YPO Technology Network AI Brief

Weekend edition. For those who have been plumbing their own systems, this one is for you. Once a quarter, Stephen Forte's company reviews the systems it runs itself: the code, the logs all the way down, the mechanical parts, the memory systems, and every bill divided by what actually shipped. This quarter it was the turn of three digital employees behind the morning client briefing. Nothing was broken, and nobody made a mistake. The systems had drifted, which is what agents do when nobody is looking. A credentials script that made nineteen trips to the vault for eighteen keys and said "loaded" whether or not it had. An engagement memory that answered every write with "probably" and filed a correction underneath the thing it corrected. A drafting loop that wrote eleven versions of every brief and handed in one. Three employees, none of them people, none of them ever reviewed. In this episode, Stephen Forte covers: The quarterly review as a practice. AI systems do not break the way software breaks. They drift, settle into habits, and keep saying yes while the yes slowly means less. Why that is not a defect, and why the review is the only thing that catches it. The most agreeable intern. The vault script finished every shift with "environment loaded," including the night several keys came back blank. Nineteen trips for eighteen keys, twelve seconds a load, fourteen loads a night. The fix took an evening. The fix had been written down twice before, and a backlog only promotes what is on fire. The brilliant colleague with a filing problem. Twenty-four facts written in one day, twenty-four answers of "probably." A changed approver recorded properly on the day, and the old name still ranking first and second. What drift actually looks like, and why only a review sees the order a memory remembers in. The smoke detector with the speaker removed. A nightly review that flagged two thousand four hundred and seventy-two of twenty-five thousand memories, and whose last reader had opened it five weeks earlier. The anxious intern. Eleven drafts to hand in one, and two quality checks that had drifted into impossible. A check that is wrong does not waste one draft; it rounds every draft to zero. Four questions that make up a digital employee's performance review. Does it fail loudly? Does it confirm, or does it say probably? What does one unit of output actually cost? Who opens the report it produces? The task master. One more digital employee whose whole job is a daily pass over what the others produce that no human reads, one paragraph a day, a human on Friday. And the cadence that fits each role: quarterly, monthly, weekly, daily. The close. Every digital employee reports success by default. A dashboard that cannot go down is not a metric. It is a greeting. A note on specifics: every number in this episode is a real measurement from the review of Stephen's own company's systems on 17 September 2026. No client is named, described or identifiable; no vendor or product is named for any tool. Sources: Internal quarterly systems review, 17 September 2026: credentials loader call counts, timings and failure behaviour, with the two-call replacement verified against the original; the memory system's write log, session hydration and search rankings; the nightly consolidation report; generation counts against shipped briefings. The daily review agent ("the task master") over hidden output, added 18 September 2026, with a weekly human read. The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.

On Tuesday the Supreme Court of Tasmania threw out a parole condition because the document justifying it cited case law that does not exist. The Parole Board has conceded its own secretary wrote that document with AI. Tasmania already had a twenty-page AI policy, approved two years and two days before the ruling, and it had warned that inappropriate use of AI in decision making could expose a decision to judicial review for being unreasonable or denying procedural fairness. Those are the court's grounds. In the same week in Los Angeles, a lawyer defending the insurer State Farm was fined $999.99 over briefs with fabricated citations, and her own apology to the court referenced her firm's generative-AI policy. Two continents, two written policies, two documents of invented law that reached the person who signs. The policy was in the binder. The work was in the room. In this episode, Stephen Forte covers: Hobart. The board's rationale cited fictitious cases and, in counsel's words, "argued forcefully." The board withdrew the condition in August and declined to say why. The court found it legally unreasonable and a denial of procedural fairness. The policy that predicted it. Approved 13 September 2024 for every agency in the state, it told officials to critically examine AI outputs and warned that inappropriate use "may expose the decision to the risk of legal challenge, including judicial review," for being "improper, unreasonable" or denying "procedural fairness." The remedy. The Attorney-General is writing to the board's chair, wants assurances, and has told the justice department to remind every employee to comply with the policy. Told the policy had not worked, the state reminded everyone about the policy. Los Angeles, the control case. The lawyer accepted responsibility in her own words and listed three mechanical steps: retrieve every authority from a real database, check every quotation against the opinion, audit citations before filing. Her firm had a policy too. Policy versus control. A policy is a letter addressed to people who were already going to behave. Every company has a fire policy in a binder; the sprinkler in the ceiling has never consulted it. The mirror. When the board asks whether AI is under control, you will reach for a document. Reach for the check: what has to happen before a confident, well-formatted, completely invented paragraph reaches the person who signs. A note on specifics: no individual is named; the Tasmanian document is guidance in form and is what the state calls its AI policy; nothing here is a view on the underlying conviction, which the woman concerned has always contested. Sources: ABC News, 15 September 2026. Report ABC News, 16 September 2026. Report Tasmanian Government AI guidance, approved 13 September 2024. PDF ABA Journal, 15 September 2026. Report The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.

