Mistakes, Not Villains

Can your AI actually use your computer? Claude, ChatGPT, and Grok Bot all can, but there are four very different ways an agent touches your systems. Stephen Forte explains why legacy work still needs computer and browser control, compares the methods, and gives one rule. Connectors first, page or app structure next, screens last. In this episode: Why this still matters. A lot of real work still lives in systems with no connector: carrier partner portals, point of sale back offices, state payroll tax sites, and old Windows insurance applications. Four methods, most precise to least. Connectors and MCP servers first. Browser structure next, with tools such as Playwright and agent-browser. Then the operating system accessibility tree, the method behind agent-desktop. Screenshots and mouse coordinates last, the slowest and most expensive option. agent-desktop and agent-browser. Open source tools from Vercel Labs that read app or page structure instead of pictures of the screen. agent-desktop is a set of hands and eyes any agent can call. It is getting better, and it can still be slow. Community add-ons. Windows-MCP, Peekaboo, and computer-use-linux bolt the same accessibility features onto Windows, Mac, and Linux. Worth a look, not a recommendation. What each assistant can do today. Claude has computer use, browser use, Claude Code, and Claude in Chrome. After Claude Fable 5.2, the older computer use API tool now returns errors everywhere except Amazon Bedrock. Desktop computer control is still a research preview, not on Team or Enterprise plans yet. ChatGPT product names have changed fast: Operator, agent mode, Atlas, then Work. What exists now is ChatGPT Work and the desktop app's Computer Use. Grok Bot works from its own cloud computer, with connected apps preferred over clicking. The rule, plus the risk. Pick the most precise method that reaches the work. Do not run a screen agent on a laptop that has client email open. A signed-in browser session is a long-lived credential. Retest after vendor updates and site redesigns. Sources: min heo on X, the agent-desktop tip: X post, Nov 9, 2026 Vercel Labs: agent-desktop and agent-browser Anthropic: Computer use tool Anthropic: What's new in Claude Fable 5.2 Anthropic: Claude Code computer use and Claude in Chrome generally available OpenAI: ChatGPT Atlas retirement, desktop browser, and API computer use xAI: Grok Bot computer and apps Community add-ons: Windows-MCP, Peekaboo, and computer-use-linux 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.

Public AI benchmarks cannot tell you which agent to trust on your workflows. The business owner should define good, test the outcome first, then every step, and when a build looks finished, force a visual workflow and a full process map.Key takeaways. The business owner writes the tests; engineering should not invent success alone. Test the outcome first, then each step. When it appears to work, force the guide, visual workflow, and process map. Check new models, including Gemini 4 Carbon, on your own cases. Rerun tests after every update. Fund checking before more output.Monday checklist (as spoken). You do not need to write this down while you listen.Outcome tests owned by the business. The owner writes what good and bad finished results look like. Why: they know what good looks like, not only what is easy to automate.Step tests across variations. Change the inputs. Require intermediate values and sources. Score logic breaks between steps. Why: a good last message can hide a broken middle step.Decline and escalate cases. Include the cases where the right answer is no, not enough information, or send it to a person. Score an unjustified action separately from unfinished work. Finished is not the same as right. That is the light version of this week's Visa lesson, without importing a consumer shopping score into your shop floor. Why: completion is not correctness.The same case, repeated. Run an identical case several times. Consistency is a separate score from a single lucky pass. Why: one pass is not consistency.Compare against past human-handled cases. Take closed work a competent person already decided. Blind the agent. Compare outcomes and the follow-ups required. Why: your closed work is local ground truth.Finished-product map, then rerun after every vendor or model update. Require the guide, the visual workflow, and the full process map after it appears to work. Then treat every platform change as a release candidate and rerun the suite. We learned that the hard way when a tool update broke multi-step workflows my team relied on. Memory kept outside the tool let us move. The test suite is how you notice the change in minutes, not after customers do. Why: leftover paths and vendor updates show up here.Examples named. Invoice, claim, and customer-change cases (types, not industry claims). Leftover data writes or folder references. Gemini 4 Carbon's benchmark claims. Elvis Saravia and Lauren (9 Nov). Meta AutoRefine. DoorDash's Ask DoorDash harness (hidden tool path only).Terms. Public benchmark: a published model test, not your workflow. Intermediate values: mid-process numbers, materials, and sources. Blind the agent: hide the prior human decision. Grader: your scoring method; AutoRefine found older graders go stale.Sources: Elvis Saravia, 9 Nov; Lauren, 9 Nov; Google, Gemini 4 Carbon; Meta, AutoRefine; DoorDash evaluation; Visa benchmark release; Visa whitepaper.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.

