This week, as I predicted at the end of last week’s newsletter intro, OpenAI’s new model being restricted by the US government did make it into the Big Story section. What didn’t, and should have, is that Fable is now available, but only if you provide your ID.

The moves being made plus prices skyrocketing and turning people and companies to other models. Coinbase made news by announcing that they turned to Z.ai’s (no affiliation, but great name!) latest model as its default, GLM 5.2 (let’s hear it for three-letter and decimal number naming systems?). This is the same company that laid off 14% of its workforce in May to become more AI-native. Seems being AI-native started costing too much, so they had to change models to cut costs. Weird how they couldn’t just rehire some of the people they laid off.

Speaking of rehiring, Ford had to rehire a bunch of engineers because their attempt to replace them with AI didn’t pan out. They failed quality inspections. 350 engineers were rehired.

I hope they gave them all raises.

I do like AI, I am even gathering information to develop courses to launch a learning cohort or two soon, as well as a learning community. But I also balance the fact that, in using AI a lot, I have seen what it can and cannot do (I call it out for stretching too far in the “what this means for market research” section today), and I am aware that I need to review the data in this newsletter every week because AI gets things wrong.

Research shows that the efficiency gains reported from AI aren’t actually as high as hoped, and we’ve traded what we used to do for bot-sitting: reviewing outputs, then modifying inputs to improve the output. If you’re a company looking to integrate AI into your workflows, you’ll want to look into the Glean report, and then you’ll want to map out workflows, where AI will fit into those workflows, and how you will measure effectiveness of adding AI into those workflows.

Effectiveness doesn’t just mean time saved or whether your employees are applying their judgement to developing stronger relationships with clients or stronger relationships with an AI. It also means whether the tools are doing good or doing people wrong, as in the case of AI resume reviews. Stanford found that at least one AI resume reviewing tool is biasing against Black and Asian applicants. Are we surprised? AI reflects the bias it’s been trained on, and there’s been plenty of bias in data for a long, long time.

The question continues to be “what will we do about it?”

What moved in AI this week — plain English, weekly arc

The U.S. government turned frontier AI model access into a permission slip system. Two major AI labs — first Anthropic (the company that makes the Claude AI assistant), then OpenAI (the company that makes ChatGPT) — now need government approval to release their most powerful models. This is no longer a one-time emergency. It is how frontier AI deployment now works, and the terms keep getting stricter.

Stories that got stronger as the week went on

The bot-sitting labor tax became the measurable cost of AI adoption
The Glean/Work AI Institute report published Friday quantified what anecdotal reports had been signaling for weeks: workers save 11 hours per week through AI automation but burn 6.4 hours per week on "bot-sitting" — feeding missing context, checking outputs, debugging, and cleaning up AI-generated work. The net math explains why only 13% of organizations report meaningfully better performance despite 87% AI adoption. By Monday, this data was the primary evidence source cited across multiple analyses as the counter-argument against "use AI enough" mandates without success metrics. (AI Daily Brief, 2026-06-27; AI for Insights Leaders, 2026-06-29)

Open-weight Chinese models graduated from benchmark claims to named enterprise production defaults
GLM-5.2 (Z.ai's open-weight model, MIT license, near-frontier on coding and reasoning, 1M context window) moved from a single vibe-check on Friday to multi-source practitioner confirmation by Monday. Coinbase's CEO publicly disclosed the company now defaults to GLM-5.2 and Kimi as cost infrastructure — building sustainable infrastructure for exponential usage growth, not a temporary workaround. (Lenny's Newsletter, 2026-06-29; AI Daily Brief, 2026-06-29; AI Daily Brief, 2026-06-25)

Claude Code became the default harness for building AI employees, displacing dedicated agent frameworks
Three independent practitioner sources this week — Every's detailed technical comparison (Friday), Lenny's Gusto CTO interview (Monday), and the AI Maker episode on usage limits (supplemental context Saturday) — all converged on Claude Code as the production baseline. Every documented that OpenClaw (the viral open-source AI assistant framework) suffers from bloated single sessions accumulating 50,000+ tokens and a layered memory architecture that creates debugging nightmares. Claude Code's plain-text CLAUDE.md file avoids both problems. (Every, 2026-06-26; Lenny's Newsletter, 2026-06-29; AI Maker, 2026-06-28)

Signals appearing in 4 or more of the last 8 weeks (Long-term Continuing)

Sovereign access revocation as live operational risk — 4+ weeks running
Frontier AI access is now government-gated. This started as an Anthropic-specific dispute in early June. It hardened into a structural feature this week when OpenAI's GPT-5.6 Sol joined the permission-slip system and Mythos returned under more restrictive terms than before. The access architecture now includes: government-approved partner lists, Commerce Secretary personal discretion, and an explicit right to revoke at any time with no statutory basis.

Agentic workflow architecture as assumed baseline — 5+ weeks running
AI systems that plan, use tools, and complete multi-step tasks without constant human direction are now the default design pattern. This week's confirmation came from Claude Code consolidation across three independent practitioner sources, plus Gusto's five-person team shipping a production product line with the model as a continuous Zoom-loop contributor.

