Above the Noise

The ROI Number Isn't a Number

Tue, 21 Jul 2026 04:41:13 GMT

Somebody in your building is about to justify a nine-figure AI decision using a stat from a slide from a deck they never opened — and the stat is a misquote of an MIT paper that says something different. This week, the practitioner's read on how AI ROI numbers actually get made, why the ones on your slide are almost never measurement, and the three questions to ask in the meeting that turn any headline number back into something you can use.

This episode of Above the Noise — the unbiased AI brief for enterprise leaders:

News Brief: HuggingFace disclosed the first end-to-end autonomous-agent breach of an AI infrastructure provider — with a guardrail-asymmetry twist your CISO should know about · OpenAI shipped GPT-5.6 and ChatGPT Work, and quietly rewrote the enterprise ROI pitch on the same launch page · The $8B forward-deployed-engineering arms race (Microsoft Frontier Company, AWS, Anthropic-UST, Accenture Edge) landed in your quarter — and it's happening because in-house ROI isn't working the way the decks said it would.
Expose a Lie ⭐: "MIT says 95% of AI pilots fail to deliver ROI." Three breaks — it's not what MIT said, the methodology can't carry the weight (52 interviews + 153 conference-attendee surveys + a slide-deck report that was never peer-reviewed), and the same number is being used to slow down AI investment in one room and sell you a new platform in the next.
The Playbook: How to read an AI ROI number the next time one lands on the table. Three moves: what was actually measured (and in what unit), who paid for the study, and what population the number came from. Backed by receipts — METR (developers thought they were 20% faster with AI, were actually 19% slower), Brynjolfsson at the Fortune 500 call center (14% overall, ~34% for novices, ~0% for the most experienced), the GitHub Copilot "55.8% faster" vs. the Accenture enterprise follow-up (real gains, sixfold smaller). The Line for the Meeting: "Show me the sample, the measurement, and the sponsor — then I'll tell you what to do with the number."
The question to sit with: If the ROI number went to zero tomorrow, would you shut the program down? If the answer is no, the number was never the reason — say the real reason out loud.

No vendors. No hype. Just the signal.

👉 Follow Above the Noise wherever you listen — new episode every two weeks. Send it to one leader who has the "MIT 95%" bullet on slide three of their next deck, and leave a rating so more people find it. Between episodes, find Shaun on LinkedIn — come argue with him there. Especially on this one. The MIT number has fans.

🔗 Show Notes & Sources

Every stat in this episode is sourced. Check the work yourself:

HuggingFace autonomous-agent breach (July 16, 2026)

HuggingFace security incident disclosure, July 2026 (HuggingFace blog)

OpenAI GPT-5.6 + ChatGPT Work (July 9, 2026)

The $8B forward-deployed engineering arms race (July 2026)

Hosted on Acast. See acast.com/privacy for more information.

You Can't Govern What You Can't See

Mon, 22 Jun 2026 12:00:00 GMT

Something goes wrong in production, someone senior asks "walk me through exactly why the AI did that" — and the room goes quiet. That silence is the most expensive sound in enterprise AI right now. This week, the tested read on the two least exciting words in the whole stack — governance and observability — and why the part nobody demos is the part that ends careers.

This episode of Above the Noise — the unbiased AI brief for enterprise leaders:

News Brief: The EU AI Act high-risk deadline (Aug 2) and the Article 12 logging language nobody's ready for · SR 11-7 just got rescinded — and U.S. regulators handed AI governance back to you · the half-billion-dollar "AI governance" land-grab, and what its existence quietly admits.
Deep Dive: What governance and observability actually are, why almost no platform ships them (four structural reasons), and why you cannot bolt them on in "phase two."
Expose a Lie: "Our logs are our audit trail." Three breaks, each fatal — and the one question that turns the room honest.
The question to sit with: Could you prove, to a hostile outsider, exactly why one of your AI systems made one specific decision ninety days ago?

No vendors. No hype. Just the signal.

👉 Follow Above the Noise wherever you listen — new episode every two weeks. Share it with one leader who keeps saying "we log everything," and leave a rating so more people find it. Between episodes, find Shaun on LinkedIn — come argue with him there.

🔗 Show Notes & Sources

Every stat in this episode is sourced. Check the work yourself:

EU AI Act — high-risk deadline & logging

SR 11-7 rescission & financial-services AI governance

The governance/observability market

Production reality & observability

"Our logs are our audit trail" — the receipts

Topics covered: AI governance, AI observability, EU AI Act, high-risk AI, Article 12 logging, SR 11-7, OCC Bulletin 2026-13, model risk management, FINRA generative AI, AI audit trail, agentic AI, LLM observability, OpenTelemetry, regulated industries, enterprise AI strategy

Hosted on Acast. See acast.com/privacy for more information.

