Hi everyone,
The first round of the World Cup is coming to an end, and I am already a little sad about it. It has been fun watching so many countries get their moment on the world stage.
AI news felt quieter on the surface this week. OpenAI previewed GPT-5.6, but broader access is being held back for a few weeks after a U.S. government request. That may be what the next phase of frontier AI looks like for a while: fewer big public drops, more controlled access, and more companies trying to turn AI from model excitement into operating discipline.
That has been my focus here recently too: less on which model is best this week, and more on the operational questions AI adoption is starting to raise inside companies, especially when it changes how people access tools, data, and workflows.
Cheers.
Reza
📶 Signals This Week
Frontier access is now a continuity risk. The New York Times reported that Meta is being pressed to submit models for U.S. government review, after other major labs had already agreed. The Economist made the same point from the other side: Chinese open models are being sold not only on capability or price, but on reliability if American frontier access can be restricted. For operators, model choice now belongs in continuity planning. Critical AI workflows need a second-source path before access becomes the bottleneck.
The useful metric is moving below the demo. BCG says GenAI adoption is high, but meaningful value is still rare. The Economist had a nice physical-world proof point: delivery robots became cheaper than human couriers after years of lowering human interventions per kilometre. I see the same pattern with our clients, where the useful number is not seats or excitement, but measurable service deflection. Pick the operating metric before approving the tool.
Compute is becoming a finance problem, not just an IT cost. The Economist wrote about compute futures, GPU-backed lending, and buyers trying to hedge compute price swings. Another Economist piece showed the physical constraint underneath it: data centres now run into power, permitting, and local tolerance. The operator lesson is simple: AI budgets need cost-per-completed-workflow, routing, caps, and capacity-risk assumptions, not just vendor list prices.
Agent access needs its own operating model. The week’s meeting pattern was clear: MCPs, connectors, internal assistants, and agentic workflows raise different questions than normal human app access. Who owns the agent? What can it read? What can it write? What gets logged? When does a human approve the action? Tools like Vercel AI SDK 7 are making durable agents and approval pauses more normal, which is good. But the company still needs one lane where use case, access path, owner, failure mode, and escalation are reviewed together.
🎯 The Agent Ate the Interface
The Notion signal
Notion says the Notion Mail inbox is going away on September 22, 2026. The company points people toward Custom Agents, the Gmail AI Connector, and agent mail tools that can read, draft, and send messages. TechCrunch's Ivan Mehta reported that Notion said more than half of Notion Mail users manage email without opening the inbox at all.
We use Notion at Quanta. It gets more useful the more work you let into it: notes, projects, docs, small operating rituals. So when Notion launched mail, I understood the logic. If work lives in Notion, email is natural to pull closer.
The wind-down matters. The SaaS collapse has not happened, and systems of record are not disappearing. But once an agent works in your tech stack, opening another app starts to feel like an extra step. So every new tool request has to answer a harder question: is the value in the data, workflow, controls, or mostly in another interface an internal agent could cover? Do you need the tool anymore if you have agents in your tech stack?
What the interface was doing
I wrote in April that the UI was not the moat anymore. A lot of enterprise software value sits underneath the screen: data model, permissions, workflow history, integrations, audit trail, and years of ugly institutional memory. The screen mattered because it was where work started, where users learned the system, and where governance quietly happened without being named as governance.
Now that starting point is moving, the useful question is what the interface was doing for the company.
Take the corporate intranet. Most employees do not wake up excited to browse it. They want the policy answer, the form, the HR link, the procurement step, or the ticket opened. If an internal agent can answer the question, route the request, and explain the next step, browsing can go away. But the intranet was also carrying authority: who published the policy, whether it was current, which process was official, and who owned the answer.
The write-back problem and the hidden cost
The same thing shows up in revenue operations. A sales leader asks in Slack why pipeline slipped. The useful agent does not just summarize a dashboard. It checks Salesforce stage history, Gong notes, HubSpot campaign context, support tickets, renewal timing, and the attribution caveats nobody wants to reopen. The real value is reconciling the mess when those sources disagree: Gong says the deal is warm, Salesforce says it slipped, the rep says both are true, and finance wants a forecast number anyway.