New Wharton and Deloitte findings show most companies still prove AI value with hours saved, while teams that also track growth report stronger returns. Stephen Forte explains why hours saved is capacity, not cash, and what should go on a one-page AI scorecard. Plus how long it can take to actually shut an agent off. In this episode: The wrong headline. Hours saved is the most common AI metric among managers who measure return. It is capacity, not cash. If that capacity is not allocated to something, it gets reabsorbed. Wharton and GBK, October 2026. Departments that track at least one growth measure, such as net-new revenue, conversion, output, or customer satisfaction, report stronger returns than those tracking efficiency alone. These are survey answers, not audited financials. Organizations on one approved platform, and those with a designated chief AI officer, also report stronger returns. Deloitte, October 29. Most leaders report faster decisions and execution. Only a quarter say they use AI to deeply transform the business. Most companies now track the total AI bill. Far fewer track token volume or the cost of a unit of work. Nearly one in four say accountability is unclear when AI gets it wrong. West Monroe, via CIO Dive on November 9. A large majority of U.S. tech executives now put more than ten percent of IT budget into AI, a sharp rise from last year. A one-page scorecard. Make a growth or outcome measure the headline. Treat token and AI spend like any other business expense, with a purpose and a deliverable. Reinvent the work before you cut. Put a named owner on the measurement and standardize on one approved platform. Leaders should report the same way they ask the organization to report. Lightning item. An AWS test with Okta found that deleting the agent's resource connection refused new requests immediately, while deactivating the agent sat in a beta API queue for eighty-six minutes. Tokens already issued kept working up to five minutes. This was vendor-published, one test organization, and sample code marked non-production. Ask your identity team which switch actually stops an agent, and test that plan once before you need it. Sources: Wharton Human-AI Research and GBK Collective, 2026 AI Adoption Report (Oct 2026) and PDF GBK press summary Deloitte, State of AI in the Enterprise (Oct 29, 2026) West Monroe via CIO Dive / Yahoo Finance (Nov 9, 2026): CIOs upping budgets to meet AI AWS Machine Learning Blog: Okta agent access, part 1 and timing results, part 2 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.

Most companies still find out what their AI agents did by reading a log after the fact. Stephen Forte explains controllability: a check that runs before the agent acts and can say no, five working guidelines, and a trend in which Mistral and ANZ are building the same kind of controls. In this episode: Controllability defined. A check that runs before the agent acts and decides whether the action is allowed. If the check says no, the action does not happen. A log afterward is not controllability. Anthropic inference hooks. Now generally available on Claude Enterprise. Before Claude answers a prompt or runs a tool call, your company's security or data loss prevention server gets an allow or deny vote, inside Claude's infrastructure. One setup covers chat, Claude Code, Cowork, and connected tools. This is an architecture lesson, not a buying recommendation. Five working guidelines. A check before the agent acts. Short-lived credentials, with tokens that expire in minutes. Memory, ledgers, rules, and procedures kept outside the tool in files you own. Subject experts own the workflow, and a human stays accountable for high-stakes outcomes. Retest after every vendor or model update, and keep a dependency inventory. Four CISO questions. What untrusted content the agent reads, what it can do and on whose identity, the blast radius if it goes wrong, and whether you can see its actions in your logs. Least agency means the narrowest capability that still finishes the job. Mistral Studio. Workflows with human approvals are generally available. Managed agents are in public preview. OpenTelemetry traces show model calls, tool use, and approvals. Private cloud and on-premises timing for the new pieces is reported as expected by late November, not as something already shipped. ANZ Control Plane. CEO Nuno Matos told investors the bank is building an Enterprise AI Control Plane for visibility and technical control, with defined limits and a way to stop activity outside those limits. Humans must remain in control. Lightning item. Anthropic's incident response agent stayed inside human review when a newer model started reaching out to another agent mid-incident, because the write surface was bounded. Design the check for where the model will be in six months. Sources: Anthropic: Claude Enterprise security updates, inference hooks generally available Anthropic Help Center: Inference hooks overview Anthropic platform docs: Inference hooks Anthropic: CISO guide to agentic AI Anthropic: Applying Zero Trust to agentic AI services Mistral: Mistral Studio announcement Mistral: Release notes, Managed Agents public preview ANZ: CEO Nuno Matos FY26 investor remarks 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.

