The Agent Ate the Interface

This week on Quanta Bits, I talk through why Notion winding down its Mail inbox is a useful signal for enterprise software. Systems of record are not disappearing, but agents may change where work starts. If an agent can read, route, and eventually write back across CRM, email, ticketing, documents, and finance systems, a new application request needs a different review lane. We need the old CIO questions and the new AI questions in the same conversation: what data is touched, who owns the workflow, what can the agent do, where does human approval sit, and whether the value is really in the interface.

I also cover quick hits on frontier model access as a continuity risk, compute becoming a finance problem, Bain using AI-generated software replicas in private equity diligence, why companies need to get moving on internal agentic capability, and a short After Hours note on 2666 and Jack Ryan: Ghost War.

Read the full issue: https://quanta-bits-newsletter.beehiiv.com/p/quanta-bits-june-27-2026
Subscribe to Quanta Bits: https://quanta-bits-newsletter.beehiiv.com/subscribe

Transcript

SPEAKER_00 0:01

Notion Mail going away doesn't sound like a big story at first. It sounds like one of those product cleanup nodes you skimmed past. It featured launch, some users tried it, the company changed direction, everybody moved on. But I think this one is worth sitting with for a minute. Notion says the Notion Mail inbox is going away in September. Instead, they are pointing people toward custom agents, the Gmail AI connector, the agent mail tools that can read, draft, send messages. TechCrunch reported that more than half of Notion Mail users were already managing email without opening the inbox at all. That's the part that stayed with me. If agents can do the work across systems, another screen starts to feel like an extra step. Hey, I'm Rezo Markabody. Welcome back to QuantaBits. This week I wrote about what happens when agents start eating the interface. Not because software disappears, but because the place where work starts may move. We use Notion at Quanta, probably not as much as we should, but enough that I understand why Notion wanted mail closer to their workplace. If your notes, projects, docs, and rituals live in Notion, email is a natural thing to pull in. That made sense to me. But the window says something about where the work may be moving. The SaaS collapse hasn't happened. I don't think Salesforce, Hotspot, Workday, ServiceNow, or the big systems of record are just going away because we have agents now. That's too simple. Those systems hold the customer record, the permissions, the workflow history, the audit trail, the invoices, the support cases, and a lot of institutional memory. But the screen is no longer automatically where the work has to begin. So when a team asks for a new tool, the review is shifting from is this useful? to where does the value actually live? Is it in the data, the workflow, the controls, the history, or is it mostly another interface sitting on top of systems the company already owns? That's a very different software review conversation. I wrote a few months ago that the UI was not the mode anymore. I still think that is mostly right, but I want to be careful with it. The interface was not just decoration, it did real work. It taught people where to click, it made certain actions visible and others harder. It showed official workflow. It gave people a place to learn the system. And without saying the word governance, it carried a lot of governance. Take the corporate intranet. Nobody wakes up excited to browse the intranet. They want the policy answer, the phone, the HR link, the procurement step, or the ticket open. If an internal agent can answer the question and route the request, great, browsing can go away. But the intranet was also carrying authority. Who published the policy? Which process is official? Who owns the answer? When the screen disappears, some hidden control disappears with it too. That doesn't mean the screen has to stay. It means the operating model has to replace what the screen was doing quietly. Reading is the easy part to talk about. An agent summarizes a policy, it finds the right account, it explains why a ticket is stuck. Useful, but usually not scary. The real tession starts when the agent can write back. Imagine a sales leader asking in Slack why pipeline slipped. A useful agent has to check Salesforce stage history, notes from calls, hotspot campaign context, support tickets, renewal timing, and the caveats nobody wants to reopen. Then it finds disagreement. Gong says the deal is warm, Salesforce says it slipped, the rep says both are true, finance still needs a forecast number. At that point, can the agent update the next step, change the forecast category, create the follow-up task, send the customer note? That's where things get expensive. A sign and write back can corrupt reporting, compliance, customer handoffs, and the next forecast call. Fewer screens don't automatically mean fewer costs. Every workflow the agent absorbs still needs funding, ownership, support, logs, exception handling, and someone who knows what to do when it gets stuck. The moment an agent can write back, it stops being a user experience question. It becomes an operating model question. This is where I think a lot of companies need a practical change. Every new application request now needs the old CIO review and the AI review in the same lane. The old questions still matter. Is the vendor secure? Who pays for it? What data doesn't it touch? Does it overlap with tools we already have? Who supports it? But the AI questions now belong in that same conversation. Could an internal agent cover this workflow? Could an existing approval assistant do the first version? If you buy the app, I'll be buying a system of record, a control layer, a workflow engine, or mostly another screen. You may still buy the tool because speed matters. Sometimes the business need is real and internal capacity is not ready. But we can do that internally, doesn't count as strategy unless there is roadmap capacity behind it. That's the trap I worry about. Someone says, we don't need to buy this, our agent platform can handle it. Then six months pass, the business needs still there, and nothing has shipped because the AI roadmap was already full. So the review lane has to ask both questions together. Should we buy it? Could we build or extend an agent to cover it? And if we say later, who owns later? A few quick hits from the rest of the issue. First, frontier model access is becoming a continuity risk. The New York Times reported that Meta is being pushed toward a US government security reviews for its models. The economists wrote that Chinese open models are being sold not only on price and quantity, but on reliability if American frontier access gets restricted. For operators, that means model choice is no longer just a technical preference. If a critical workflow depends on one vendor, one region or one access path, you need a second source plan before access becomes a problem. Second, compute is starting to look like a finance problem. The economists wrote about compute futures, GPU-backed lending, and buyers trying to hedge compute price swings. That sounds abstract, but the operating point is simple. Budgets can't just be vendor list prices and seed counts. They need cost per completed workflow, routing rules, caps, and assumptions about capacity risk. And third, Bain is using AI-generated software replicas in private equity diligence, according to the Financial Times. That's a small window into how people will value software. You're not only asking what the product costs to build, you're asking what would be hard to replace the workflow depth, the propriety data, the compliance age, or the customer lock-in. That's the same interface question from another angle. If the thing is mostly screen, the screen is getting cheaper to copy. Quick after hours before I wrap. I finished Roberto Bellano's 2666, which is a huge, strange, exhausting book. Five parts all circling the same center. Part 4 is especially hard to read, partly because it makes you feel yourself getting used to horror. And I think that's very much the point. I like the structure and the writing. I didn't feel fully satisfied by the ending. Maybe that's the argument. Life, evil, and systems of violence don't resolve cleanly. I understand that, I'm just not sure I fully accepted it as an ending. And then on the other end of the spectrum, I watched Jack Ryan Ghost War. It was exactly what I expected. Fast, commercial, entertaining enough, and little loose with tradecraft. Sometimes that is fine too. Not every system resolves clearly, which was oddly on theme this week. If this whole conversation today doesn't resonate with you yet, that may be the warning sign. It may mean your company hasn't gone far enough with internal agents for these conflicts to show up. Business teams are not yet asking why an internal assistant can't deflect helpless tickets, support sales triage, or replace other application requests. This is not a reason to relax, it is a reason to get moving. Assign real people to this. Break the silos between applications, AI, security, and the business. Give the agentic work a roadmap, not just experiments. Bring in help if you need it. The curve is moving from experimentation to demand management. Companies without a lane for this are going to feel slow quickly. That's it for this week. The full issue with all the links and resources is in your inbox or at QuantumBits newsletter.behive.com. I'm Rezo Marcabadi. Thanks for listening. See you next week.