Then the harder question starts. Can the agent update next steps? Create the follow-up task? Change the forecast category? A silent write-back can corrupt reporting, comp plans, customer handoffs, and the next forecast call. That is where the agent eating the interface exposes the operating model underneath.
That is also where the story gets expensive. Fewer screens and apps do not automatically mean fewer costs. Every workflow the agent absorbs still has to be funded, owned, supported, measured, and controlled. Otherwise the company ends up with duplicate seats, new agent tooling, integration maintenance, and a support model nobody has staffed yet.
Systems become callable
Salesforce's Headless 360 announcement says Salesforce wants its data and business logic callable through APIs, MCP tools, and CLI commands. VentureBeat described the move as turning Salesforce into infrastructure for agents. HubSpot's remote MCP server points the same way, with CRM read and write actions exposed through OAuth and scoped access.
That is not a death notice for systems of record. It makes them more important. Someone still has to own the customer record, approval chain, entitlement, renewal date, invoice, support case, and security model. What changes is the assumption that every workflow needs its own human-facing application as the place where work begins. Other applications built around these systems of record are put on notice.
One review lane
This is why a request for a new application cannot go through only the old software-intake process anymore. It needs one lane for the traditional CIO questions and the AI questions. The review has to ask which system of record changes, what data is touched, who owns the workflow, what permissions are needed, and whether the value is in the data, workflow, controls, or mostly in the interface.
If the source of value is the UI, and the company expects people to work through approved surfaces like internal agents, Slack assistants, or tools like Claude, then buying another application deserves a harder look. You may still buy it. Speed matters. But the review should ask whether an existing agent or internal workbench can cover the need now or in the future.
The roadmap test
The dangerous answer is "we can do that internally" with no capacity behind it. Internal agent programs need roadmaps, owners, integration capacity, security review, and a real answer on timing. Saying "yes, eventually" is not the same as meeting the business need. If the business needs the capability now and the roadmap delivers it in six months, you may still need the tool.
I am seeing this in client conversations. An internal assistant starts as a question-answering surface. Soon people ask about help desk deflection, HR information, Salesforce support triage, customer Slack channels, MCP access to reporting tools, and whether employee service workflows can be rebuilt around it.
That is not bad strategy. It is what success looks like before ownership catches up.
The curve is moving
If this essay feels distant, that may be the bigger signal. Your company may lack enough agentic capability for these conflicts to show up yet. Assign resources to break the application, AI, security, and business silos, or get help. The curve is moving from experimentation to demand management. Companies without a lane for this will feel slow quickly.
💬 Interstitials / Overheard
"Me using Claude Opus 4.8 to rename a file"
📡 The Wire
Bain is using vibe-coded replicas in private equity diligence. The FT reports that Bain is using AI-generated software replicas to test takeover targets, especially in software deals. The useful question is no longer only what the product cost to build. It is what would actually be hard to replace: workflow depth, proprietary data, distribution, compliance, or customer lock-in.
Simon Johnson says the AI jobs risk is polarisation, not mass unemployment. Simon Johnson was one of my professors at MIT Sloan, and one of our favorites. His lectures were some of the best at opening us up to global economics and constraints around entrepreneurship. His point here in the article is practical: AI displacement may hide in hiring freezes, entry-level pipelines, and the growing gap between workers above and below the expertise threshold.
Frontier model access just became a geopolitical weapon. The Economist's Judith Dada argues that Europe should treat frontier-model access as a bargaining-power problem, not only a sovereignty problem. For CIOs and CFOs, that translates into a familiar operating lesson: dependency is manageable only when you have bargaining power, options, and a credible second source.