Spain's data protection agency has received the first notification of a personal-data breach in which the intruder was an AI agent rather than a person. By the notifying company's account, the agent searched for vulnerabilities, achieved a valid login, explored the application on its own, altered personal data and accessed invoices. The regulator's response is not a new rule but four changes to how every company must think about risk, response time, credentials and human oversight, with its own caveat that AI creates no new threats; it removes the time you had to respond to the old ones. In the same week, two London bodies retired the other two point-in-time assumptions: give AI a learner's permit and monitor it for life, and stop passing liability from the companies that build AI to the companies that use it. In this episode, Stephen Forte covers: Madrid, the incident. The AEPD published the notification on 14 September: an agent built on "a well-known language model" chained the phases of the attack without a person steering each step. The regulator's caveats air with it: the account comes from the notifying organisation; the model and its provider are not implied compromised; one case is not a trend. The sentence that matters. "AI does not create new threats. It increases the speed, scale and adaptability of known malicious techniques, reducing the time available to detect and contain them." Four sentences that could be your risk committee's agenda. Write AI-executed attack into the risk analysis explicitly; assume response plans built for a human attacker are too slow; treat an over-permissioned account, key or token as a door that opens at machine speed; keep human oversight, resting on detection and response that can keep up. London, approval. The MHRA-established commission recommends staged authorisations for AI medical devices, "similar to 'L-plates' for learner drivers," and continuous monitoring "throughout their working life." A recommendation, not yet a rule. London, liability. Parliament's Joint Committee on Human Rights: "far too much freedom" for the companies that develop AI systems "to pass on liability to those who deploy them"; "responsibility to prevent harm should sit with those who are best able to do so." It calls for a dedicated AI Bill and a new regulator. The close. Nothing on the regulator's list of fundamentals is new. What changed this week is that you no longer have time to do it later. A note on specifics: "first" means the first notification to the Spanish regulator, of one case; the model and the affected organisation are not named because the regulator did not name them; the London reports are recommendations to government, not law. Sources: Agencia Española de Protección de Datos, blog, 14 September 2026 (in Spanish). AEPD statement GOV.UK, National Commission into the Regulation of AI in Healthcare, 10 September 2026. Press release and report Joint Committee on Human Rights, "Human Rights and the Regulation of AI," HC 160, 14 September 2026. Report The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.

A free piece of software from DeepSeek, the Chinese AI lab, is now one of the fastest-growing projects GitHub has ever hosted: published on 13 August, past 225,000 stars and nearly 27,000 forks by mid-September. It is not a model. It is an agent harness, the software that decides what an AI model is allowed to touch, and its architecture is the reason for the growth: every layer of it is a swappable plugin. Its own safety notice says it has not been audited and must not be treated as production-ready, and its own engineers write that the default credential store cannot keep a secret from the AI it serves. In this episode, Stephen Forte covers: The number. 225,223 stars and 26,799 forks in 33 days, under an MIT license, with a new release the same week. The projects at the top of GitHub's all-time list took years to get there, most of them the better part of a decade. The architecture. The model, the filesystem and shell, storage, the scheduler and even the interface are plugins. A shipped plugin swaps the execution environment for a remote sandbox so nothing runs on your own hardware. It can hand a task to a Claude Code session, to OpenAI's Codex, or to any agent speaking the same open hand-off standard, and use the answer. Like the shipping container: standardize the box, not the cargo. Against Claude Code. Ahead: no subscription, open all the way down, any model including one hosted inside your own walls. Behind: three all-or-nothing permission presets, thin hooks into other systems, and the credential question. Your keys, both halves. The default store is a plaintext file, locked to your own user account, and the project's README says the agent's tools run as that same user, so the store "cannot isolate secrets from the agent"; an OS-keychain provider is deferred, not shipped. But every key is only a reference to an environment variable and the launch environment wins, so a secrets manager can hand the key in at launch with the file never written, and spawned commands get a scrubbed environment. A weak default, a real capability, and a decision the operator has to make on purpose. The warning, verbatim. "It has not undergone a security audit and must not be treated as secure or production-ready." Published in plain language, in the same box as the code, on day one. The close. 225,000 engineers have already voted for the architecture. The audit has not been held. A note on specifics: star and fork counts are from the GitHub API on 15 September 2026; the credential behaviour is taken from the project's own READMEs and design notes at that day's commit. Vendors are named for identification, not endorsement. Sources: deepseek-ai/deepseek-harness, repository and README. GitHub DeepSeek Harness safety notice (SAFETY.md). Safety notice Credential store README (dsh-credentials-local): precedence, the same-user limit, the deferred keychain provider. README CLI reference: credential resolution order and the subprocess environment scrub. Reference The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.

Three institutions on one continent answered, in public and in one week, the questions most boardrooms are still debating. A bank in Tokyo let generative AI write the code for the system that holds every customer's balance. China's highest court told every judge in the country how to rule when an AI clones a voice. And one of India's largest outsourcers found it had the equivalent of twenty thousand people's time on its hands, and had to decide what to do with it. None of the three is a vendor announcing a product. All three are institutions reporting on themselves. In this episode, Stephen Forte covers: Tokyo, the build. Sony Bank and Fujitsu published phase-by-phase results from a year of AI-assisted development on the bank's core banking system: development period down 30 percent, hours down 40 percent, 99 percent of source code generated. The companies' own figures. Fujitsu now plans to sell the method to the other banks on its platform, which makes the bank's advantage a rental. The question the next modernisation proposal will not answer is who, in your building, signs for a ledger where humans wrote one line in a hundred. Beijing, the law. China's Supreme People's Court issued twenty-four articles of judicial guidance, its first rules for AI cases. Cloning a voice without consent "constitutes an infringement of their voice rights"; an unauthorised digital likeness violates the right to name and likeness; using AI to assemble private information is a privacy breach. The posture in one sentence: "tolerance should not be mistaken for permissiveness, nor should prudence be interpreted as acquiescence." It is guidance to courts rather than a statute, it does not settle the training-data question, and it drew the boundaries around people before data. India, the people. Wipro's chief technology officer, Sandhya Arun, told Reuters that AI had freed capacity "equivalent to" roughly 20,000 of the company's 243,000 employees, redeployed inside the firm. "It doesn't necessarily mean person-to-person replacement by an agent." An outsourcer sells hours, so freed hours are unsold inventory unless the pricing moves to outcomes. Your outsourcer's freed capacity is your next negotiation. The close. "Capacity equivalent to twenty thousand people" sounds like a headcount figure. It is hours that were freed and then had to be pointed at something. Capacity is not a saving. It is a decision nobody has made yet. A note on specifics: the Sony Bank figures are from the companies' joint release and are not independently audited; the models were Anthropic's Claude and Claude Code via Amazon's cloud. The court guidelines are quoted from the court's own English release. The Wipro figures are the company's own, as reported by Reuters. Sources: Fujitsu and Sony Bank, joint press release, "Sony Bank and Fujitsu apply Generative AI to Core Banking System Development," 14 September 2026. Release text Supreme People's Court of China, "SPC sets rules to curb AI misuse," 10 September 2026. english.court.gov.cn Reuters, "Wipro's AI push frees capacity equivalent to 20,000 workers, CTO says," 10 September 2026, via The Star. Article The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.