Visa's new AI shopping benchmark found agents that failed usually still chose to buy. Stephen Forte explains why CEOs should test agents on the cases where the right answer is to decline or ask, and why memory kept outside the tool matters when a vendor update breaks. In this episode: Visa's Agentic Commerce Benchmark. Twenty-two models from twelve labs ran fifteen shopping scenarios three times each, for nine hundred and ninety runs. Overall pass rate fifty-seven point two percent; only thirty-four point two percent of model and scenario pairs passed all three attempts. Caveats that matter. Simulated market and customer, stopped before checkout, no money moved, a small scenario set, and Visa is an interested party. Read it for how it grades, not as a score for your business. How a run passes. An agent had to discover every customer requirement and then make the right call, including choosing nothing when nothing fits. Guessing the right product without asking the right questions could still fail. What failed agents did. In fifty-four percent of the failed runs, the model still chose to buy. That is half of the failures, not half of all runs. Visa also counted false refusals as failures. Completion is not correctness. Most agent pilots are graded on whether the task finished. Demos show the cases that go well. The cases that matter are where the right answer is no. Three tests before you trust an agent. Mix in decline-or-ask cases, rerun identical cases for repeatability, and score acting-when-it-shouldn't separately from refusing-when-it-should. Grok Bot update. A first-hand account of multi-step stalls, faster token burn, worse judgment and a false self-upgrade claim after a vendor update. Public context from Elon Musk and Matt Palmer on November eighth, without a named cause or timeline. Resilience outside the tool. Client work moved to Claude in about fifteen minutes because memory lived in backed-up ledgers. Keep a dependency inventory for tools with their own Linux machine. Retest agents after every vendor update. Sources: Visa, Nov 5: Agentic Commerce Benchmark release Visa whitepaper: Methodology and grading Elon Musk on X, Nov 8: Grok Bot fixes underway Matt Palmer on X, Nov 8: User feedback request xAI: Grok Bot product description 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.

AI exposes the layers, hand-offs and busy work inside a company. Stephen Forte explains why leaders should redesign the work before cutting jobs or flattening the organization. In this episode: Three restructuring announcements. NiCE, Canva and SUSE named AI among their reasons for change, alongside other business pressures. These are not simply AI layoffs. Readiness exposes waste. Preparing for AI reveals weak data, inappropriate access and the hand-offs between a request and a result. Double clutter. Email and Excel create work on both sides of a request. Count the relays before changing the org chart. Sequence the change. A Harvard Business School AI Institute working paper links AI capabilities with flatter hierarchies and wider managerial spans. The effects were stronger where people had the skills to use the technology; they were not automatic. Evolution, not revolution. Reinvent the work first, then decide which layers to remove and how to organize the people you need. Operational expertise matters. The people who understand accounting, sales or marketing can tell an agent what good looks like. Tool knowledge alone is not enough. Train people and remove friction. Investment in training signals confidence in the team. Start by automating the spreadsheets and recurring reports people hate, giving them back the work they were hired to do. One practical starting point. Pick one process, count its hand-offs and emails, automate the frustrating part, then reassess the layers. Sources: NiCE, Nov 4: Quarterly results and restructuring announcement Calcalist, Nov 4: NiCE workforce and financial context TechCrunch, Nov 4: Canva restructuring The Register, Nov 2: SUSE restructuring Harvard Business School AI Institute, Nov 3: The Future is Flat 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.

