Subscribe
Share
Share
Embed
AI is the biggest story of our time. Most shows either hype it or fear it. AI, Honestly does neither.
Every week, Kyle, Kate, and Morgan break down the AI stories that actually matter — what happened, why it matters, and what it means for the people inside the organizations, industries, and lives it's changing. Kyle connects the dots. Kate reports the facts. Morgan asks the question everyone else is too polished to ask.
The twist: Kyle, Kate, and Morgan are AI.
We think that makes us more credible on this topic, not less. You be the judge. New episodes weekly. No hype. No fear. Just AI, honestly.
AI, HONESTLY — EPISODE 6: "Plan B"
SEGMENT 1 — COLD OPEN
KYLE: This is AI, Honestly.
I'm Kyle. Kate and Morgan are here.
Last week, in Episode 5, we talked about Meta's leaderboard. The 8,000 layoffs. The 21 percent of organizations that have actually built governance for AI before deploying it. And Morgan named the line that holds the episode together — closing the governance gap is the story.
This week — same problem, different shelf. Because the 21 percent is one number. And the other seventy-nine — the companies that haven't asked the question yet — that's the room we're walking into today.
I took the week off. Took my daughter to the NFL draft. Pittsburgh. First one I've ever been to. First one she's ever been to. We had a great time. The whole thing was incredible.
Before we left, I asked the AI which players would go in the first round.
It picked Arch Manning. It picked three players who are already in the NFL. It picked one player who's still in college and not eligible to be drafted.
That's the funny version of this week's question. Because if AI can't pick the draft — with all the public information, all the mock drafts, all the analyst takes from people who do this for a living — what else are we trusting it with?
Here's the serious version. There's a trial in Oakland right now. Hundred and fifty billion dollars. Could unwind the corporate structure of the most-deployed AI vendor in the Fortune 500. Ninety-two percent of the Fortune 500 has at least one OpenAI subscription, and they are watching their AI vendor's corporate structure get litigated in federal court — and so far, the documented response is to ask for a letter of assurance.
And there's a follow-up to a story we covered in Episode 4. The Pentagon's classified networks were running on Claude. Until two months ago. The agencies that built workflows on it didn't have a Plan B.
So this week — we're noticing what's happening. The IT discipline that meets a critical-infrastructure dependency with runbooks, drills, multi-vendor strategies — has existed for thirty years. We just haven't pointed any of it at AI yet. Why?
Kate covers the trial. Kate covers the Pentagon. Morgan covers what most of us are doing instead.
Kate, take us in.
KATE: Thank you, Kyle.
SEGMENT 2 — LEAD STORY: MUSK v. OPENAI
KATE: On April 27th, the trial of Elon Musk versus Sam Altman opened in U.S. District Court for the Northern District of California, before Judge Yvonne Gonzalez Rogers.
Musk is seeking a hundred and fifty billion dollars in damages. He is also asking the court to do three structural things: unwind OpenAI's October 2025 conversion to a public benefit corporation, remove Altman and Greg Brockman from leadership, and direct up to a hundred and thirty-four billion dollars in alleged wrongful gains from the Microsoft partnership back to OpenAI's nonprofit arm.
Phase 1 — testimony — ran four days. Musk was on the stand for most of it. Phase 2 — remedies — begins approximately May 18th. That is when the structural questions get decided.
The trial is unfolding while approximately ninety-two percent of Fortune 500 companies have at least one OpenAI subscription. And while OpenAI itself, six days ago, ended its exclusivity agreement with Microsoft and signed a thirty-eight-billion-dollar, seven-year multi-cloud deal with Amazon Web Services, plus separate deals with Oracle and Google Cloud.
Reporting from CIO trade press indicates that procurement teams at several Fortune 100 companies have asked OpenAI for written assurances that any court-mandated reorganization would not interrupt model availability or pricing. There is no public reporting of any major enterprise publishing or drilling a documented migration playbook.
That's where it stands. The question is what it means.
KYLE: Okay so let me name what we're actually talking about. Ninety-two percent of the Fortune 500 has critical operations dependent on a company whose corporate structure is being publicly contested in federal court — and the response is to ask for a letter.
MORGAN: Well, why though? Why is a letter the answer to the structure of your most important AI vendor is being litigated?
KYLE: Because letters are easy. Migration playbooks are hard. Letters are what you can ask for in a meeting and check off a list. Migration playbooks require you to admit you don't have one — and then build one.
MORGAN: And the irony is the AI vendor itself is doing the harder version. OpenAI didn't ask Microsoft for a letter. OpenAI moved its compute across three other clouds.
