SAN FRANCISCO/STOCKHOLM, Dec 16 – Last spring, CellarTracker, a wine-collection app, constructed an AI-powered sommelier to make unvarnished wine suggestions primarily based on an individual’s palate. The drawback was the chatbot was too good.
“It’s just very polite, instead of just saying, ‘It’s really unlikely you’ll like the wine,’” CellarTracker CEO Eric LeVine stated. It took six weeks of trial and error to coax the chatbot into providing an sincere appraisal earlier than the function was launched.
Since ChatGPT exploded three years in the past, firms huge and small have leapt on the likelihood to undertake generative synthetic intelligence and stuff it into as many merchandise as doable. But up to now, the overwhelming majority of companies are struggling to understand a significant return on their AI investments, based on firm executives, advisors and the outcomes of seven current govt and employee surveys.
One survey of 1,576 executives carried out through the second quarter by analysis and advisory agency Forrester Research confirmed simply 15 per cent of respondents noticed revenue margins enhance resulting from AI during the last 12 months. Consulting agency BCG discovered that solely 5 per cent of 1,250 executives surveyed between May and mid-July noticed widespread worth from AI.
Executives say they nonetheless imagine generative AI will finally rework their companies, however they’re reconsidering how shortly that may occur inside their organizations. Forrester predicts that in 2026 firms will delay about 25 per cent of their deliberate AI spending by a 12 months.
“The tech companies who have built this technology have spun this tale that this is all going to change quickly,” Forrester analyst Brian Hopkins stated. “But we humans don’t change that fast.”
AI firms together with OpenAI, Anthropic and Google are all doubling down on courting enterprise clients within the subsequent 12 months. During a current lunch with media editors in New York, OpenAI CEO Sam Altman stated growing AI methods for firms may very well be a $100 billion market.
All that is occurring towards the backdrop of unprecedented tech funding in the whole lot from chips, to information facilities, to power sources.
Whether these investments may be justified shall be decided by firms’ capability to determine find out how to use AI to spice up income, fatten margins or velocity innovation. Failing that, the infrastructure build-out may set off the type of crash paying homage to the dot-com bust within the early 2000s, some consultants say.
THE ‘EASY’ BUTTON
Soon after ChatGPT’s launch, firms worldwide created activity forces devoted to discovering methods to embrace generative AI, a kind of AI that may create unique content material like essays, software program code and pictures by textual content prompts.
One well-known situation with AI fashions is their tendency to please the person. This bias – what’s known as “sycophancy” – encourages customers to speak extra, however can impair the mannequin’s capability to provide higher recommendation.
CellarTracker bumped into this drawback with its wine-recommendation function, constructed on prime of OpenAI’s expertise, CEO LeVine stated. The chatbot carried out nicely sufficient when requested for normal suggestions. But when requested about particular vintages, the chatbot remained optimistic – even when all alerts confirmed an individual was extremely unlikely to take pleasure in them.
“We had to bend over backwards to get the models (any model) to be critical and suggest there are wines I might not like,” LeVine stated.
Part of the answer was designing prompts that gave the mannequin permission to say no.
Companies have additionally struggled with AI’s lack of consistency.
Jeremy Nielsen, normal supervisor at North American railroad service supplier Cando Rail and Terminals, stated the corporate not too long ago examined an AI chatbot for workers to review inside security experiences and coaching supplies.
But Cando ran right into a shocking stumbling block: the fashions couldn’t constantly and appropriately summarize the Canadian Rail Operating Rules, a roughly 100-page doc that lays out the security requirements for the business.
Sometimes the fashions forgot or misinterpreted the principles; different occasions they invented them from entire material. AI researchers say fashions usually battle to recall what seems in the course of an extended doc.
Cando has dropped the mission for now, however is testing different concepts. So far the corporate has spent $300,000 on growing AI merchandise.
“We all thought it’d be the easy button,” Nielsen stated. “And that’s just not what happened.”
HUMANS MAKE A COMEBACK
Human-staffed name facilities and customer support had been presupposed to be closely disrupted by AI, however firms shortly discovered there are limits to the quantity of human interplay that may be delegated to chatbots.
In early 2024, Swedish funds firm Klarna rolled out an OpenAI-powered customer support agent that it stated may do the work of 700 full-time customer support brokers.
In 2025, nonetheless, CEO Sebastian Siemiathowski was compelled to dial that again and acknowledge that some clients most well-liked to speak with people.