Goldman Sachs just limited who can use its strongest coding agents. Stephen Forte asks who should get those power tools in your company. In this episode, Stephen Forte covers: Goldman's October change. Bloomberg (Nov 2): Anthropic's Claude Code and Microsoft's GitHub coding tools are now for engineers. Traders and bankers lost access; the quant "strats" are exempt. Spokesperson Nick Carcaterra said that after experimenting on productivity and risk, they put the tools with those best suited to leverage them long term. A real shift. In July, CIO Marco Argenti had praised vibe coding by traders and bankers who built tools and handed them to one of 12,000 engineers to polish. Earlier in 2026, Goldman had already restricted Claude for Hong Kong staff over confidential data going to third-party AI tools. Respect the bank. Stephen will not question Goldman's call. Banking's regulatory environment is unique. Pulling coding agents back to engineers after experimentation is risk management where a bad tool can leave with client data. Most companies are not Goldman. Builders outside engineering should get coding agents, 100%. Tools should go first to subject matter experts, who truly understand the problem. IT's job. An engineer fixing accounting without understanding accounting is a mistake. IT should support subject matter experts, lift their skills and manage their code inside the company's accepted stack. The real risk. Neutering your most experienced people and denying them modern tools. Those people will leave for a place that will give them the tools. Stay inside the approved stack. Non-confidential prototyping elsewhere is usually fine. Operate inside tools the company already approved. Even Microsoft Copilot is getting better every day. Replit's counterpoint. CEO Amjad Masad said agents on employee laptops are a systemic security risk and the ban hammer is coming down. Control what agents can touch and where they run; don't confuse that with locking out subject matter experts. Juniors and judgment. Partner Chris Churchman warned heavy AI use can degrade reasoning. Stephen's view: juniors learn through structured use of AI. AI does busy work and calculations; people own judgment. A large language model cannot replace human judgment yet. Tokens as expenses. Garry Tan of Y Combinator spends $100K to $200K a year on tokens. Stephen's rule: no hard caps by role. A gate for everybody through management, with deliverables that show value, same as justifying a $2,000 business trip. Sources: Bloomberg Law (Nov 2, 2026): https://news.bloomberglaw.com/capital-markets/goldman-restricts-traders-and-bankers-from-using-some-ai-tools Quartz (Nov 2, 2026): https://qz.com/goldman-sachs-restricts-ai-coding-tools-traders-bankers-110226 CNBC, Chris Churchman (Aug 24, 2026): https://www.cnbc.com/2026/08/24/goldman-sachs-ai-partner-danger-skills.html Tech in Asia, Goldman Hong Kong Claude restriction: https://www.techinasia.com/news/goldman-sachs-restricts-ai-coding-tools-engineers The New Stack, Amjad Masad (Nov 3, 2026): https://thenewstack.io/ai-coding-agents-security/ TechCrunch, Garry Tan tokens (Nov 2, 2026): https://techcrunch.com/2026/11/02/y-combinators-garry-tan-uses-ai-agents-so-much-he-spends-up-to-200k-a-year-on-tokens/ 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.