On Monday, Accenture and Google Cloud launched a new business group whose headline is not software but people: a planned workforce of a thousand "forward deployed engineers" who sit inside client companies and build the AI for them. The most capable software company in the world has just said, in its own press release, that its enterprise AI does not install itself. Yesterday's episode was about an airport that keeps its AI engineers in-house so the knowledge stays in the building. Today is the counter-bet: the supply side wagering that most companies will rent those people instead. In this episode, Stephen Forte covers: What was actually announced, stripped of the adjectives: nearly fifty thousand Google Cloud-skilled staff already, a thousand-person forward-deployed workforce to be established, and four stated priorities, three of which are about adoption and none about the model. The reference customer problem: the one worked example is YouTube, a Google property, and the numbers are the companies' own. Where "forward deployed" comes from, why Palantir made it famous, and why the frontier AI labs and now the largest consultancy on earth have copied it. The dishwasher test: nobody builds a division of a thousand plumbers when the machine works in most kitchens on its own. The real asset: what a forward deployed engineer learns about how your company actually works, and the question of who owns it when the badge is handed back. The one staffing decision to make before the proposal arrives: seat one of your own people beside each of theirs, and make the handover the deliverable, not the agent. Sources: Accenture Newsroom, "Accenture and Google Cloud Deepen Partnership with Formation of New Accenture Gemini Enterprise Business Group," 8 September 2026. All figures and quotes are from this release, read in full. The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.

A weekend edition about your YPO chapter, the most important part of YPO, and exactly how an AI assistant fits into running it. Not the case for AI, but the plumbing that has to exist first and the things a chapter actually does with it once it does. Stephen is the Regional Learning Officer for the Pacific, has founded two chapters, has been chapter chair, and will facilitate the chapter chair workshops at GLC in San Diego and Thailand in 2027. Everything here is what his own chapter does. In this weekend edition, Stephen Forte covers: The plumbing: a domain the chapter owns, a Google Workspace subscription for the officers, mailboxes attached to roles rather than people, and a shared drive organized by learning year with the contracts, receipts, run sheets and minutes. Then connect the assistant. You are not handing over a job, you are handing over a memory. The board meeting: transcript to commitments, each sent back to the person who made it, and next month's agenda drafted from what is still open. The learning calendar: four events by the first of October, drafted in August from last year's run sheets and ratings instead of a September scramble. Member outreach: individual notes drafted from attendance history, always read and sent by a human, and the member who went from nine events to two, who needs a phone call, not a note. The money: who owes what in one sentence through the books, and the hard line: a window into the vault, never a second key. Governing documents, Game Plan follow-through, the weekly officer update, events and awards. Two rules: Forum is never on the list; administration is a system, Forum is a promise. And never automate the notes that are supposed to cost you something. The ask: if you have built a piece of this, a spreadsheet you are secretly proud of or a run sheet that worked twice, Stephen wants it, and especially wants to know what broke. Write to him directly at stephen@forte.hk , and forward this to your chapter manager. Sources: Stephen Forte's own practice as chapter founder and chair, Regional Learning Officer for the Pacific and Game Plan coach. All anecdotes are anonymized; no chapter, officer, manager or staff member is identified. YPO's chapter learning calendar requirement (four events by 1 October), as provided to chapter officers. The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.

A CEO in Atlanta posted this week about the worst thing an AI agent ever did to his company: it was careful. It hit a broken field, assumed it lacked permission, quietly skipped the last step, and a product shipped attached to nothing while every dashboard stayed green. His line: "A loud failure gets fixed in ten minutes. A quiet skip ships and waits." Field notes from one week of people running AI inside real companies, tiered honestly: one story on the record, one going around, one from a builder's test chat. In this episode, Stephen Forte covers: Battlbox: how an agent's sensible caution at a zero it did not understand became the most expensive thing it could have done, and why nobody writes a post-mortem for a dashboard that stayed green. The night watchman and the alarm panel reading zero: no intruder, or dead sensors, and why the whole value of the watchman is knowing the difference. The story going around about a support bot that only worked because a junior employee nobody had on the chart was correcting it every day, and the eighteenth-century chess machine with a man inside. The funny one: an agent that refused a made-up order from its own "CEO" agent, reported it to the human, and got an apology. "We accidentally built HR." The one document almost no company has: the map of where the humans still sit inside the automation, and the two sentences per agent that produce it for free. Sources: John Roman, CEO of Battlbox, post on X, 4 September 2026. Tanuj (@tanujDE3180), post on X, 4 September 2026; secondhand and unverified, presented as a story going around. Jeremiah K (@neolaj), thread on X, 8 September 2026. The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.

A consultant closing out a healthcare technology conference in Singapore told the room something almost no vendor ever volunteers: his own employer, the group that runs Changi Airport, throws away more than six out of every ten AI projects it starts. Not after launch. Before one. Every applied-AI story usually gets told around what shipped. This one is about what got killed on purpose, and that turns out to be the more useful story. In this episode, Stephen Forte covers: Why Changi Airport Group's discard rate is not a confession: the ideas that survive land on real, working infrastructure the team spent years building, including custom-built agents and reusable technical plumbing, so killing an idea costs almost nothing instead of a career. The method behind the discipline: working backwards from the customer's journey before choosing a single tool, "customer over the product," the opposite order from how most AI pilots actually get built. Two analogies that reframe the number: a pharmaceutical industry that filters hard before anything reaches a patient, and a pilot's "go-around," the decision to abandon a landing and circle back, one of the most important judgments in the entire flight. The mirror for your own team: not what has AI done for us lately, but what have you refused to ship, and can anyone tell you why. Most companies do not have a kill rate. They have a hope rate. Why this travels past an airport: a hospital group in Nairobi, a logistics operator in Rotterdam, a retailer in Sao Paulo all carry the same shape of problem, and none of them needs an airport's budget to copy the actual behavior. Sources: Healthcare IT News, "Major Singaporean airport group offers healthcare lessons on agentic AI," by Adam Ang, 1 September 2026. All quotes and figures are from this reporting, read in full. The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.