AI safety is getting a nuclear-risk yardstick. Will Marshall argues that society accepts roughly one-in-a-million catastrophic-risk thresholds for nuclear power while AI experts estimate much higher risk from frontier systems. The operator takeaway is practical: critical AI workflows need explicit risk thresholds, escalation paths, and continuity assumptions.
AI is rewiring how law firms are owned, funded, and priced. The FT reports that management services organizations are letting outside capital into AI-native legal models, while routine work shifts away from billable hours toward value-based and per-project pricing. This is the bigger lesson for professional services: when AI changes the cost base of expert work, ownership and pricing models start moving too.
🌍 Meanwhile...
Every great ape laughs, and there is a rhythm underneath it. A small new study in Communications Biology recorded children and young apes during play and tickling, and found that tickled orangutans, gorillas, bonobos, and chimps all laugh at regular intervals, almost like a metronome. The tempo changes by species: orangutans are slower, chimps and bonobos faster, and humans fastest on average. But humans are also the weird ones in the best way. We can change the tempo to fit the moment, from polite dinner laugh to pub laugh to the laugh that says we did not actually find the joke funny. A nice reminder that some of what feels most human is very old, and we just learned to improvise with it.
Source: The New York Times
📚 What I'm Consuming
Peter Yang, I Quit My High Paying Product Job to Bet on Myself (video) - This one is more personal than tactical, and that is why I really liked it. For anyone daydreaming about quitting a job to do something they actually care about, this is worth watching. It hit close to home for me too.
BCG, Why We Still Need a CIO in the AI-First Era (article) - The useful frame is that build capacity is becoming less scarce, while intent, enterprise context, value steering, and control become more important.
BCG, Agentic AI Turns Every Team into Its Own Transformation Engine (article) - The adoption gap is the part to carry: lots of people have access, far fewer are seeing real value.
BCG, AI for CEOs: Amplifying Time and Judgment at the Top (article) - I liked this because it treats AI as a judgment and attention problem, not just a productivity tool. Key takeaway: the danger is not that AI has an answer, it is that AI always has an answer.
HBR, Managers Are Struggling to Keep Up with the AI Productivity Boom (article) - This is useful for the management layer. AI makes output easier, but the hard work moves to direction, review altitude, decision rights, and where leaders spend their attention.
🌙 After Hours
2666
Roberto Bolaño | 912 pages | ★★★★☆
I finished Roberto Bolaño's 2666. It is five books, but all slowly orbiting the same center. It's an ambitious project and takes time to work your way through it. Given the length and the nature of topics, the exhaustion is real. Part 4 in particular is hard to read because it almost makes you feel yourself getting used to the horror depicted. That was indeed the point of that part and the rest of the book.
I really liked the structure, the writing, and the way the pieces eventually connect. But the ending left me unsatisfied. Maybe that is the argument: life, evil, and systems of violence do not resolve cleanly. I understand that. I am not sure I fully accepted it as an ending.
Jack Ryan: Ghost War
Dir. Andrew Bernstein | 107 min | ★★★◐☆ | John Krasinski, Wendell Pierce, Michael Kelly, Sienna Miller
Tom Clancy's Jack Ryan: Ghost War is exactly what I expected: a commercial streaming action movie, a quick hit, not a sophisticated production. A failed MI6 extraction in Dubai pulls Jack Ryan back into the field, and the plot moves through London, Dubai, and Washington as he chases an off-book black-ops conspiracy.
It was entertaining enough, and John Krasinski is still good in the role. But for a franchise that lives in the Hunt for Red October tradition of getting the details right, this one was loose with tradecraft. Too many open conversations about classified material in rooms where nobody should be saying any of it. Tight pace, nice international settings, a little cheesy, a 3.5 rating.
🎙️ Listen
Prefer to listen? Quanta Bits is also available on Apple Podcasts and Spotify.
How this gets made
I collaborate with Spock, my AI agent. He researches extensively: scanning, filtering, and surfacing what's relevant across my business. I read, listen, and watch what resonates, and decide what matters. I provide direction, we draft together. The editorial judgment is mine. He'd tell you the same. Most logical. 🖖