Are companies backing away from AI? Some are, and the reversals have names. Klarna rehired human agents after its assistant took on the workload of roughly 700 of them. Commonwealth Bank of Australia called its AI-driven job cuts an error and apologized to the staff involved. McDonald's shut down the drive-thru voice ordering it had been testing with IBM. What you rarely see in the headline is what each did next, because all three still run AI today. They changed how, not whether.

Key takeaways

• The retreat is real and measurable. The share of companies abandoning most of their AI initiatives before those reach production climbed from 17% to 42% in a year.
• It isn't a retreat from AI as a category, but from one deployment pattern: broad, unsupervised and pitched as a headcount replacement.
• MIT's Project NANDA report puts the dividing line at approach rather than model quality, and its own authors call the finding directional.
• Retail's version of the lesson is narrow scope, an existing workflow and a person who still presses send.

Are companies backing away from AI?

Some are, and the numbers aren't small. Surveying more than 1,000 respondents, S&P Global Market Intelligence found the share of companies abandoning most of their AI initiatives before they reach production rose from 17% to 42%, with the average organization scrapping 46% of its proofs of concept. Cost, privacy and trust led the obstacles. The same research found adoption accelerating rather than slowing, with 60% of AI-investing organizations having generative AI implemented.

The bigger number comes from MIT. Project NANDA's 2025 report on the state of AI in business reviewed more than 300 publicly disclosed AI initiatives and surveyed 153 senior leaders, and found 95% of organizations reporting no measurable profit-and-loss return, against $30 billion to $40 billion of enterprise spend. Read the fine print, though, because the authors flag their own limits: the figures are "directionally accurate based on individual interviews rather than official company reporting". That's a finding about direction, not a census.

So yes, plenty of companies are moving away from AI projects. Almost none are moving away from AI. The gap between those two sentences is where the useful information sits.

Which AI rollouts got pulled back?

Three reversals between 2024 and 2025 tell the story better than any survey. Each was reported as a defeat for AI. Each ended somewhere more interesting.

Look at the pattern rather than the drama. In every case the company put AI in front of the customer at volume, with a thin human backstop, and treated it as labor substitution. That bet lost. It isn't the same bet as using AI to tell an associate which twelve customers to text this morning.

The sequels are instructive too. Commonwealth Bank rebuilt its platform with Microsoft, and by mid-2026 it was cutting outsourced chat roles again, with its assistant resolving close to nine in ten conversations. So much for tidy morals. The second attempt succeeded where the first one failed, and what changed was the system, not the ambition.

What separated failures from successes?

Scope did most of the separating. MIT's researchers concluded that the divide "does not seem to be driven by model quality or regulation, but seems to be determined by approach", and the approaches that worked shared a shape: narrow use case, deep integration into a workflow people already used and a system that learned from feedback instead of repeating the same mistakes.

Three findings from that report are the ones to keep:

• Internal builds failed roughly twice as often as deployments run with an outside partner.
• Only 5% of custom enterprise tools reached production, while off-the-shelf assistants sailed through pilots and delivered personal productivity rather than profit-and-loss impact.
• Top-performing mid-market companies went from pilot to full implementation in about 90 days. Large enterprises took nine months or more and converted less often.

A mid-market manufacturing COO gave the researchers what they call the prevailing sentiment:

That isn't an argument against AI. It's an argument against buying AI as a strategy instead of applying it to a job.

Why do retail AI pilots stall?

Retail sits mid-table in MIT's industry grouping, where Consumer and Retail is characterized as support automation with limited impact on loyalty or on which brands lead. Read that as a clue rather than a verdict. Retail's most visible AI spending has gone into deflecting contacts, which is the use case where customers push back hardest and the one that produces no revenue when it succeeds.

The tone at NRF's 2026 Big Show pointed elsewhere. Dick's Sporting Goods' executive chairman described AI there as "a productivity tool more than a replacement of personnel". REI's chief executive argued her green-vest associates remain a differentiator because AI can't test outdoor gear in the field. Walmart's incoming chief executive described warehouse staff moving into bot-tech roles, a change to the work rather than a removal of it. Plenty of the real-world AI applications already running in retail never touch a customer directly.