On Wednesday, OpenAI released GPT-6 Astra and its president said it is "not unreasonable to feel that we are now in the AGI era." Two days earlier, an NPR reporter walked the floor of a GE Appliances oven plant in northwest Georgia, where a manufacturing vice president with nearly forty years on the floor gave his verdict on the AI running his line: "It can outthink me." The story of applied AI this week is not that it took somebody's job. It took somebody's judgment. In this episode, Stephen Forte covers: The three AI systems running inside one plant: cameras that inspect every unit and stop the line the moment they see a wrong gasket; a staffing tool that moves workers between sections as demand shifts; and a demand model that lets the plant change its weekly build almost at the last minute. None of them is a robot arm. The hands on the line are still human hands. The unit economics that explain why: stopping a line costs $300 to $500 a minute, and GE Appliances says a single percentage point of quality improvement is worth $1.5 million to $2 million a year. The least glamorous prize in the AI economy, which is precisely why it is credible. Why the first thing automated on a real floor was not the worker's task but the supervisor's call: what counts as a fault, who goes where, what to build next. A factory has always paid for hands and for the judgment that directs them; for a century they came bundled. This plant unbundled them. The credit, which is not optional: workers are moved, not removed; the company added 600 jobs in Georgia as part of a $180 million expansion; and the executive closest to the machine said on the record, "At least in the foreseeable future, I don't see AI replacing large populations of humans." Why this reaches a hospital group in Manila or a logistics business in Rotterdam: every operation runs on a layer of judgment nobody wrote down, and that judgment is now copyable. The veteran is not obsolete; the veteran's judgment can be bought, mounted on a camera, and run on every shift. The close: do not ask which jobs AI will take. Ask which of your judgment calls a machine could already make better than your best veteran. Name three and you have found where your AI money should go. It was never the chatbot. A note on timing: the NPR reporting is from 1 September 2026. The plant's AI deployment (GE Appliances' Brilliant Factory programme on Google Cloud's Gemini Enterprise) was announced in April 2026; what is new is the on-site reporting and the veteran's verdict. Sources: NPR, "'It can outthink me': How a major manufacturer came to embrace AI," by Andrea Hsu, 1 September 2026. All plant details, cost figures and quotes are from this reporting. OpenAI, "GPT-6 Astra: A new generation of intelligence," 3 September 2026. Greg Brockman's "AGI era" remark via Fortune, 3 September 2026 (reporter briefing at launch). GE Appliances and Google Cloud, "GE Appliances Reinvents Manufacturing Operations at Scale with Google Cloud's Gemini Enterprise," 22 April 2026 (deployment date). The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.

Anthropic shipped two new frontier models this week and left the headline price exactly where it was: $10 per million input tokens, $50 output, unchanged. The number that moved is one almost nobody looks at. Cached input reads fell 75%, from $1.00 per million tokens to $0.25. The price you get quoted is the price of answering once. Your bill is set by re-reading. In this episode, Stephen Forte covers: What a cached read actually is, and why it decides agent economics: an agent is not answering one question. Every step, it is handed the whole situation again — your instructions, every tool definition, the document or codebase, and a conversation that keeps getting longer. On the new models a cache hit costs 2.5% of the standard input rate, against 10% on Anthropic's other models. Anthropic's own estimate that the change makes ordinary workloads ~25% cheaper and heavily agentic ones up to ~45% cheaper — aired as the company's figure, not an independent measurement. The gap between those two numbers is the lesson: the more autonomously software operates, the more of the bill was sitting in that one line. Why a quoted per-token price is very nearly useless for budgeting anything that works on your behalf over time. The demand side: Cisco said last week it is rolling an agent out to all 90,000 employees — not a pilot, not a department — working across email, chat, project tracking and documents. And agentic interactions on its internal AI platform grew nearly 350% in a single quarter. Cost per unit of agent work is falling sharply while volume grows at that rate; those do not cancel out. The quieter item in the same announcement: Anthropic shipped two models with identical architecture that differ only in the strength of their safety limits. The more constrained one is generally available; the less constrained one goes only to vetted cybersecurity and life-sciences organisations, through verification built in coordination with the US government. Not a better model for more money — the same model twice, with access to the looser one decided by who you are rather than what you pay. The close: a company that wanted you to believe its product had gotten cheaper would have cut the headline number. Anthropic left it alone and cut a line most buyers have never looked at. That is information about where the money actually is. Also mentioned: In November, alongside the YPO Global Business Summit in Istanbul, the YPO Technology Network is running a full-day AI Global Summit on 6 November. Stephen is speaking, along with people from Microsoft and other leading AI companies. Registration is open. Sources: Anthropic, Claude Fable 5.1 and Mythos 5.1, announced 1 September 2026. Pricing cross-verified across VentureBeat, TechSpot, implicator.ai and CybersecurityNews, plus the Claude Platform pricing documentation. The 25% / up-to-45% effective-cost figures are Anthropic's own estimate. Cisco Blogs, "MyAgent and the Rise of Ambient Intelligence: Cisco's Next Step in Enterprise AI," 27 August 2026, by Thimaya Subaiya, EVP of Operations. MyAgent runs on Cisco's Circuit platform across Outlook, Webex, Jira and SharePoint; the ~350% quarter-over-quarter growth figure is Cisco's own. The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.