Build vs. Buy — The Most Expensive Identity Crisis in Enterprise AI

Tue, 09 Jun 2026 11:30:00 GMT

In a single 48-hour stretch, Microsoft announced seven in-house frontier AI models and their own GitHub division flipped Copilot to token billing — sending some users from $29 a month to $750. The universe basically scripted this episode of the build-versus-buy debate. The tested take from someone who builds enterprise AI for a living: this isn't one decision, it's three — and almost every failed AI program made one call instead of three.

This episode of Above the Noise — the unbiased AI brief for enterprise leaders:

News Brief: Microsoft Build 2026 — the MAI in-house model family, and why the largest AI buyer on Earth deciding to build at only the foundation layer tells you exactly where the build line is for the rest of us · GitHub Copilot moves to token billing (users seeing 10–50× cost jumps) — the "buy" side's hidden tax: you don't own the pricing model · Ignore pile — SAP Sapphire's "Autonomous Enterprise" and its €100M partner-payment fund.
Deep Dive — Build vs. Buy: The three-layer cut (model / platform / workflow) that fixes the conversation. MIT NANDA: vendor-led wins 67% vs. 22–33% for pure internal builds. Build cost: $300K–$1.5M+ upfront + 20–30% annual maintenance, forever. Plus the four taxes every spreadsheet undercounts — and why the right answer in 2026 is almost always hybrid.
Expose a Lie: "If we build it in-house, we own it." Steelmanned, then dismantled. When you build, you don't own the AI — you own the obligation to keep it alive. Maintenance bill, model-aging risk, governance liability, talent risk, all on you. Plus the data-moat sub-myth that drives half of build proposals.

🔗 Show Notes & Sources

Every stat in this episode is sourced. Check the work yourself:

News Brief — Microsoft Build 2026 (MAI models)

News Brief — GitHub Copilot token billing flip

News Brief — SAP Sapphire 2026 (ignore pile)

Deep Dive — Build vs Buy

Expose a Lie — "If we build it in-house, we own it"

Hosted on Acast. See acast.com/privacy for more information.

Agents, Hype, and What Actually Ships

Fri, 22 May 2026 20:10:31 GMT

Most AI advice comes from people who've never had to ship the thing. This first episode is the practitioner's read — from someone who builds enterprise AI for a living. The tested take on what's real, what's hype, and what to actually do. Case in point: the "autonomous agentic workforce" you're being sold is a roadmap slide, not a 2026 reality — and here are the receipts.

This episode of Above the Noise — the unbiased AI brief for enterprise leaders:

News Brief: Google's flagship launch was the cheaper model — the frontier race quietly became a price war · Every major vendor (IBM, ServiceNow/Accenture, Cognizant) pivoted to selling "agents in production + governance" — aka the cleanup crew for agents that didn't deliver.
Deep Dive — Agentic AI, hype vs. ships: Agents work in narrow lanes (one health system cut documentation time 42%). But Gartner says 40%+ of agentic projects get canceled by 2027, Deloitte says only 21% have mature governance, and MIT says 95% of GenAI pilots show no measurable return. What to actually do about it.
Expose a Lie: "AI agents will replace your team this year." Steelmanned, then dismantled with evidence and production experience — and why it persists.
The question to sit with: Of every AI initiative in your org — how many would survive "show me the line on the P&L where this shows up"?

No vendors. No hype. Just the signal.

👉 Follow Above the Noise wherever you listen — a new episode every two weeks. If it was useful, share it with one leader who's tired of guessing, and leave a rating so more people find it. Between episodes, find Shaun on LinkedIn.

🔗 Show Notes & Sources

Every stat in this episode is sourced. Check the work yourself:

News Brief

Deep Dive — Agentic AI: hype vs. ships

Expose a Lie — "AI agents will replace your team this year"

Background / Context

Topics covered: enterprise AI, agentic AI, AI agents, AI hype, AI ROI, AI governance, build vs buy, agent washing, Gartner, MIT, Deloitte, Google Gemini, IBM watsonx, CIO, enterprise leadership

Host: Shaun Gehring — engineering leader who builds enterprise AI and sells nothing. Opinions his own.

Hosted on Acast. See acast.com/privacy for more information.