Quanta Bits - June 20, 2026

Hi everyone,

Love the World Cup vibe: all these nations centered on a single event. Sometimes it feels like the World Cup is a stronger version of what the Olympics tries to be. All in all, it has been an exciting week of watching football/soccer at various times of the day.

At the same time, it has been a slow week for frontier models. Since Fable was shelved due to government restrictions, OpenAI and Anthropic seem to have gone into a cautious release mode. Very little has come out for Claude Code or Codex these past few weeks.

That's a welcome relief for organizations in some ways. Instead of chasing the latest, we can continue testing and piloting. But a pattern that emerges, while I'm all on board with pilots, is that production tests are drifting toward "automate what exists" instead of "first, figure out the ideal end state." We have powerful technologies at our fingertips, but under management pressure to go fast, with plates already full of normal day-to-day work, we tend to take the shortcut just to show we are doing something. That won't scale. This is the topic for this week's essay.

Cheers.
Reza

The old automation mistake is back, and AI just made it easy enough to make it casually. I'll admit the appeal: when a thing is suddenly this easy, the temptation is to point it at whatever annoys you most and let it rip. I've felt that pull. But the most annoying process is usually the messiest one, and pointing an agent at a mess doesn't clean it up. It just runs the confusion faster.

Take deal review. A team wants an agent to handle the deals that need Finance sign-off: anything over a threshold, discounts outside policy, thin margins, odd payment terms, anything that touches revenue recognition. On paper it's a perfect AI job: read the opportunity, check the quote, flag the risk, draft the finance packet, route the deal.

Then the details show up. The quote says $62K ARR; the order says $48K net ARR. Finance has its own spreadsheet to check the ARR it cares about. The order form has quarterly billing. Legal slipped in a termination-for-convenience clause. And the threshold question is rarely just "does this cross $50K?" It's which $50K: ARR, ACV, TCV, net-new, or total contract value. A $45K deal on a three-year term with odd payment terms may need more scrutiny than a clean $60K one.

AI can make all of that faster. It can't make it clean on its own. None of this is new, by the way. Michael Hammer was warning companies about it back in 1990, with a line I still love: stop paving the cow paths (that should resonate with Boston residents!). What's new is the entry cost. Any team can now wire an agent, sanctioned or not, across Salesforce, Slack, Gmail, and the document store. So the risk is no longer a bad summary. It's inherited permissions and triggered downstream work.

A framework I learned for scaling a process from my operations days at EMC, a company I still consider one of the best operations organizations I've seen, was: make it consistent, make it simple, then decide how to scale. I've leaned on it for years, long before agents showed up, and it still works.

Consistent means the trigger, the handoff, the decision rule, and the result mean the same thing every time. Simple means as few steps, inputs, outputs, and exceptions as the work can safely carry.

AI doesn't remove that test. It raises the stakes on it. A vague process doesn't get clearer when you point an agent at it; the agent just moves the confusion faster. A political handoff gets its politics routed faster. An untrusted source field gets a clean summary nobody believes.

Back at the deal desk, the tempting shortcut is to point an agent at the Salesforce quotes and ask it to infer the review process from the data exhaust. And it'll give you something. It'll look confident. But it's guessed around the actual operating question, and a confident guess is the most dangerous kind. Worse, it learns the company's workarounds and rubber stamps as if they were the process. Let it write back to Salesforce, Slack, and CPQ, and now you have inconsistent commercial data across systems with no obvious owner to clean it up.

The better path usually means Finance, Sales Ops, Legal, RevOps, and business applications redesign the review path together. Finance owns the ARR definitions, RevOps owns the CRM and quoting data, Legal owns the terms exceptions. Someone has to own the seams between them.

An HBR piece from OpenAI and Bain calls the pattern the "micro-productivity trap": tasks get faster, but value stalls because the surrounding workflow still depends on manual handoffs, tacit knowledge, and legacy systems. Their example was a quote process: the company changed who did the early bid work first, then used AI to support the new workflow. The quotes got faster because the process changed first, not the other way around.