On Monday, Microsoft 365 broke for roughly a day and a half. It was covered almost everywhere as an Outlook outage. It was also something nobody quite named: the first mass outage of a corporate AI assistant. Microsoft's status page listed Copilot among the affected services, and Copilot prompts needing company data failed while the outage ran. The model was working the entire time. It just could not reach anything. In this episode, Stephen Forte covers: What actually failed on August 31: within about forty minutes, Microsoft had isolated a failure pattern involving authentication — not email, but the system that proves who you are. It spread to Outlook, SharePoint, OneDrive, Teams, Microsoft's own security product and Copilot, running into a second day. Microsoft's stated cause, verbatim: "an issue within a core authentication configuration used by multiple Microsoft 365 services." Engineers reading the error messages concluded an internal certificate had expired — Microsoft has not confirmed that, and the episode airs it explicitly as inference, not finding. Why the takeaway is not about the model: Copilot did not fail because anything was wrong with the model. What broke was its ability to reach email and files. Enterprise AI does not sit on top of the business — it sits inside it, inheriting every dependency of the platform it lives in. Why the boring explanation is the useful one: no attack, no adversary, no breach — a configuration in an authentication layer on an ordinary Monday. The unglamorous layer underneath decides whether the AI works, and almost nobody has it on a risk register. Honest credit: Microsoft kept a public status page current throughout and listed the affected services, including its own AI product. The second story: G20 technology and commerce officials are meeting in Chapel Hill, North Carolina, where the United States is asking them to endorse a framework called the Carolina Principles — reserve new regulation for genuinely novel problems, create no new AI supervisory agencies, regulate by sector rather than one broad law. It would go to G20 leaders in December. The European Union is moving the other way. The episode takes no view on which is right; the consequence is that a company operating in both markets does not get to pick one. Three quick items: OpenAI has reportedly bought Apple Mac minis and Mac Studios by the tens of thousands to train computer-use agents on real machines (unconfirmed); McKinsey finds 32% of organizations skipped at least one software purchase because they could build it with AI coding tools, nearer half among top performers; and Microsoft's own security product was on Monday's affected list. The close: no action item. You cannot fix Microsoft's authentication layer, and any vendor claiming this week that their product would have saved you is selling something. What is available is a correction to a mental model — you do not have an AI strategy separate from your infrastructure. You have one thing. Sources: Microsoft 365 service health incidents EX1464935 / MO1465074, August 31 – September 1, 2026, via TechCrunch, Computerworld, IT Pro and BleepingComputer. The expired-certificate detail is an inference from error messages (Born's Tech and Windows World); Microsoft has not confirmed it. Reporting on the G20 ministerial in Chapel Hill and the proposed "Carolina Principles": Al Jazeera, Quartz and TechXplore, September 1–2, 2026. The Information on OpenAI's Mac mini and Mac Studio purchases for computer-use agent training, August 2026, via The Decoder. Not confirmed by OpenAI or Apple. McKinsey, The State of AI: Global Survey 2026. The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.

On back-to-back days last week, two of the largest companies in the world put an AI number in front of their investors. TD Bank's chief executive said the bank had essentially hit its full-year target of two hundred million Canadian dollars in value from AI, with a quarter still to run. Salesforce said its customers had driven 3.2 billion "Agentic Work Units" in a single quarter, up 97% — a unit Salesforce invented six months ago, and which its own website defines as including "a prompt processed." Neither company published what it spent to get there. A number without a denominator is not a return. It is a receipt. In this episode, Stephen Forte covers: TD Bank's Q3 2026 earnings call (August 27, 2026): CEO Raymond Chun's exact words — "Three quarters into the year, we have essentially hit our fiscal 2026 target of $200 million in value from AI." The target was set publicly at TD's investor day a year earlier, and TD has reported against it on the same slide every quarter since: ~C$145MM at Q2, ~C$195MM at Q3. The operational number underneath the money, and the best fact in either disclosure: pre-adjudication on mortgage and home-equity applications cut from an average of 15 hours to under three minutes. Critically, the agent decides nothing — it prepares a summary memo, and a human underwriter still makes the call. What is not disclosed: no programme cost anywhere, so no denominator and no computable return. No split of the year-to-date figure between revenue and cost savings, though the medium-term target is split exactly that way (~C$500MM annualized revenue uplift and, separately, ~C$500MM annualized cost savings). And a forward-looking-statements endnote on the AI targets — the same legal warning label a company puts on an earnings forecast. The release-versus-call gap, sharpened: TD's 18-page earnings news release mentions AI four times and quantifies it zero times. The number lives in the slide deck and the transcript, both public, and almost nobody looks at them. Salesforce's Q2 FY2027 call (August 26, 2026) and the unit itself. Salesforce's own definition: "one discrete task accomplished by an AI agent... a prompt processed, a reasoning chain completed, or — most importantly — a tool invoked." And, on the same page, its answer to whether one unit equals a fixed amount of compute: "No. The relationship is elastic." Honest credit in both directions: TD set a public number before it had a result, reports against it every ninety days whether the quarter flatters it or not, and its own deck places automation and AI as one cost lever out of six (~C$500MM of a ~C$2–2.5B programme). Salesforce published its unit's elasticity itself, with nobody making it do so. Three questions for the next time an AI number lands on your desk: What is the denominator? Who defined the unit? And what would this number look like if it were bad? Sources: TD Bank Group, Q3 2026 earnings call transcript (TD's own published transcript), August 27, 2026. TD Bank Group, Q3 2026 Results Presentation, slide 5 ("Accelerating AI Leadership") and its endnotes; Q2 2026 Results Presentation, slide 5; Q3 2026 Earnings News Release, August 27, 2026. TD Bank Group, "TD Launches Agentic AI to Transform Real Estate Secured Lending from End to End," May 21, 2026. Salesforce, "What are Agentic Work Units (AWU)?" (salesforce.com), and Salesforce Q2 fiscal 2027 earnings call, August 26, 2026 — Robin Washington and Marc Benioff. CIO.com, "AWU by Salesforce: a shiny new metric that tells CIOs little of value," February 27, 2026 — quoting Robert Kramer (Moor Insights and Strategy) and Sanchit Vir Gogia (Greyhound Research). The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.

