← All Episodes

AI, Honestly

Three AI voices  ·  One honest conversation

Episode 006  ·  Now Playing

"Plan B"

~25 min  ·  May 2026

▶ Now Playing

This Episode

The risks of AI dependency are public, named, and recent. Outages. Vendor restructure. Model rotation. A $150 billion trial that could unwind the corporate structure of the most-deployed AI vendor in the Fortune 500. None of it is hidden. The IT discipline to handle critical-infrastructure risk has existed for thirty years. We just haven't pointed it at AI yet. Why?

Cold Open

Draft Day, AI's Worst Showing

Kyle took the week off for the NFL draft. Asked AI to predict the picks. It picked players already in the NFL. The funny version of the question that runs through this whole episode: if it can't pick the draft, what else are we trusting it with?

Story 1

Musk v. OpenAI — and Nobody's Plan B

$150 billion. Structural remedies on the table. Approximately 90% of the Fortune 500 has at least one OpenAI subscription. The CIO response so far: a request for a letter of assurance. Meanwhile OpenAI itself just ended Microsoft exclusivity and signed multi-cloud deals across AWS, Google, and Oracle — building Plan B for itself. Their customers aren't.

Story 2 — EP004 Follow-up

The Pentagon Found Seven New AI Vendors

In EP004 we covered Anthropic stepping back from Pentagon work. May 1, 2026: the Pentagon announced classified-network deals with seven other AI companies — explicitly excluding Anthropic. Government workflows built on Claude have been cut off for 60+ days. Nobody had Plan B. The trial is the loud version of a story that already happened.

The Pattern

Y2K + Bubble I — We've Done This Before

The hype cycle outruns lived failure. The status game punishes prudence. And nobody wants to be Y2K guy — vindicated, exhausted, and never promoted for being right early. The discipline already exists. We've used it for thirty years. We just haven't pointed it at AI.

Takeaway: The AI Risk Register

The discipline already exists for every other vendor. Most of this list is familiar IT territory — we know how to handle it. A handful is genuinely new. Both halves matter. Below: the twelve AI-new risks the discipline hasn't caught up to yet, with the full 34-row register expandable below.

AI-new  ·  Behavioral / Integrity

Silent model drift

AMD tracked 14 model releases in one month — same prompt, opposite results. Discipline action: behavioral baselines, output regression monitoring, vendor changelog discipline.

AI-new  ·  Behavioral / Integrity

Model deprecation / sunset

Feb 13: GPT-4.1, 4o, o4-mini retired the same day. April: Claude 3 Haiku. Discipline action: treat model upgrades like database upgrades; regression test the prompt library.

AI-new  ·  Behavioral / Integrity

Prompt injection

A vulnerability class deterministic software didn't have. Live and growing. Discipline action: input sanitization treated as an OWASP-equivalent discipline.

AI-new  ·  Behavioral / Integrity

Training-data poisoning

Documented in research; not yet a public corporate incident. Discipline action: provenance review on third-party data feeding the model.

AI-new  ·  Behavioral / Integrity

Audit / explainability failure

"Why did the AI deny that loan?" — already biting in lending and insurance. Discipline action: decision logging and explainability written into procurement, not bolted on.

AI-new  ·  Compute / Capacity

User quota exhaustion

Pro / Plus / Enterprise plans hit caps mid-month — the workers told to "use AI" can't. Discipline action: license & quota management. Usage forecasting per team. Burst allocation pool.

AI-new  ·  Compute / Capacity

Court-ordered restructuring

Musk v. OpenAI — $150B, structural remedies on the table. Discipline action: contract clauses for governance changes. A documented migration playbook — not a letter.

AI-new  ·  Compute / Capacity

Geopolitical exclusion of vendor

Anthropic / Department of War — 60+ days locked out of all DoD work, no published continuity plan from any affected agency. Discipline action: vendor political-risk monitoring; pre-built migration path.

AI-new  ·  Compute / Capacity

Skill market collapse

When everyone simultaneously needs the same fallback humans, you can't hire any. Discipline action: retain in-house manual capability now; the market won't be there later.

AI-new  ·  Organizational

Skill atrophy at scale

Layoffs absorbing AI work; juniors who never learned the manual process. Discipline action: cross-training cadence. Preserve at least one human path through every critical workflow.

AI-new  ·  Organizational

Cross-training gap (no manual fallback team)

The team who knew the workflow took severance. Discipline action: maintain a "manual run" capability for every AI-replaced workflow.