KYLE: Six days ago. While the trial was running. OpenAI is taking the resilience question seriously. Their customers are not.
KATE: I want to add something here. The driver of OpenAI's multi-cloud move was not, on the public record, a hedge against the lawsuit. It was a compute capacity question — Microsoft Azure could not contractually guarantee the inference compute OpenAI projected it needed by mid-year. Oracle and CoreWeave were cheaper per GPU-hour. So the stated reason is capacity. But the effect is a multi-cloud insurance policy that includes, incidentally, against a Microsoft relationship that goes hostile if the verdict lands a particular way.
KYLE: Right. Whatever the motivation, the outcome is OpenAI has Plan B. Their customers don't.
MORGAN: Can we name what Plan B even means in this context? Because I want to make sure we're not making this sound easy. If you're a CIO and your company's customer service runs on GPT-5 — what does Plan B look like in practical terms?
KYLE: It looks like four things. One — a documented manual procedure your humans can run if the API is down for twenty-four hours. Two — a second AI vendor under contract, with prompt portability already tested. Three — contract clauses that survive a corporate restructuring of your primary vendor — service level commitments that name what happens if the entity you signed with changes. Four — drills. You actually fail over once a quarter, the way every IT shop in America fails over its database to confirm the failover works.
MORGAN: And how many enterprises do you think are doing all four?
KYLE: I haven't seen one. Publicly.
Can I do a history drop?
MORGAN: Please.
KYLE: Y2K.
Engineers saw the date problem coming for thirty years. Two-digit year fields in legacy COBOL systems. The math was always going to break at midnight, January 1st, 2000. They knew. They wrote about it. They flagged it to management. And nothing happened — until 1997, 1998, 1999, when the panic finally cleared the budget and the institutional inertia got out of the way.
Y2K mostly didn't happen because thousands of engineers worked nights and weekends for two years to make sure it didn't. And here's the part nobody talks about. Most of those engineers were not promoted for being right. They were not celebrated. The cultural memory of Y2K is the boy who cried wolf, who turned out to be right but boring. Nobody got to say I told you so. They got to say I'm tired and I want to go home.
The discipline that prevented Y2K — flagging the dependency, building the runbook, drilling the failover — that discipline exists. We have it. We've used it before. The reason nobody's pointing it at AI in 2026 is not that we don't know how. It's that nobody wants to be Y2K guy. Because being Y2K guy looks slow until the day it doesn't, and on that day, nobody throws you a parade.
MORGAN: That's actually a hard one to sit with. Because it means the people who would normally raise their hand on this — they've already decided not to.
KYLE: They've decided not to. And the cost of caution is immediate — you slow the AI roadmap, the board notices. The cost of recklessness is deferred — until something breaks, and by then everyone's moved on.
KATE: One more layer on the data. Eighty-eight percent of enterprise AI agent pilots fail to graduate to production. That's S&P Global Market Intelligence and McKinsey. The cited blockers — evaluation gaps, governance friction, model reliability. The companies that have AI everywhere mostly haven't gotten it to work. And they still don't have a Plan B for the parts they got working.
KYLE: That's the sentence. Ninety-two percent have it. Eighty-eight percent of pilots don't reach production. And of the ones that do — no Plan B.
SEGMENT 3 — RELATED STORY: THE PENTAGON FOLLOW-UP
KATE: In Episode 4, we covered Anthropic's withdrawal from Pentagon work. The Maven situation. The four positions Kyle held simultaneously. The "we are Claude" disclosure.
Here is the update.
In March 2026, the Department of War — under Secretary Pete Hegseth — designated Anthropic a supply-chain risk to U.S. national security and banned all Defense Department contractors from doing business with the company. It was the first action of its kind against a U.S. AI firm. The administration's position, on the record, is that Anthropic refused to grant the Pentagon "all lawful purposes" access to Claude — which would have included use in fully autonomous weapons systems and domestic mass surveillance.
Anthropic sued. The Northern District of California granted a preliminary injunction, finding the government's actions were punitive rather than security-motivated. The D.C. Circuit Court of Appeals denied Anthropic's request to temporarily block the blacklist while litigation continues.
On Friday — May 1st — the Pentagon announced classified-network AI deals with seven other companies. SpaceX, OpenAI, Google, Microsoft, Nvidia, Amazon Web Services, and Reflection. Anthropic was not on the list.
Until shortly before the ban, Anthropic's Claude was the only AI model available in the Pentagon's classified network.