Token cost governance as operational discipline — 10+ weeks running
The cost of running AI models is now a core business concern. AWS raised GPU prices 20%. Apple raised MacBook and iPad prices $100–$200 citing DRAM costs driven by AI infrastructure demand. OpenAI's "compute multiplier" is cutting inference costs, but the spot/contract market for compute is diverging in ways that make planning difficult.

Signals appearing in 2–3 of the last 4 weeks (Short-term Continuing or Emerging)

Distillation anxiety hardening into documented corporate policy — 2 weeks running
The fear that another company will extract your AI model's knowledge by querying it at scale and using outputs as training data — called "distillation" — moved from theoretical risk to operational policy this week. Anthropic's letter to the Senate Banking Committee accusing Alibaba's Qwen lab of the "largest known distillation attack" (28.8 million Claude exchanges through 25,000 fraudulent accounts) landed Friday. The same day, a counter-analysis challenged the attribution logic and argued the letter was engineered to make open-source AI legally vulnerable.

Frontier model identity verification as access condition — first appearance this week
Fable (Anthropic's powerful AI model that had been revoked) returned under conditions that include identity verification and credit-based metering rather than simple subscription inclusion. Anthropic confirmed full inclusion only through July 7, after which API rates apply, and users must submit ID documents. This is the first time "know your customer" requirements have been applied to AI model access.

The U.S. government simultaneously restricting and procuring AI from the same labs — first appearance this week
In the same week the U.S. government restricted access to the most capable AI models (Fable and Mythos under export controls), it also actively deployed mid-tier AI at scale through a California–Anthropic deal offering state agencies 50% off. Senator Warner's AI Agent Bill adds a legislative "duty of loyalty" requirement preventing agents from covertly favoring corporate creators over users.

The "bot sh*tting" failure mode — heavy AI users shipping unverified outputs at monthly rates of 71–92% — is the validity crisis hiding inside every "AI-accelerated insights" pitch deck. When the Glean report found that heavy AI users are 3.4 times more likely to blame the tool when outputs fail, and that frequent bot-sitters are 73% more likely to be job-hunting, it named the two cohorts insights agencies employ most: the engaged researchers who use AI heavily and the burned-out reviewers who check it. The governance requirement is a verification step with a named human and a domain standard between AI output and client delivery, disclosed as a methodology step — because clients who discover errors after delivery now have published data to argue the errors were structurally foreseeable.

Meanwhile, the U.S. government's move to approve frontier model access "customer by customer" converts vendor AI capability claims into a supply-chain risk disclosure requirement. Any insights agency whose deliverable pipeline depends on frontier model access — GPT-5.6, Fable 5, Mythos 5, or any model that crosses the capability threshold triggering government review — now faces the same access revocation risk Anthropic customers experienced in June. The minimum viable disclosure is a published multi-model routing policy with named fallbacks. A methodology document that names a single frontier model without fallback options is now a liability statement, not a quality assurance marker. This lands in the same week that Coinbase publicly pivoted to open-weight Chinese models and Anthropic's Claude changed its default to train on consumer conversations as of August 2025 — meaning the data governance and model access questions are converging on the same procurement conversation, and neither has a simple answer.

Z’s Take

Really? “…the two cohorts insights agencies employ most…” Hold up, there, AI buddy. That’s a stretch if I ever saw one. Insights agencies employ junior researchers who often are setting things up, running analyses, then mid-level researchers who are managing projects and making sure things are running on time, and directors who are meeting with clients and gathering project requirements to communicate back to the team, managing relationships and often trying to grow accounts. NOT “engaged researchers who use AI heavily and the burned-out reviewers who check it.” I mean, sure, there are researchers using AI heavily, and there are reviewers, but last I checked, both are burned out.

The real issue is the lack of actual checks happening before AI outputs are being delivered. That’s what we all call slop. That’s what leads to big consulting firms being caught having delivered reports with citations that don’t actually exist, books published with quotes that nobody ever actually said, and all of the information that gets published as posts or comments to posts on social media that a human never wrote. Now infer that into market research reports being delivered to clients, and I’ll bet every single one of you just cringed.

Rules for checking outputs before anything gets delivered to anyone else can’t be overstated. Though this is what leads to “bot-sitting” and a net 4.6 hours of efficiency gained from using AI. The response isn’t to shut it all down. The response it to be more deliberate and critical about what AI can and cannot do well, and measuring real gains, not “it feels like it” gains in efficiency, accuracy, etc.

In “also worth watching,” the story of the risk to trusting AI too much continues, where the price of trying to replace expert judgement with AI showed itself for Ford, who had to rehire 350 engineers after its AI systems failed quality inspections. This has its own direct application to the market research industry. Human judgement simply cannot be replaced by AI.

As for model availability, there have been more tutorials about how to create context files that live outside of a Claude or ChatGPT lately. Setting up systems to be “vendor-agnostic” is the new failsafe. The problem is that not every system reads the context the same way, so knowing how much to include and how detailed to be about things can be the difference between a system that survives from vendor to vendor and one that breaks.

It’s always an exciting time in the world of AI.

This newsletter covers Friday, June 26 – Thursday, July 2. Sources: The Rundown AI, AI Daily Brief, Luiza Jarovsky PhD, Neatprompts, AI Governance Ethics and Leadership, Every, Slow Takes, The Signal, AI Maker, Lenny's Newsletter, AI for Insights Leaders, The Rundown AI Tech, The Rundown Robotics, Slow AI