existing Google, Apple or Facebook account.

1. Put an approval step in front of anything you automate. Okta found approval outperformed undo on all 11 actions. If your assistant books, orders or commits anything, make it propose first and let the customer tap yes.

1. Keep the first transaction small. Six in ten consumers would never let an agent make a purchase over $100, and 48% would never allow one under $25. If you sell a small first job — a diagnostic, a tune-up, a detail — that is where you test this.

1. Check what your ad reports are actually counting. Okta found a third of organizations reporting bots that pad engagement metrics or abuse promotions. Pull last month's form fills and call the first ten. You will know in twenty minutes.

1. Write your approval rules on one page. What the AI can do alone, what it must ask about, and who on your team can override it. 52% of senior leaders called AI "a must-win capability" for competing over the next two to three years. That is a reason to be deliberate, not fast.

1. Run one test for two weeks. Pick the single step above with the biggest gap and change only that. Then compare booked appointments, not clicks.

Three things you can copy today:

Home services: Try a prompt like: Write three short text replies for a homeowner who asked for a furnace quote. Each offers two appointment windows and asks them to reply with a number. Keep each under 40 words.

Car dealer: Try a chat prompt: Write a reply for a shopper asking about a used SUV. Offer two test drive times, ask them to pick one, and do not ask for a phone number until they pick.

Phone script, either business: Before I go further, do you want me to text you the numbers, or walk you through them now? Either way, you decide what happens next.

Mistakes to avoid

• Do not gate the price. Three in four consumers say they would walk away from a form in front of the price. You are paying for that click either way.
• Do not let a bot commit. Only 6% of consumers want AI making final decisions and acting completely on its own. Approval is the feature, not the friction.
• Do not read the 80% stalled figure as 80% failed. The survey measured how each respondent saw their own organization's pace, not a headcount of projects, and it covered four industries, not just retail.
• Do not treat AI-referred traffic as free money. Walmart said ChatGPT purchases converted at a rate three times lower than click-out links, and Shopify's growth came from a base it called small.