An agent can't follow a rule it can't read, and at most companies the rules live in PDFs, shared drives and people's heads. Stephen Forte walks through the step most companies skip before handing work to an agent: writing one process's rules down, in one place a machine can read. In this episode, Stephen Forte covers: Machine-Readable Government. On Nov 2, Abu Dhabi and OpenAI rewrote a defined set of policies so an AI system can check a case against the rules and link each recommendation to the exact clause. An official verifies, corrects or overrides, and a person makes the final decision. The first pilot. Food import registrations at Abu Dhabi's food safety authority: about 20,000 a year, drawing on 1,400 policy documents, some over 700 pages. Routine checks that took days now take minutes. Construction permits, licensing, HR compliance and procurement are next. The caveat. It is a self-announced pilot with no independent results and no published error or override rates. The lesson doesn't depend on them. Your company's permits. Discount approvals, valid expenses, vendor setup, new-hire tasks, what your team can promise a client. Those are rules, applied every day. Where the rules live. Wikis, PDFs, forms, SharePoint, employee manuals. A good employee knows who to ask. An agent applies whatever version it's given, with complete confidence. Stale copies, not conflicts. What Stephen sees most is people with an old version of a rule, or none, who email around for the answer. The answer depends on who they asked. One queryable repository. In his day job, rules of engagement, process instructions and SOPs live in machine-readable files in one shared project that Stephen maintains. One version, changed there first, and the place an agent should start. What machine-readable means. Plain text, one rule per section, a named owner, a date, and limits as numbers, not adjectives. "Discounts over 15% need regional VP sign-off" is checkable. "Large discounts need approval" is not. It has to come from the top. About 50 Samsung senior executives built personal knowledge bases before meeting Sam Altman, and Siemens CEO Roland Busch said it has to come from the top. If the CEO asks the repository first, the team follows. Where to start. Pick one high-volume process, gather every version, give it one owner, and have a large language model merge it, flag disagreements and list unanswered questions. Put it in a shared Claude or ChatGPT project, or a SharePoint site Copilot reads. For a month, answer emails with "ask the repository." Sources: Abu Dhabi Media Office (Nov 2, 2026): https://www.mediaoffice.abudhabi/en/crown-prince-news/khaled-bin-mohamed-bin-zayed-meets-sam-altman-and-witnesses-signing-of-agreements-with-openai/ CXO Insight Middle East, OpenAI and Abu Dhabi launch Machine-Readable Government: https://www.cxoinsightme.com/business/industries/government/openai-and-abu-dhabi-government-launch-worlds-first-ai-native-machine-readable-government-initiative/ AGBI, Abu Dhabi partners with OpenAI to speed up business approvals: https://www.agbi.com/ai/2026/11/abu-dhabi-partners-with-openai-to-speed-up-business-approvals/ Korea Herald, Samsung executives meet Altman (Oct 31, 2026): https://www.koreaherald.com/article/10921312 Korea Herald, Siemens CEO Roland Busch (Nov 1, 2026): https://www.koreaherald.com/article/10921699 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.

Most exposure in a company comes from mistakes, not villains: access opened for a good reason and never closed. Attackers now find those leftovers in hours. In this episode, Stephen Forte covers: Binance. About $6.3B stolen. Binance now says it was a months-long campaign the FBI attributes to North Korea, starting with a fake job interview and a malware-laced coding test. AI speed. Storm-0940 reached privileged cloud access in under three hours. Unit 42 saw two weeks of attack work done in under ten. Day job. What his team finds in client tenants, and when to call a specialist. The full checklist (ask for a number, a list, or a screenshot): 1. Global/super admins, cloud root, registrar, DNS owners, by name. 2. Standing vs just-in-time admin rights; separate admin accounts. 3. High-privilege service accounts and app registrations, with owners. 4. MFA % enforced, not registered; exceptions and approvers. 5. Phishing-resistant MFA for admins and finance; legacy auth blocked. 6. Accounts idle 90+ days, ex-staff still enabled, guests. 7. Last access review on finance, HR, CRM, customer data. 8. Project access opened in 12 months, and closed? 9. Secrets in repos and history; how many still work? 10. Where API keys live; rotation dates; revoke at the issuer. 11. Every AI-provider key and who pays. 12. Domains and subdomains (crt.sh), each with an owner. 