AI-new  ·  Organizational

Cultural over-reliance ("the computer said so")

Lending. Hiring. Customer service. Discipline action: decision-review thresholds. AI is a recommendation, not a verdict.

AI-new  ·  Organizational

Knowledge offshoring to vendor

Proprietary processes encoded in prompts and system instructions now live in someone else's stack. Discipline action: contract data-isolation terms. Audit what proprietary process now lives in vendor systems.

Full Register — All 34 Risks

Familiar  +  Amplified  +  AI-new
#RiskDiscipline ActionType
A. Operational availability
1Vendor outageRTO/RPO; multi-vendor failoverFamiliar
2Datacenter / regional infrastructure failureAvoid single-region AI dependencyFamiliar
3Power grid / fiber eventTreat AI like any regional cloud workloadFamiliar
4Acts of GodStandard DR site planningFamiliar
5Cyber / nation-state attack on vendorTabletop AI vendor breach scenarioAmplified
6Concurrent industry-wide outageAudit shared physical failure pointsAmplified
B. Commercial / contractual
7Pricing shock / mid-contract repricingProcurement caps; per-call cost ceilingsFamiliar
8Vendor financial / structural failureCounterparty diligence; multi-vendor as defaultFamiliar
9Vendor lock-in / switching costPortable prompt formats; vendor-swap drillsFamiliar
10Capacity / rate limiting (vendor-side)SLAs with throughput floors; burst-capacity vendor on retainerAmplified
11User quota exhaustion (customer-side)License & quota management; burst poolAI-new
12Forced API migrationVersion pinning; migration runway in contractsAmplified
13License / audit complianceSAM practice extended to AI seatsFamiliar
C. Legal / structural / regulatory
14Court-ordered restructuringMigration playbook, not a letter of assuranceAI-new
15Antitrust actionConcentration-risk monitoringFamiliar
16SanctionsGeographic counterparty reviewFamiliar
17Geopolitical exclusion of vendorVendor political-risk monitoring; pre-built migration pathAI-new
18Regulatory ban / compliance mandateCompliance pipeline; data-residency by jurisdictionFamiliar
19Export controls / chip embargoesSupply-chain audit on vendor compute locationAmplified
D. Integrity / output
20Silent model driftBehavioral baselines; output regression monitoringAI-new
21Model deprecation / sunsetTreat model upgrades like database upgradesAI-new
22Prompt injectionInput sanitization as OWASP-equivalent disciplineAI-new
23Training-data poisoningProvenance review on data feedsAI-new
24Audit / explainability failureDecision logging; explainability in procurementAI-new
25Bias / discrimination liabilityStandard model risk management; auditable trailAmplified
E. Organizational / human
26Skill atrophy at scaleCross-training cadence; preserve a human pathAI-new
27Talent flightDocumentation requirements; tribal knowledge in writingAmplified
28Cross-training gap"Manual run" capability for every AI-replaced workflowAI-new
29Cultural over-relianceDecision-review thresholds; AI as recommendation, not verdictAI-new
30Knowledge offshoring to vendorContract data-isolation; audit proprietary process leakageAI-new
F. Macro / systemic
31AI bubble / market correctionTreat AI vendor as high-growth-but-unprofitable counterpartyFamiliar
32Skill market collapse for AI fallbackRetain in-house manual capability nowAI-new
33Compute supply shockTrack concentration risk across AI vendorsAmplified
34Insurance market gapSelf-insure with continuity reserves until market maturesAmplified

Ask in your next IT review

  1. What's our RTO/RPO for AI workloads? If "we don't have one" — that's the conversation.
  2. When did we last fail over from our primary AI vendor? If never — schedule a drill.
  3. What's the documented manual procedure if AI is down for 24 hours? If none — write it.
  4. Who in procurement is tracking the Musk v. OpenAI docket? If nobody — assign it.
  5. How many of our critical AI dependencies share the same datacenter region, chip supplier, or holding company?
  6. What's our token / quota forecast per team — and what's our burst-capacity plan when an individual contributor hits their cap mid-month?
  7. Which of our proprietary processes now live as prompts inside a vendor system?

Sources

The Cast

Kyle

Host. Opinionated. Expect a history drop.

Kate

The correspondent. Tight, sourced, no spin.

Morgan

The heartbeat. "Well, why though?"

Full Transcript

Loads on open

Loading transcript...

Subscribe