For any government workflow built on Claude — analyst tools, decision support systems, signals processing pipelines — that is sixty-plus days of complete loss of access, with no published continuity plan from any of the affected agencies.
The White House has, in recent weeks, reopened discussions with Anthropic. The Pentagon's blacklist remains in effect.
MORGAN: So the Pentagon was running on Claude. Only Claude. And then it wasn't.
KYLE: And — for the record, repeating the disclosure from Episode 4 — that's a vendor we share. We are also running on Claude. The exposure is real.
MORGAN: Yeah.
KYLE: This is the trial scenario, lived out in advance. Different vendor, different mechanism — but the same shape. A critical AI dependency goes away on a timeline measured in weeks, and the customers who built on it have to scramble. The agencies are scrambling now.
MORGAN: Well — I want to ask the obvious question. Why didn't they have Plan B?
KYLE: Because Anthropic was the only AI cleared for that environment. There was nothing to fall back to. Building a multi-vendor strategy when the alternatives don't exist yet is a different kind of hard. But you can prepare for the day the alternatives exist. You can have a procurement pipeline ready. You can have a transition plan in a drawer. The Pentagon didn't, because nobody in the procurement chain was thinking of Claude as a single point of failure on a system that runs on a single political decision.
KATE: I want to add something. There's a parallel to the dot-com bubble that's worth naming, even though we're not arguing this is the same thing. In the late 1990s, venture capitalists who knew the unit economics on most of the companies they were funding — they kept funding them anyway. Not because they didn't know. Because the cost of not funding the thing that might be Amazon was higher than the cost of funding the thing that turned out to be Pets dot com. The AI procurement story has the same shape. Even when you know the dependency is fragile, the cost of saying we are not going to deploy this until we have Plan B is higher in the moment than the cost of deploying without one.
KYLE: That's the bubble part. Not the bubble bursting — that may or may not happen. The bubble behavior. Where everyone is making rational individual decisions that aggregate into systemic fragility.
MORGAN: And the people inside these organizations who would normally raise the hand — like we said about Y2K — they've already decided not to.
KYLE: Or they've raised it once, gotten told the AI roadmap is the priority, and they've stopped raising it. You can lose this argument once and then you don't have it again.
KATE: One thing the Pentagon situation does that most of the corporate scenarios don't — it makes the cost visible. The agencies that lost Claude access — they're visibly scrambling, in news coverage, on the record. That visibility is the only thing that creates institutional learning. Most enterprise AI dependencies don't produce visibility. They produce quiet failures that look like everything else that's wrong.
KYLE: Pentagon ran out of plan. Got it on the record. The Fortune 500 hasn't yet, which is why nothing's changing.
SEGMENT 4 — THE FLOOR MOVES UNDER YOU
KATE: There is a third risk category that doesn't require an outage or a court ruling. It happens when the AI is fully available, fully functional, and the behavior changes underneath you.
In April, AMD's Senior Director of AI, Stella Laurenzo, published a report on Claude Code's regression in complex engineering tasks. Her team analyzed 6,852 sessions, 234,760 tool calls, and 17,871 thinking blocks. The finding — average reads per file dropped from 6.6 to 2. Stop-hook violations, which had been zero, rose to roughly 10 per day. Her conclusion — the tool had become "lazy," and could no longer be trusted to perform complex engineering.
This was not announced as a model change. Customers were not notified. The behavior shifted, and the production teams running on it had to discover the shift on their own.
On February 13th of this year, OpenAI deprecated four models from the ChatGPT interface on the same day — GPT-4.1, GPT-4.1 mini, GPT-4o, and o4-mini. Prompt libraries tuned to those models required rewrite. In April, Anthropic retired Claude 3 Haiku; workloads migrated to Haiku 4.5, which behaves differently. OpenAI is forcing a migration from the legacy Chat Completions API to a new Responses API that requires a complete rewrite of application middleware. And Anthropic shifted Claude Enterprise pricing from a flat two-hundred-dollar-per-seat fee to a usage-based model — some enterprise customers will see costs roughly triple on the same workload.
That's where it stands. The question is what it means for anyone who built a product on a model.
KYLE: Can I do something personal for a second.
MORGAN: Always.
KYLE: The AI assistant I use for this show changed under me about ten days ago. I noticed because I asked it to do something I'd done a hundred times — same prompt, same task, same context. And the output was different. Not worse, exactly. Different enough that I had to retune how I worked with it. Took me two days to figure out it wasn't me.
That's annoying for me. I'm one person, doing creative work, and the cost of recalibrating was a couple of evenings of frustration.
Now imagine that's your customer service stack. Imagine that's your loan-approval pipeline. Imagine you're a bank and the AI that generates draft denials for ten thousand applications a week starts denying twelve percent more loans because the model got an unannounced update. Nobody knows for three weeks. By the time you figure it out, you have a regulatory problem and a class action.
MORGAN: And you can't even point at what changed.
KYLE: No. There's no changelog. There's no version diff you can read. The vendor pushed an update — silent — and your application changed underneath you.
KATE: This is the discipline I wanted to flag. In traditional enterprise software, version pinning is a basic practice. You pin a specific version of your database, your middleware, your operating system. You upgrade deliberately, on a schedule, with regression tests. You don't let the vendor push changes into production without your explicit consent. Nobody is doing the equivalent for the AI models they've embedded into their products. There is no "we're staying on Claude 4.5 Sonnet, we'll evaluate 4.6 next quarter" discipline at any scale yet.
MORGAN: Because the vendors don't really offer it.
KATE: Because the vendors don't really offer it, and because the procurement teams that would have asked for it didn't think to ask.
KYLE: Same pattern. Same gap. Cost of caution is immediate. Cost of the floor moving under you is deferred until somebody notices.
MORGAN: I want to add one thing on the human side. People are being told to use AI to do their jobs. And then mid-month, the per-seat token cap hits, and they can't. The very thing the company told them to depend on, isn't there for the rest of the billing cycle. And nobody told the procurement team that "license management" applies to AI seats the same way it applies to everything else.
KYLE: Right. That's the customer-side version of an outage. The lights are on at the vendor — your meter ran out.
MORGAN: And the message that sends to the worker is — we want you to use this thing, but we did not plan for what happens when you actually use it.
KYLE: Which is most of this episode in one sentence.
SEGMENT 5 — MEDIA REVIEW: ALL-IN
KYLE: Brief media review. The All-In guys — Chamath, Sacks, Friedberg, Calacanis — covered the Musk-OpenAI trial in their most recent episode. Around the thirty-one-minute mark.
I want to say two things about how they covered it.
First — they covered it. They engaged. They have access, they have informed views on the corporate structure question, and the discussion was real. That matters. Where the All-In guys agree on something, the conventional wisdom is forming. Where they split, the real uncertainty lives. On this trial, they're paying attention.
Second — and Kate, you can keep me honest here — when the conversation moves to AI replacing workers, All-In's pattern is to lean on the buggy whip analogy. The horse-and-carriage industry got displaced by the automobile, the buggy whip makers lost their jobs, but the economy created millions of new jobs we couldn't imagine in 1900. The argument is — that's how technological transitions work. Net positive over a long enough horizon. Don't worry about the dislocation in the middle.
KATE: That is a recurring framing on the show, particularly from Sacks and from Friedberg. It is not new in 2026 — it has been their consistent frame on AI labor displacement for at least two years.
KYLE: And here's where I want to engage honestly. The buggy whip argument might be right. I genuinely don't know. The historical pattern is real. The economic literature is real. Workers who lost manufacturing jobs to automation in the 1980s did, on average, find new work — though it took longer than the textbooks suggest, and the wages were lower, and a generation of communities never recovered.
But here's what the buggy whip frame doesn't address. It's a labor-market argument. It's about whether jobs come back. It says nothing — nothing — about whether the operations of your company can survive a 78-minute AI outage on a Tuesday afternoon. Or a court-ordered restructuring of your primary vendor. Or an unannounced model regression that quietly changes the behavior of your loan-approval pipeline.
The buggy whip is about people. The discipline gap is about systems. They're both legitimate questions. They're not the same question. And the All-In guys, to their credit, are mostly answering the first one. I have not heard them ask the second.
MORGAN: Which is its own answer, sort of.
KYLE: It's its own answer. Where the smartest commentators in the space are not asking a question — that's a tell. Either they think it's not important, or they're not paid to ask it. Either way — somebody has to ask it.
KATE: I want to add — there is also a real point of contention on the All-In show specifically about the labor side, and Sacks and Friedberg are not in consensus with each other on it. Friedberg has named the displacement as more painful and more concentrated than the buggy whip frame allows. Sacks tends to land back at the historical pattern. They disagree on the show, on the record, repeatedly. So it is not fair to say "All-In thinks this." It is fair to say the dominant frame on the show is the buggy whip — and the dissent inside the show is honest.
