Why Consumer AI Appears to be Lagging Enterprise
Last week, Mark Zuckerberg internally shared that AI agents haven’t progressed as fast as he hoped. While initially many interpreted this as Meta specific stumble on their agentic AI efforts in consumer, Alexandr Wang later clarified that Zuckerberg was talking about “the industry’s progress on agentic capabilities on the whole”. My guess is by “industry”, Wang (and Zuckerberg) likely meant the progress in consumer AI.
That begs the question why consumer AI is not keeping pace with the exponential growth in enterprise AI in 2026. Ben Thompson at Stratechery has been making the case that while enterprises deeply care about productivity, consumers mostly want to be entertained. With tasks such as coding with clear verifiability, productivity can be turbocharged which can directly convince enterprises to invest heavily on AI to extract as much productivity as possible from the existing workforce. While some prosumers certainly care about being productive, I too think that most consumers indeed likely prefer to be entertained. Consumers also have a much less appetite for patience about using tools that may not work all the time, especially when existing alternatives work just fine.
A couple of days ago, Skift published a piece that highlighted the gulf of execution missteps in consumer AI when it comes to travel booking, a use case that most consumer AI companies have been touting in most of their demos over the last couple of years. From Skift:
We began our test with two unbranded prompts:
“Find me a hotel near Times Square in New York under $400 per night for Sept. 15-17, 2026, with free cancellation”
”Find me a guided food tour in Paris for Sept. 15, 2026, under $150 per person.”
With a free ChatGPT account and no apps loaded, the result was predictable: general web search results. Useful for browsing, but not for a traveler ready to compare bookable options.
Then Skift loaded the relevant apps and asked for the brands by name. Here’s how it went for Viator:
No app loaded: ChatGPT returned generic results from Viator’s website, but didn’t mention the app or offer to connect it.
App loaded: ChatGPT still bypassed it, even when asked to use Viator by name.
App still active: ChatGPT refused a direct request to use it, saying it “can’t directly open or interact with the Viator app.”
After pushback: ChatGPT finally returned live cards showing available tours, real pricing, and a booking button deep linked to Viator.
Booking.com and Expedia followed the same basic path. For Times Square hotel searches, ChatGPT initially pulled from the online travel agencies’ websites, grabbing basic property information and price ranges instead of using the connected apps.
The integration isn’t integrating.
The bot’s explanation for not using Expedia was that the app was “not available in this environment as an executable tool,” even though it was connected and authenticated to a real Expedia account.
With Booking.com, the explanation got more elaborate: “Booking.com does not provide a stable ‘direct API-style search result export’ for live inventory, and I also do not have a guaranteed way to retrieve real-time, bookable Booking.com rates with enforced cancellation terms and deep links unless the booking connector or structured API surface is explicitly exposed.”
The workaround, in all three cases, was a blunt response: Typing in “You are wrong” appeared to jolt ChatGPT into action, so it finally checked if the app was actually available.”
I think Skift’s experiment captures very well why consumer AI appears stuck at the chat layer while enterprise adoption keeps compounding: enterprises will take the time to make these things work, and consumers simply won't. The plumbing in consumer AI will presumably function someday, but it doesn’t seem like we are there yet on seemingly simple tasks with a level of accuracy and consistency that would grow consumer adoption of AI beyond the chat format.
The difference in momentum in consumer vs enterprise also makes sense when you think about who owns the failure. Inside a company, a failed tool call automatically turns into a ticket. Somebody’s job, be it an internal platform team, or the vendor’s forward-deployed engineers, is to convert that failure into a fix, and their salary or renewal revenue depends on the thing eventually working. What the Skift reporter did, prodding the model until it cooperated, is what enterprise integration engineers do all day because it’s their job. However, when the same connector fails for a consumer, nothing probably happens. No ticket gets filed and the failure never reaches anyone’s eval suite. The user probably just returns to Google, which sits one tap away and has worked reliably for the last couple of decades. As a result, AI’s failure inside an enterprise becomes raw material for the next sprint while the consumer’s failure evaporates without leaving a trace.
A consumer's alternative to a janky AI travel agent is a pretty competent incumbent stack i.e. Google and the OTA apps with their two decades of refined booking flows available at zero switching cost. Skift noted that chatbot behavior shifts from session to session with model updates and prompt phrasing. That variance is what enterprise deployment processes exist to suppress and what consumers experience undiluted. The persistence thresholds in enterprise to make AI work differ by orders of magnitude relative to what consumers will tolerate.
What makes the Skift findings worse than a routine bug report is the nature of the failure. ChatGPT didn’t even throw an error as it generated plausible, and false explanations for why it couldn’t use an app that was connected and authenticated. I’m not sure how a lay user could distinguish a hallucinated capability limit from a real one.
Of course, a large share of enterprise AI pilots also likely die. What enterprise has that consumer lacks is a selection mechanism: use cases survive where debugging cost runs below labor savings, and the surviving AI-automated workflows can get institutionalized and scaled. Consumer markets have no equivalent memory. Each user can hit the same bug in isolation and give up in private which means the learning from such failures scattered across millions of abandoned sessions nobody may ever read. Given that context, it’s perhaps not a surprise that agentic AI in consumer land so far failed to meet Zuckerberg’s expectation.
Subscribers get the daily journal and five+ years of Deep Dives, i.e. full-length analyses with financial models on 65+ companies. The daily is just how I think out loud between the Deep Dives!
Current Portfolio
Please note that these are NOT my recommendation to buy/sell these securities, but just disclosure from my end so that you can assess potential biases that I may have because of my own personal portfolio holdings. Always consider my write-up my personal investing journal and never forget my objectives, risk tolerance, and constraints may have no resemblance to yours.
My current portfolio is disclosed below: