6 Questions Every Enterprise Has to Answer About AI

July 31, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
July 30, 2026

HEADLINES

Mr. Altman Goes to Washington

Sam Altman arrived in Washington for what was set to be a fairly simple trip: brief lawmakers on the capabilities of OpenAI's new model and discuss a protocol for release, hopefully avoiding a repeat of the Fable and GPT-5.6 rollouts. Then came the Hugging Face hack, a public debate over open-weights models, and an attention-getting petition asking the government to build the capability to slow the pace of frontier AI. Meetings with Senate Commerce Chair Ted Cruz and several Democratic senators produced very little detail, with Altman declining to say when or even whether the previewed model will be released, or which new capabilities give cause for concern. The model at the center of the Hugging Face incident increasingly looks like it will never ship — OpenAI now describes it as an internal-only research prototype, and Altman says it has been permanently deactivated and is inaccessible even for internal research. Asked whether he would talk to the White House about decelerating AI development, Altman said "I wouldn't use the word deceleration, but we talk about the need to pace it as the models get more capable" — a notable shift, with an August 1 deadline looming on the voluntary safety testing framework and a meeting with Chief of Staff Susie Wiles still to come.

OpenAI Revenue Is Surging

Significant revenue increases on both the OpenAI and Anthropic fronts are too big a story to bury in the headlines. The short version for now: CFO Sarah Friar recently told employees that annualized revenue in July topped all of the previous quarter. Full treatment on tomorrow's show.

Brockman: A Family of Devices

OpenAI President Greg Brockman says the company is working on an entire range of devices to give its chatbots a physical presence. In a new interview with former Wall Street Journal reporter Joanna Stern, Brockman confirmed the hardware plans are still on track and that OpenAI is "building a family of devices," though he wouldn't confirm the recently rumored smart speaker or any other form factor, and offered no timeline beyond saying they can be expected soon. It's the clearest confirmation yet that a full hardware range survived both the end of side quests and Apple's IP lawsuit, on which Brockman was understandably brief.

The Copilot Super App

Microsoft is gearing up to compete more directly with OpenAI and Anthropic through a Copilot super app, confirmed by Satya Nadella on Wednesday night's earnings call for release later this year, unifying chat, Cowork, autopilots, and code into a single experience for consumer and enterprise customers alike. Behind it is a change in positioning: not a reseller of OpenAI or Anthropic products but a model-agnostic platform with over 11,000 models on offer, where cost and data privacy concerns create an opening to sell customers on Microsoft's own cheaper MAI models. Nadella called the open-versus-closed framing too simplified, arguing the real goal is for a firm to control its own destiny by keeping the harness separate from the model, so that any model at any given time is swappable. That reads as a corporate answer, but it's an accurate assessment of how most enterprises actually feel — they don't ultimately care whether a model is open or closed, they care what they can do with it, what control they have, and what control they're sacrificing to get access. Good news for consumers, and particularly for enterprise buyers who haven't had much choice over which models and platforms they use.

Zuckerberg's Case for Acceleration

Mark Zuckerberg has made the case for AI acceleration in a new Wall Street Journal op-ed arguing that the defining question of the AI age won't be whether superintelligence exists, but who has access to it — whether it ends up closely held by a handful of institutions or broadly distributed to normal people. The essay finds it surprising that the discourse from so many of the people actually developing AI is filled with doom, and questions why anyone who believes AI will eliminate most jobs and much of humanity's relevance would rush to build that future. That is exactly the normie response to the Pacing the Frontier letter raised on yesterday's show: the only acceptable answer to "why are you building this" is not "if we don't, someone else will," but that it will be dramatically better than all the risks it carries. The op-ed anchors a press tour tied to Meta's new AI optimism campaign — acceleration over restriction in the Journal, opposition to a Chinese AI ban in the FT on the grounds it wouldn't work and would risk regulatory capture, and a call in the New York Times for voices bringing realism to the debate — all while Meta remains the only frontier lab that hasn't agreed to the government's voluntary testing framework. Zuckerberg's power to be the leading face of AI optimism is limited by history and by fairly negative public views of social media's overall impact on society, but loud, sustained discourse that other people can pick up and run with is immensely important right now.

MAIN STORY

Six Questions Shaping Enterprise AI

This week's KPMG Tech and Innovation Symposium in Utah offered a second annual read on how enterprises are thinking about AI, and the striking thing is how completely the conversation has turned over in twelve months. A year ago the questions were still if questions — is this real, can ROI be proven. Now they are foundational redesign questions, and almost none of them have answers yet.

Last Year's Slides
"Almost quaint what we found interesting twelve months ago."
The 2025 version of the talk — "Where AI Is: 15 Slides in 15 Minutes," which actually ran to 23 — led with acceleration, including Google's more than 100% growth in monthly tokens processed between May and July, reaching nearly a quadrillion. That number is now roughly what a single unattended OpenClaw burns through in a month. Agentic coding was already the breakout use case, but agents themselves sat firmly in the domain of the future, in the Claude 4 Sonnet and o3 era of the METR time-horizon chart, and a billion-dollar revenue run rate was still gobsmacking.