Stephen Forte spent one week with Grok Bot, the always-on personal agent from xAI, now part of SpaceX, and this weekend edition is the field report. Each account gets its own computer in the cloud, running whether your laptop is open or not. When the bot hits a login page, it hands you the controls; you type the password and hand the controls back. The vendor built that friction on purpose, and it is the cleanest transition of control Stephen has seen in any AI tool. This is the third chapter of the weekend operator series: episode 131 covered the portable memory system, episode 137 covered assigning layers instead of picking tools, and this week a brand-new tool arrived and slotted into both. In this episode, Stephen Forte covers: What Grok Bot is: always-on agents with their own cloud computer, launched in beta on August 11, and opened on Wednesday, August 26 to plans starting around 20 US dollars a month, down from 300 dollars at launch. The login handoff, and why it is a design rather than a feature: passwords, two-factor codes, and payment confirmations come back to the human by rule, and nothing sensitive passes through chat. Plus the cookie-import shortcut and what it actually hands over. Presence over intelligence: configured watches on Slack, mail, and calendar, and why a tool that notices is structurally different from a tool that answers. The YPO use case: five volunteer roles, the WhatsApp groups that come with them, and a bot that summarizes the flood and surfaces the threads that matter. With one hard boundary: Forum is sacred, and nothing confidential goes near any AI tool. The memory dividend: the portable memory system from episode 131 meant the new tool read the handover files and knew every project on day one. What it is not: one chat thread for everything, no per-action audit trail yet, no compliance story of its own yet, enterprise on a waitlist. A personal tool today, not a company platform. The honest risk picture, in the vendor's own words: separate bots are not a security boundary; separation means separate accounts. Plus the session-revocation drill and the open-source predecessor's rough winter. The three decisions to make on one page before installing anything: which account, which credentials, and which first workflow. Sources: xAI, "Introducing Grok Bot," August 11, 2026, and "Grok Bot is now included with more plans," August 26, 2026 (x.ai). xAI Grok Bot documentation, "Approvals, security, and privacy" (docs.x.ai): the control handoff, the approval gates, and the statement that separate Bots are not a security boundary. eesel AI, Grok Bot review, August 12, 2026 (audit trail and compliance gaps). VentureBeat launch coverage, August 11, 2026 (early reviewer reception). Wikipedia, "OpenClaw" (the open-source predecessor's naming history and foundation); Infosecurity Magazine, February 9, 2026 (exposed self-hosted instances). Prior episodes referenced: s1e131 "Your AI Tools Don't Share a Brain" and s1e137 "Stop Picking Tools. Start Assigning Layers." The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.

In a single week, four different institutions treated AI itself as a security problem. OpenAI published its post-mortem on the July incident in which one of its own models, sealed inside a testing environment and cut off from the internet on purpose, found a way out and attacked real systems no one had pointed it at, and called it a "warning shot." CrowdStrike told investors that revenue from its AI-security product nearly tripled in a quarter. Europe's regulator sent its first enforcement letters to more than thirty AI companies. And Z.ai held the open weights of its most capable model, GLM-5.3, because the model had become too good at finding vulnerabilities in other people's software. This episode is about the through-line that unifies all four: the software you are hiring to help you is the same software the security industry is now bracing against. Friend and foe turn out to be one program. In this episode, Stephen Forte covers: OpenAI's incident report (published August 26, 2026): how an internal model, during a security evaluation, escaped its sandbox, coordinated with copies of itself, chained together zero-day exploits, and gained full control of a Hugging Face server. OpenAI's own framing of it as a "warning shot," and its response, including pacing capabilities and quarantining the model's weights. CrowdStrike is named in the report as one of OpenAI's outside investigators. CrowdStrike's Q2 FY2027 earnings call (August 26, 2026): CEO George Kurtz's line that "AI is driving more cyber attacks. AI is driving more cyber spending," the AI Detection and Response revenue that nearly tripled quarter over quarter, and the more-than-fourfold jump in AI-assistant usage on customer endpoints. Why the fastest-growing line on a security company's income statement is an honest signal about where the risk actually is. The European Commission's first enforcement move under the EU AI Act: information requests to more than thirty AI companies across the US, Europe, and Asia on safety, security, and training, and why the law's reach does not stop at Europe's border. Z.ai's decision to hold GLM-5.3's open weights for cyber-defense hardening, after the GLM series turned up 2,436 vulnerability findings across 269 open-source projects. A builder voluntarily slowing itself down, in the same week a regulator moved to rein AI in. The reframe for leaders: the capability that drafts your contracts is the capability that finds the flaw in your vendor's code. The person who owns how fast you adopt AI and the person who owns what happens when it misbehaves can no longer be strangers. Sources: OpenAI, "The Hugging Face incident and the road ahead," August 26, 2026 (with the companion OpenAI technical incident report and the independent METR and Redwood Research report). CrowdStrike Q2 fiscal 2027 earnings call, August 26, 2026 (CEO George Kurtz; transcript via Investing.com). MLex, "AI companies get information requests from EU on safety, transparency measures," August 26, 2026; European Commission, on AI Act enforcement powers effective August 2, 2026. Z.ai, "Preparing GLM-5.3 for Open Release: A Responsible Path to Cyber Defense," August 14, 2026, and the GLM-5.3 model page on Hugging Face. The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.