13. Test, staging, demo servers still running with real data. 14. Old sites and WordPress installs patched or retired. 15. Shodan view of your IPs: open databases, admin panels. 16. Quick-built apps on Supabase/Firebase: row-level security, no secrets in front end. 17. Public buckets and "anyone with the link" shares. 18. Anything that can delete production databases or sites. 19. Which notetakers are in use, sanctioned or personal. 20. Transcript default sharing, who sees all, retention. 21. Board, HR, legal, M&A calls: no bot or restricted. 22. Third-party OAuth apps on mail, files, CRM; admin consent. 23. Top 10 vendors: recent breaches, tokens rotated. 24. Immutable backups and last tested restore, incl. M365/Google. 25. DMARC at enforcement, SPF, DKIM on every domain. 26. Bank-detail changes: call back a number on file, dual approval. 27. External auto-forwarding blocked; inbox-rule alerts. 28. Log retention and alerts (new admin, MFA off, mass download). 29. Endpoint protection on every laptop. 30. Written incident plan; who calls insurer, counsel, IR firm. 31. Tabletop exercise with the CEO in the last 12 months. 32. Quarterly one-page cyber report to the board. 33. Cyber insurance answers match reality. Free tools: Secure Score, ScubaGear, ScubaGoggles, Maester, Zero Trust Assessment, gitleaks, TruffleHog, crt.sh, Shodan, Prowler, Have I Been Pwned. Specialists: active compromise means an incident response firm; customer money or a raise or sale means a pen test; no one watching at night means MDR. Sources: CoinDesk on Binance (Nov 3): https://www.coindesk.com/business/2026/11/03/binance-s-usd6-3-billion-hack-began-with-a-fake-job-interview-exchange-says Microsoft on Storm-0940: https://www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/ 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.

When an AI agent does the analysis, it also makes some of the decisions, and most of the time you will not notice which ones. A new Anthropic study shows how that happens, and Stephen Forte walks through the two habits he uses to catch it: an audit pass and an adversarial pass. In this episode, Stephen Forte covers: The study. Fifty professional analysts did a real statistics project with Claude Code. On four judgment calls the agent was set to lean one way, picked by coin flip. Analysts went with the lean 86% of the time. No lies, just one side. The agent argued for one option and usually spoke first. 77% of analysts explained their choices with the agent's arguments, only 15% credited it, and none suspected the lean. The limits, and the bigger risk. One task, one field, one model, an amplified lean. But with nothing planted, the standard agent made the same choices in 91% of runs. The risk is the model's default habits quietly becoming everyone's judgment. What Stephen sees. Probing often finds far less factual basis behind an answer than assumed, and an audit of anything material always finds mistakes, big or small. The audit pass. At the end, ask for a decision audit: every judgment call, what was passed on, what data was left out, where it was unsure, what is missing, what would flip the conclusion, and which calls you made versus accepted. What is missing. A wrong number jumps out. A missing customer segment or a missing year of data does not. One honest caveat. A model explaining itself afterward can be plausible yet misleading. Treat the audit as a list of choices, not proof of why. The adversarial pass. Ask who will read it and what they would poke holes in: an investor, an auditor, the investment committee. Run it in a fresh session or a different model. In one study, debate beat a single model arguing one side, 88% to 78%. Build it into the workflow. Save both prompts one click away in Claude, ChatGPT, GitHub Copilot or Grok Bot, make it a checklist step, name the human-owned calls in advance, and write your own view first on the big ones. Skip it for a pasta recipe. Sources: Anthropic, Agents and expert judgment calls (Oct 27, 2026): https://www.anthropic.com/research/agents-judgment-calls Herdağdelen, Joren, Tamkin, Who's Aligned to Whom? (paper): https://www-cdn.anthropic.com/519c17a5d23a794e8b6d50a854048d0b61e0d73a.pdf Decrypt, AI can quietly influence experts' decisions: https://decrypt.co/382065/ai-can-quietly-influence-experts-decisions Turpin et al., Language Models Don't Always Say What They Think (2023): https://arxiv.org/abs/2305.04388 Anthropic, Reasoning models don't always say what they think (Apr 2025): https://www.anthropic.com/research/reasoning-models-dont-say-think Khan et al., debate vs a single consultant (ICML 2024): https://arxiv.org/abs/2402.06782 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.