Do shoppers still prefer human help?

Shoppers still prefer a person for anything carrying complexity or feeling, and the evidence for that predates the current backlash. A 2024 survey commissioned by the customer experience vendor Callvu found that 81% of respondents would wait at least some amount of time for a live agent rather than start immediately with a bot, with 16% willing to hold eleven minutes or longer. It's vendor-commissioned and two years old, so weigh it accordingly.

McKinsey's 2024 customer-care work carries two findings that sit badly with the automation case. In a separate McKinsey survey of 3,500 consumers, live phone conversations ranked among the most preferred contact methods for respondents of every age, Gen Z included. And in its fourth global survey of more than 340 care leaders, revenue generation had climbed from about one in 20 leaders naming it in 2016 to a third of them by 2024.

If your customer conversations are a revenue channel rather than a cost center, "how few humans can we get away with?" was always the wrong question. Clienteling starts from the opposite premise.

How should retailers deploy AI now?

Start where a mistake is cheap and a person still presses send. Points one and four come from MIT's success pattern; the rest is our reading of it.

1. Pick one repeatable decision, not a department. "Who should my team text today?" is a workable scope. "Automate customer service" isn't.
2. Point it at your staff first. Internal-facing AI fails privately. Customer-facing AI fails on a review site.
3. Keep approval with a human. A draft an associate edits in ten seconds carries your brand voice. An autonomous sender carries your risk.
4. Buy before you build. Internal builds failed about twice as often as deployments run with an outside partner, according to MIT's report.
5. Measure revenue, not deflection. Contacts avoided is a cost metric that hides a loyalty problem.

None of this requires a view on whether AI will one day handle the hard conversations. Your associates' jobs aren't what's under threat in this model, because the model only works if they stay in it.

What Endear learned from running this ourselves

Endear built AI Opportunity Engine on exactly that constraint. The system watches purchase and browsing behavior, surfaces the customers most likely to engage, and drafts a message with the reason attached. Then it stops. Every message requires a team member's review and approval before it sends, which both the help center and the product FAQ state as a design principle.

The early numbers are ours and they come from one specialty retailer, so weigh them accordingly. Outreach went from 60 messages a week to 360 over six weeks with no extra headcount and no drop in conversion rate, and stores that had previously done almost no personalized outreach became some of the most active. That same retailer, using Endear's AI Opportunity Engine, earned $35 in attributed revenue for every $1 spent on actioned opportunities. That's attributed revenue, not incremental profit. Endear supports over 2,000 stores in more than 19 countries.

The same logic runs through SalesChat, where a live associate handles the online conversation instead of a bot fielding returns queries. AI does the noticing. People do the talking. That division of labor is what every reversal in the table above was missing, and it's why the question for 2026 isn't whether to use AI. It's where to put the human.

FAQs

Are companies really abandoning AI?

Some abandon individual projects, and very few abandon AI altogether. S&P Global recorded 42% of surveyed companies abandoning most of their AI initiatives before production, up from 17%, while generative AI adoption kept climbing over the same period.

Why did Klarna rehire human agents?

Klarna's leadership concluded cost had become too dominant a factor, and service quality suffered for it. Human agents came back for complex and premium conversations while AI kept the high-volume, simple ones. It rebalanced rather than reversed.

Do customers prefer humans over chatbots?

Most customers will wait for a person rather than start with a bot, particularly when they're frustrated or the issue is complicated. McKinsey's 3,500-consumer survey put live phone conversations among the preferred channels in every age group, Gen Z included, though shoppers still want digital options alongside it.

Should retailers pause AI until it matures?

Pausing probably costs more than piloting narrowly, though that's our judgement rather than a measured finding. What the research shows is that the deployments which succeeded were narrow ones. A single well-scoped use case with human approval carries little downside.

Companies aren't backing away from AI so much as from AI that leaves people out of the loop. To see how Endear keeps your team in charge of every message, request a demo.