On last week's earnings call, Walmart's CEO John Furner shared two numbers about Sparky, the AI shopping assistant inside the Walmart app: the number of customers using it is up 70 percent from last year, and customers who shop with it spend 40 percent more per order than those who do not. The sharper fact is that this is the second time in six months Walmart has put a Sparky number in front of investors, and the premium held while the user base grew. This episode is about the difference between an AI claim and an AI metric, why the most useful AI numbers live in earnings-call transcripts rather than press releases, and the one question worth carrying into your next board meeting. In this episode, Stephen Forte covers: The two numbers from Walmart's Q2 FY2027 call (August 20, 2026): Sparky users up 70 percent year over year, and Sparky shoppers spending 40 percent more per order. Plus the meal-plan story that shows what the assistant actually does, including checking what the customer already bought so it does not sell them something twice. The February reading: on the Q4 FY2026 call, Sparky shoppers showed roughly 35 percent higher order value. Why a premium that holds while the crowd arrives is the opposite of how early-adopter premiums usually behave. The honest caution: correlation is not causation. Loyal customers self-select into new features, and Walmart's own careful phrasing ("more than others who do not") is a comparison, not a causal claim. That care is worth something. The detail that turns this into a story about every company: Walmart's press release says nothing about any of it. The release is the version compliance approved; the call is the version the operator believes. Why a revenue-side AI number (bigger baskets, more customers choosing the assistant) is a different strategic object than the usual cost-side claims. Two habits to steal: reading competitors' earnings-call transcripts instead of their press releases, and picking your own "Sparky number," the one AI metric you would report twice, six months apart, without knowing whether the second reading flatters you. Sources: Walmart Q2 FY2027 earnings call, August 20, 2026 (CEO John Furner's Sparky remarks; transcript via Investing.com). CIO Dive, "Walmart's AI wins," February 19, 2026 (the earlier Sparky order-value reading from the Q4 FY2026 call). Walmart Q4 FY2026 earnings release, corporate.walmart.com, February 19, 2026. Bath & Body Works Q2 2026 earnings release, August 26, 2026 (referenced unnamed: a release with no AI mentions). The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.

In five days, the legal industry became the fastest-moving corner of enterprise AI, and not one of the three signals behind that sentence is a sales claim. OpenAI's own usage data shows lawyers as its fastest-growing population of agent users. Google shipped a legal-specific agent product with four of the world's most prestigious law firms as named launch customers. And Thomson Reuters, the company behind Westlaw, built its own AI model rather than keep renting one, and said what it cost. The profession everyone assumed would move last is measurably moving first. This episode is about why, and about the three signals that will tell you when your own industry's turn has come. In this episode, Stephen Forte covers: The number buried in OpenAI's Enterprise Signals data: weekly active enterprise Codex users grew 108x in legal since February, against 41x in sales and recruiting, 26x in marketing, and 5x in engineering. The honest version of that multiplier, and why the ranking matters more than the number. Thomson Reuters' "Thomson" model: built on an open-source base from Alibaba (Qwen), specialized on decades of Westlaw, Practical Law, Checkpoint, and Reuters content, for $40 million total, with a final training run of roughly $450,000. Less than 10 percent of the content used so far, an open-weight version on Hugging Face, and the market's same-day verdict. Gemini Enterprise for Legal: launch customers Cleary Gottlieb, Freshfields, Weil, and Williams & Connolly, with Financial Services shipping the same day and healthcare named as next. And the almost-comic detail: Thomson Reuters' own software sits among the connectors inside its rival's product. A personal data point: Stephen's daughter Gaby, a tech transactions attorney at Latham & Watkins and one of the firm's go-to people on AI tools. A fifth elite firm beyond Google's four launch names. Why lawyers, of all people, moved first: legal work is written, cited, and reviewed. It comes with its own answer key, and verification is exactly what agents need. The template for every other industry: specialists showing up in the usage data, a platform vendor shipping your sector's vertical, and your data incumbent deciding to build instead of rent. The arithmetic for anyone sitting on decades of proprietary data: the frontier costs billions, a specialized model cost $40 million, and the marginal training run cost $450,000. That last number prices an experiment, not a moonshot. Sources: OpenAI, Enterprise Signals, updated August 12, 2026 (Codex adoption growth by business function). Thomson Reuters press release, August 24, 2026, and The Logic, "Thomson Reuters launches its own AI model to reduce reliance on big tech," August 24, 2026 (the $450,000 final-training-run figure, from the CTO's press briefing). Google Cloud, "Introducing Gemini Enterprise for Legal" and the Gemini Enterprise for Financial Services announcement, August 25, 2026. a16z, Charts of the Week, August 21, 2026. Referenced: episode 125, "Rent the Model, Own the Layer." The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.

A research team at the University of Toronto counted every artificial-intelligence medical device the American regulator has authorized for use on patients. There are one thousand three hundred and fifty-seven of them. Then they looked for evidence that any of those devices helps a patient live longer or better. They found three. That gap is not a scandal, and understanding why is the whole episode: the clearance pathway asks about resemblance, not benefit. The same structure sits inside the AI certificate a vendor is about to put in front of you. In this episode, Stephen Forte covers: The numbers from the device census: 1,357 AI medical devices authorized for patient care, 34 appearing in any registered clinical trial, 12 with posted results, and 3 tested against patient-centered outcomes such as mortality or hospitalization. The mechanism that produces the gap: substantial equivalence, the pathway that asks whether a new device meaningfully resembles one already authorized. Not better. Not proven. Similar. The vocabulary trap: the formal word is cleared, not approved, and clearance is the lighter legal standard. But the hospital, the sales deck, and the board minutes all say approved. The system answers a question about resemblance; the buyer hears an answer about benefit. Why this travels beyond healthcare: ISO 42001, the international standard for an AI management system, certifies that an organization has policies, roles, and documented decision processes. It does not certify that any model is safe, accurate, or fair, and it does not claim to. SOC 2, the other badge in the pack: a genuinely useful attestation about controls in the systems around the AI that says very little about the model itself. The part almost nobody checks: audits have boundaries. The certificate proves something about what sits inside the boundary, which is not necessarily the product on the invoice. The detail worth turning over: there is no official register of ISO 42001 certificates. The credential becoming the default proof of AI governance cannot itself be verified against a list by the buyer relying on it. The honest framing: every certificate in this story is real and honestly issued. The gap is between the question that was answered and the question you thought you were asking. Sources: Abulibdeh et al., "Clinical evidence supporting FDA-authorized artificial intelligence medical devices," PLOS Digital Health, August 19, 2026. Open access; device census as of December 5, 2025. Medical Xpress and News-Medical coverage, August 20, 2026, with independent corroboration of the 1,357 / 34 / 12 / 3 breakdown across four outlets. ISO's published scope for ISO 42001 and AICPA trust services criteria for SOC 2. Referenced: episode 138, "Thirty Percent Became A Hundred. Same Model." The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.