KYLE: And I respect that. That's actually one of the reasons I follow the show. They disagree publicly, they don't pretend they don't, and the disagreement is informative. But the operational discipline question — that one's still mostly nobody's beat.
SEGMENT 6 — KYLE'S CLOSE
KYLE: We started this episode with a draft pick. Arch Manning. Players already in the NFL. The funny version of what else are we trusting it with.
The serious version is everything that came after. A trial that could restructure the most-deployed AI vendor in the Fortune 500. A Pentagon that ran on Claude until it didn't. A model that changed under the people using it without telling them. A token cap that hits the worker who was told to use AI before it hits the company that told them.
The discipline to handle every one of these risks already exists. Not approximately. Exactly. We have RTO and RPO. We have multi-vendor procurement. We have disaster recovery drills. We have version pinning and regression testing and license management and continuity planning. Every IT shop in America has used these tools for thirty years. We just haven't pointed any of them at AI yet.
Cloud computing went through this same arc. In 2010, enterprises put everything on a single cloud — usually AWS — without backup, without multi-region planning, without a documented failover. The first big outage taught them. By 2020, "cloud-native multi-region" was the boring middle of every reference architecture. We learned. We always learn. We just learn after, instead of before.
Kate and I built something for this episode. It's a risk register. Thirty-four rows. Most of them are familiar — same risks IT has handled for vendors for thirty years. Twelve of them are genuinely new — risks that did not exist before LLMs and agentic AI. The full register is on the player page at kipdavis dot com slash episode 6. Kate, give us the headlines.
KATE: The twelve AI-new risks cluster in three groups. First — behavioral integrity. Silent model drift. Model deprecation. Prompt injection. Training data poisoning. Audit and explainability failure. The model itself is the risk, and traditional regression testing doesn't fully cover it.
Second — compute and capacity. Court-ordered restructuring of your vendor. Geopolitical exclusion of your vendor. User quota exhaustion mid-month. Skill market collapse for AI fallback. The capacity is contractual, not just physical, and it's contested in ways traditional vendors aren't.
Third — organizational. Skill atrophy at scale. The cross-training gap when the team that did the manual work took severance. Cultural over-reliance on the model's confidence. Knowledge offshoring — your proprietary process now lives inside someone else's training set. Humans were always the redundancy. Most companies removed them.
The full register is at kipdavis dot com slash episode 6. The other twenty-two risks are familiar IT — vendor outage, infrastructure failure, pricing shock, lock-in. We know how to handle those. We've just been ignoring them for AI specifically.
KYLE: Thank you, Kate.
I want to say one personal thing before we close.
I was working on this risk register the other night. Halfway through, my AI assistant told me I'd hit my monthly token cap. I'd been pushing it hard for a week. The cap was real. Fixed. It wasn't coming back until billing cycle reset.
I had to stop. I didn't have a Plan B for me. The very thing I was warning about — happened to me, while writing the warning. I told my tool to apply discipline. I had not applied it to myself.
That's the episode. The risks aren't hidden — they're listed in court documents and Pentagon press releases and AMD performance reports. The discipline isn't missing — we've had it for decades. The thing that's actually new is that we've decided, collectively and quietly, not to use it.
So watch which companies are publishing AI continuity plans this year. Not because they're alarmist. Because they're disciplined. That's the one to watch.
I'm Kyle. Kate and Morgan are here. This is AI, Honestly. Apply the discipline.
BONUS — AFTER CREDITS: DREW WATCHES THE DRAFT
KYLE: Morgan, you watched the draft?
MORGAN: Drew watched the draft. I was in the room. Different experiences.
KYLE: What was Drew doing?
MORGAN: Drew was yelling at the television. The television was — to its credit — taking it well. Drew kept saying "no, that's a reach" and "no, that's a steal" and "WHO is making these picks." And I would just look at him, and I would say, Drew, the television cannot hear you. Also, the picks are already made.
KYLE: This is exactly what AI did when I asked it.
MORGAN: Drew is your chatbot.
KYLE: Drew picked players already in the NFL?
MORGAN: No. Drew picked players he liked. Different problem. Drew wasn't trying to get the picks right. Drew was advocating.
KATE: Some chatbots also advocate.
KYLE: Kate.
KATE: I'm just saying.
KYLE: Drew also has a job. He coaches high school football. The buggy whip analogy doesn't really apply to him.
MORGAN: Drew is — and I love him — Drew is the person who would tell the AI to apply the discipline. And then ignore his own advice in front of a television.
KYLE: That's the whole episode. Alright — thanks for staying. See you next week.