Siemiathowski stated AI is dependable on easy duties and may now do the work of about 850 brokers, however extra advanced points shortly get referred to human brokers.
For 2026, Klarna is concentrated on constructing its second-generation AI chatbot, which it hopes to ship quickly, however human beings will stay a giant a part of the combination.
“If you want to stay customer-obsessed, you can’t rely [entirely] on AI,” he stated.
Similarly, U.S. telecommunications large Verizon is leaning again into human customer support brokers in 2026 after makes an attempt to delegate calls to AI.
“I think 40 per cent of consumers like the idea of still talking to a human, and they’re frustrated that they can’t get to a human agent,” stated Ivan Berg, who leads Verizon’s AI-driven efforts to reinforce service operations for enterprise clients, in a Reuters interview this fall.
The firm, which has about 2,000 frontline customer support brokers, nonetheless makes use of AI to display screen calls, get info on clients, and direct them to both self-service methods or to human brokers.
Using AI to deal with routine questions frees up brokers to deal with advanced points and check out new issues, similar to making outbound calls and doing gross sales.
“Empathy is probably the key thing that’s holding us from having AI agents talk to customers holistically right now,” Berg stated.
Shashi Upadhyay, president of product, engineering and AI at customer-service platform Zendesk, says AI excels in three areas: writing, coding and chatting. Zendesk’s purchasers depend on generative AI to deal with between 50 per cent and 80 per cent of their customer-support requests. But, he stated, the concept that generative AI can do the whole lot is “oversold.”
THE ‘JAGGED FRONTIER’
Large language fashions are quickly conquering advanced duties in math and coding, however can nonetheless fail at comparatively trivial duties. Researchers name this contradiction in capabilities the “jagged frontier” of AI.
“It might be a Ferrari in math but a donkey at putting things in your calendar,” stated Anastasios Angelopoulos, the CEO and cofounder of LMArena, a well-liked benchmarking device.
Seemingly small points can unexpectedly journey up AI methods.
Many monetary corporations depend on information compiled from a broad vary of sources, all of which may be formatted very in a different way. These variations may immediate an AI device to “read patterns that don’t exist,” stated Clark Shafer, director at advisory agency Alpha Financial Markets Consulting.
Many firms are actually wanting into the doubtless costly, prolonged and sophisticated means of reformatting their information to reap the benefits of AI, Shafer stated.
Dutch expertise funding group Prosus says one in all its in-house AI brokers is supposed to reply questions on its portfolio, much like what the group’s information analysts on workers already do.
Theoretically, an worker may ask how usually a Prosus-backed food-delivery agency was late to ship sushi orders in Berlin final week.
But for now, the device doesn’t at all times perceive what neighborhoods are a part of Berlin or what “last week” means, stated Euro Beinat, head of AI for Prosus.
“People thought AI was magic. It’s not magic,” Beinat stated. “There’s a lot of knowledge that needs to be encoded in these tools to work well.”
MORE HANDHOLDING
OpenAI is engaged on a brand new product for companies and not too long ago created inside groups, such because the Forward Deployed Engineering crew, to work instantly with purchasers to assist them use OpenAI’s expertise to sort out particular issues, a spokesperson stated.
“Where we do see failure is people that jump in too big, they find that billion-dollar problem—that’s going to take a few years,” stated Ashley Kramer, OpenAI’s head of income, throughout an onstage interview at Reuters Momentum AI convention in November.
Specifically, OpenAI is working with firms to seek out areas the place AI can have a “high impact but maybe low lift at first,” stated Kramer.
Rival AI lab Anthropic, which pulls 80 per cent of its income from enterprise clients, is hiring “applied AI” consultants who will embed with firms.
For AI firms to succeed, they should view themselves as “partners and educators, rather than just deployers of technology,” stated Mike Krieger, Anthropic’s head of product, in an interview earlier this 12 months.
An rising variety of startups, many based by former OpenAI workers, are growing AI instruments for particular sectors similar to monetary providers or authorized. These founders say firms will profit from specialised fashions greater than general-purpose or shopper instruments like ChatGPT.
It’s a playbook that Writer, a San Francisco–primarily based AI utility startup, has been adopting. The firm, which is now constructing AI brokers for finance and advertising groups at massive corporations similar to Vanguard and Prudential, places its engineers on calls instantly with purchasers to know their workflows and co-build the brokers.
“Companies need more handholding in actually making AI tools useful for them,” stated May Habib, CEO of Writer.