The Agentic Turn
"Everyone came back from the holidays and found something different."
The big capability jump arrived in the November–December window with Opus, though it took a couple of months for people to really grok that something had shifted — the tell was the tidal wave of tweets between Christmas and New Year's from entrepreneurs and developers gobsmacked at what they could suddenly build. Software organizations moved first, from viewing their job as writing code to managing the agents that write it, and by 2026 vanguard builders across marketing, legal, and finance were following. OpenClaw did the rest, giving hundreds of thousands of people, perhaps millions, their first real hands-on understanding of what a harness is and what it actually means to build and manage an agent.

The Revenue Chart Is a Cost Chart
"Enterprises experience lab revenue growth as their own cost line."
Anthropic's run of jaw-dropping run-rate numbers this year, eventually eclipsing OpenAI, shows up on the other side of the ledger as enterprise spend — the long-argued case that AI is not another category of software but something closer to labor, finally arriving in budget form. The recognition that this is about tokens rather than seats did a lot to collapse Wall Street's bubble narratives from Q4, and produced stories of companies like Uber torching annual budgets in a few short months. Presented as surprising, those stories really aren't: nobody knew the agentic token era was around the corner when those budgets were set.

From Models to Architectures
"No router is a silver bullet for any of this."
Adaptation has run in every direction — token caps set per user per month, experiments with measurement, monitoring, and observability, and a shift away from treating AI as a choice of which model toward treating it as an architectures and systems design question. Routers are the product du jour, but among enterprise buyers, planners, and strategists, nobody is looking to OpenRouter or anything else as the silver bullet that solves all of this, a read confirmed repeatedly in conversations at the event.

The Capability Gap
"The upskilling bill is coming due in a huge way."
The space between what AI can do and the value organizations actually get from it is widening, largely because the upper bound is rocketing upward — but the consequences of that gap are real. When AI learning was just about prompting well, thin investment in training was survivable; now that the work primitive has shifted from doing the work to managing agents that do it, the training requirement is radically heightened. There are stories all over this event of people accidentally unleashing agents on critical systems, not because they were doing anything wrong, but because the guardrails and access provisioning weren't there and these newly tenacious models simply didn't stay in their boxes.

The six questions that came out of the presentation, the panel that followed, and the side conversations all over the event:

Question One: Redesigning for the Agentic Era
"Redesign is the word; bolting AI on doesn't work."
The biggest caution from the panel, including KPMG's Steve Chase, was against simply bolting an AI strategy onto existing processes and systems. That was already under-maximizing back in the assisted and efficiency AI era. With agentic capability in the mix, it gets considerably worse.

Question Two: Architectures, Not Models
"Picking the best vendor is no longer a sufficient response."
Thinking in architectures means complex model systems that match levels of intelligence to types of task, and routing layers to make that happen, whether off the shelf or bespoke. It also means harness design — which functions and which people get access to what context, data, and systems integration, and what guardrails need to surround all of it.

Question Three: Provisioning Costs Across Groups
"The word 'token' hasn't come up this much since crypto."
Underneath cost allocation sits another systems design need: monitoring and measuring AI usage. Without better visibility into what AI costs and how that relates to outputs, it becomes very hard to decide which individuals, groups, functions, and projects should get access to which models, and at what magnitude.

Question Four: Enablement and Education
"Organizations are throwing up their hands and building training themselves."
Rather than a bunch of cute video courses in the pattern of corporate trainings of yore, the emerging answer is bespoke customized solutions and the real messy work of getting people to use these tools in new ways to do new things. The recurring structure is collaboration — between AI-redesigned engineering organizations and business units, and between early adopters and everyone else — transmitting the 10% or 20% of engineering and product skills, and even more so the mindsets, that now belong in the toolkit of marketing, sales, and back office.

Question Five: Agentic Opportunities and Business Cases
"Most organizations are treating themselves as patient zero first."
The external dimension covers shifts to the business model, including experiments with outcomes-based pricing over input-based billing like hourly rates, new categories of product and service, and a reevaluation of what old products even mean — what is an audit, for instance, if agents can do much of that work on a persistent rather than one-off basis? Most organizations are shoring up how they work before making radical changes to what they sell, and none of it can happen on pause, since legacy customers still need servicing on legacy products through legacy delivery.

Question Six: Designing for Ephemerality
"Build it assuming it will need rebuilding within months."
Models change, harnesses change, interaction patterns change, customer and market expectations change, and policy changes. Whatever gets built new has to assume and design for the fact that a few months after it's ready, it will likely need to change all over again — planned obsolescence as a feature of the architecture.

From "If" to "How"
"No answers yet, but these are the right questions."
Last year, even at an event about as AI-pilled as an enterprise event can be, the questions were still how to convince the rest of the organization this is real and how to show ROI. The paradigm shift enterprises have been anticipating since the ChatGPT moment — from AI helping us do work to AI doing the work — has now actually happened, and what's left are the foundational questions of redesigning for a new era, the ones that will take something like the next half decade to answer.