The practical move: understand the process before the agent gets access, not after something breaks.

The visible pain point is usually just the top of the iceberg. A slow deal review looks like a routing problem. Underneath it may be a pricing-policy, data-trust, ownership, or integration problem.

AI can help once you understand what you're really touching. It can draft the packet, check policy, summarize risk, and route the obvious cases. But the license fee is the easy cost to see. The real bill is the plumbing underneath: process redesign, data cleanup, training, and exception handling.

So before I'd automate a process like this, I'd first find out who the process owners and stakeholders are, what their real pain points are, how involved it would be to redesign the future-state process, and whether the time and cost of automation are worth it. Ironically, you can use AI here too: feed it current process information (meeting notes, workshop transcripts, documents, and samples), and ask for a first-pass design of what the new process should be. You shouldn't need a giant consulting project for that first pass.

If the answer is yes, automate with confidence. If the answer is no, slow down and understand the work well enough to decide whether it deserves automation at all. That's what I mean when I say earn your complexity. AI lets you bolt on a layer of automation for almost nothing now, but the layer is only worth having once the process underneath is consistent, simple, and owned. Earn that part first. Then, by all means, let AI make it fly 🪁

@staysaasy on X (325.6K views, Apr 25, 2026).

Google gets a legal warning on AI summaries. A German court held Google liable for false statements in AI Overviews after the tool described publishers as scams and Google did not fix it after a cease-and-desist. The useful part is the logic: an AI summary is not just a link to someone else's words; it is a new statement by the product. For CIOs and risk teams: disclaimers are not a control if the system is confidently restating facts about customers, vendors, or people. Source: Ars Technica

Anthropic starts planning for overnight capability jumps. Anthropic's new research arm published an agenda that includes monthly worker surveys, threat research, and "fire drill" exercises for sudden AI capability surges. That language matters because it sounds less like software release management and more like critical infrastructure planning. For CIOs and operational-resilience teams: vendor due diligence now needs a question about what happens when the model you depend on changes materially between quarters. Source: Anthropic

Pew's June 2026 survey is American public opinion, not a global read. Still, the pattern is useful: about half of U.S. adults now use chatbots, up from a third in 2024, and 24% use them daily. The trust gap is the part to watch: 40% expect AI to have a negative impact on society, versus 16% positive, and 71% say AI will make their personal information less secure. If you are selling, deploying, or governing AI in the U.S., that is the room you are walking into.

Stanford HAI's Diyi Yang lab built CARE, an AI practice system where novice counselors rehearse with role-played patients and then get feedback from an AI mentor trained with help from therapist supervisors. The useful twist is that the simulated patients are designed not to be agreeable. They push back, hold back, and resist easy advice, which is closer to real counseling than a chatbot that politely accepts every suggestion. In a 90-person randomized trial, practice alone built confidence, but practice plus feedback measurably improved skill: counselors became more empathetic and more client-centered. This is the hopeful version of the AI-in-human-work story: not replacing the human, but giving under-resourced people a safer place to practice.

Source: Stanford HAI

Dir. Thomas Vinterberg | 115 min | ★★★★◐ | Mads Mikkelsen, Thomas Bo Larsen, Annika Wedderkopp

A small-town Danish kindergarten teacher gets caught in the crosshairs of an accusation that takes on a life of its own. What makes the film so effective is how it captures the mechanics of a community turning on one of its own without anyone stopping to question what they actually know. It reminded me of Middlemarch, oddly enough, where gossip and social consensus harden into accepted truth in insular communities. Mads Mikkelsen is fantastic in the lead, and Annika Wedderkopp, who plays the child at the center of it, is remarkable for a six-year-old. Tough watch, but the kind of movie that stays with you.

Prefer to listen? Quanta Bits is also available on Apple Podcasts and Spotify.

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. 🖖