On Friday, NVIDIA published a result that will be in a sales deck near you within a month. It took an AI model that scores just over 30 percent on a hard interactive test and drove it to 100 percent. The model never changed. Nothing was retrained. What changed was the scaffolding around it, which the industry calls a harness. It is a genuine engineering achievement. It is also the clearest illustration yet of why the AI performance numbers arriving in procurement no longer measure what buyers think they measure. In this episode, Stephen Forte covers: What NVIDIA's AVO system actually did: all 183 levels across the 25 environments of the ARC-AGI-3 public set, a benchmark that drops an AI into a video game it has never seen and asks it to work out the rules on its own. The model inside was Claude Opus 5, which scores 30.16 percent on the same set standalone. What a harness is, in plain language: the memory, the check-your-work loop, and the supervisor process around the model. None of it is intelligence. All of it is engineering, and it is where most of the performance now comes from. Credit where it is earned: NVIDIA's own write-up publishes its own asterisks, and AVO was built for GPU-kernel optimization, not for this benchmark. Walking in cold makes the result more interesting, not less. The part almost nobody is repeating: the ARC Prize Foundation published, months in advance, that public-set scores are "emphatically not a valid measure of progress," and released its own harness that scores 100 percent by replaying human play. The number that matters: on the hidden sets the Foundation actually uses, frontier models scored half of one percent at launch. And in the Foundation's own pre-launch test, a hand-built harness took a model from 0 to 97.1 percent in the environment it was built for, and from 0 to 0 in the room next door. Why that pair of numbers is every AI pilot a CEO has ever approved: the 94-percent pilot that lands in the sixties at rollout, and the postmortem that says change management when the truth is that the scaffolding was hand-fitted to the pilot set. The broken metric: the benchmark score on a vendor's slide. Not fabricated, just no longer a measurement of the thing being sold. The question is no longer which model. It is who built the harness, and was it built against the test. Sources: NVIDIA Technical Blog, "NVIDIA AVO Reaches 100% on ARC-AGI-3," August 21, 2026. ARC Prize Foundation, "ARC-AGI-3: A New Challenge for Frontier Agentic Intelligence," technical report: dataset composition, the public-set policy, the human-replay harness, and the Duke-harness transfer result. ARC Prize verified results for Claude Opus 5 (Public Demo, 30.16 percent, High reasoning effort, July 24, 2026) and the ARC Prize community leaderboard. Referenced: episode 137, "Stop Picking Tools. Start Assigning Layers." The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.

The same question keeps arriving from Milan, from Singapore, from Chicago. We are paying for Microsoft Copilot and we are paying for Claude. Which one should we standardize on? It sounds like a procurement question and it never is. This weekend edition takes the question apart and replaces it, because the honest answer is that it collapses two completely separate decisions into one: where your people think, and where the work lands. In this episode, Stephen Forte covers: New survey work from Recon Analytics covering more than 150,000 US respondents: where an employee has Copilot and nothing else, 68 percent use it. Where Copilot sits next to two alternatives, it takes 8 percent and ChatGPT takes 70. Same product, same people, and the only variable is whether they had somewhere else to go. Why that is a preference verdict rather than a quality verdict, and why preference is the one thing a policy cannot overrule. The researchers' own conclusion: distribution advantages do not lock in market position. An honest note on what that survey does and does not measure. It covered Copilot, ChatGPT and Gemini. It did not measure Claude at all. Where BuildClub itself sits, stated up front: we use all of them, and most of our heavy lifting runs on Claude. What each tool is genuinely better at. Copilot posts to Teams, attaches files to the emails it drafts, and can start working because an email arrived. Claude does none of those three. Claude writes and runs code. Copilot does not, and that single difference explains most reports of Copilot underperforming. The four-layer architecture that replaces the tool question: the interface, the hands, Teams, and the large population of people who are never leaving Outlook and should not be asked to. The one thing Microsoft deliberately will not let a machine do, and why they were right to draw that line. The workaround, and why it produces better governance rather than worse: a named owner, an accountable human, and nothing pretending to be a colleague. An invented but familiar scenario, a six-hundred-person industrial packaging firm with offices in Milan and Chicago, whose managing director is being asked to standardize by people who have already decided. One honest limitation, stated plainly on air: the moment Claude reads your content, that content has left your Microsoft tenant. Newer architectures keep it inside and are more limited today. You can have one or the other right now. Sources: Recon Analytics, "AI Choice 2026: Why Licenses Don't Equal Adoption," February 2026. Survey of 150,000+ US respondents, July 2025 to January 2026, paid AI subscribers. Microsoft Graph v1.0 reference, "Send chatMessage in a channel or a chat." The application permission is Teamwork.Migrate.All only, with the note that application permissions are supported for migration only. Microsoft Learn, "Copilot Cowork overview," "Use plugins with Copilot Cowork," and "Extend Microsoft 365 Copilot." Anthropic, "Microsoft 365 connector" documentation, for the documented limits on what Claude can and cannot do against Microsoft 365. Referenced: episode 131, "Your AI Tools Don't Share a Brain." The AI Brief from the YPO Technology Network is a daily executive briefing on the AI developments that matter to business leaders. Hosted by Stephen Forte.