AI in the News: Agents Lose Their Internet Access, Review Becomes the Bottleneck and Insurers Meet AI Liability

The lab learned this from a review that started in July, not from a live alarm. The fix was containment and monitoring, not a smarter model. For any company deploying agents the question is the same: what can this agent reach, who sees what it does as it happens, and how long until you would know.

LinkedIn post: 'The vendor switched off the internet.' Three questions for any AI agent in your company (what can it reach, who watches it live, how fast would we find out), closing with a ten-minute Monday exercise a leadership team can run on one tool.

Only 35% named writing code as the bottleneck; 57% named reviewing, testing and deploying. Speeding up one step moves the queue to the next one. This is not just a software story: every function that adopts AI to produce faster will discover that checking is the scarce skill.

Workshop exercise: 'Where did your queue move?' Each participant maps one AI-accelerated task in their function, marks the new bottleneck downstream, and names who owns the review step. Post the blank map as a LinkedIn document.

Read the fine print: reports are sent without human review, so some may contain errors. Even the vendor's own security product ships with a human-verification caveat. Whatever AI finds or drafts in your shop needs the same owner and the same sign-off. Also note which firms signed on: the consulting giants are positioning around forward-deployed delivery.

Short video: 'Even the scanner needs a second pair of eyes.' Use the no-human-review line as the example for a three-line policy on AI-generated findings: who reads it, who signs it, who can overrule it.

The complaint cites over 160,000 entries in WebText and more than 122 million tokens in C4. The practical lesson for a mid-market buyer is provenance: if you cannot say what data your AI tools were trained on or what you feed them, you are carrying a risk you did not price.

LinkedIn post: 'Do you know where your AI got its homework?' Offer a one-page vendor question list: training data, indemnification, what your inputs are used for, and how to exit.

This is an allegation, not a finding. But the claim is a new shape of AI risk: not a data breach, but an antitrust theory about a shared pricing tool. Any company using a vendor's AI to recommend prices, rates or quotes should know whether competitors use the same engine.

Short video: 'Same tool, same answer, same problem.' Explain algorithmic pricing risk in 60 seconds with one question to ask any vendor: who else uses this model, and how are we separated.

Demand did not vanish; it moved to experience and judgment. GitHub says AI agents touched about one in three code pull requests as of October 2026, up from under one in ten a year earlier. If juniors are not doing the reps, where do your future seniors come from? Whoever answers that first wins the hiring market in three years. (Causation is unconfirmed.)

LinkedIn post: 'Who trains your next senior?' Three options for rebuilding the apprenticeship ladder when AI does the entry-level tasks, ending with one question for the reader's own team.

Whatever you think of the politics, the mechanism is the lesson: define the new roles before the displacement arrives. Most SMBs have no written role definition for who supervises, trains or audits AI. A one-page AI role map is a cheap start.

Workshop exercise: 'Name the three jobs your AI created.' Each participant writes titles, owners and the skill each requires. Convert to a downloadable one-page role map.

The line clients draw is not 'AI or no AI.' It is disclosure and voice. Back office, yes. Their advisor's own words, no. Clients 60 and older hold most of the wealth and prefer the human touch. A written 'where we use AI, where we don't, and what we tell clients' page is now table stakes.

LinkedIn post: 'The three-column AI disclosure.' Used behind the scenes / Used with advisor review / Never used, with 79%, 87%, 40% and 33% as the frame. Offer the template to advisory firms.

Advisors are using AI to write, not to decide. The top barrier is not cost or capability; it is not knowing what compliance will allow. That is a leadership gap, and a readiness engagement fills it with a written supervised-use policy.

Short video: 'Two barriers, one fix.' Compliance uncertainty and training are both answered by the same artifact: a one-page supervised-use policy plus a practice session.

Fronting premium grew from about $6 billion in 2020 to over $22 billion in 2025. If the value chain unbundles, an insurance executive's job shifts from running a company to choosing which layer to own. The author runs an AI-native wholesaler, so read it as a thesis from someone with a stake.

LinkedIn post: 'Which layer do you own?' Draw the insurance value chain in five boxes (capital, underwriting, distribution, claims, service) and ask where AI changes the economics of each.

Your insurer is building the scorecard before your board is. Observability, autonomy and governance are the three words a carrier will use to price your AI. A company that has written answers on all three will pay less and renew easier. A timeline of agent incidents published the same week notes that liability can arise from AI activity the deployer did not intend.

LinkedIn carousel: 'The three AI questions your insurer will ask.' Observability, autonomy, governance, each with a plain-English test a CEO can run in a meeting.

The insurance market is pricing agent risk as its own line. The underwriting method is the checklist: public scan, governance questions, behavioral testing. Your readiness work is, in effect, preparation for that questionnaire. Both are two-person teams and the claims are company-reported, so treat them as direction, not proof.

LinkedIn post or talk clip: 'Insurers are writing the readiness test.' Take Klaimee's three underwriting blocks and translate them into a ten-question self-assessment for an SMB owner.

The architecture is the lesson: a supervisor agent, specialist subagents and a human underwriter who decides. Same coordinator pattern as every agent platform this year. All figures are company-reported; ask for the client's own numbers.

Workshop exercise: 'Draw the supervisor.' Map one inbox-driven workflow (quotes, claims, onboarding) as supervisor, specialists and a human sign-off, then mark where the record is kept.

The deepfake wire is a procedure problem before it is a technology problem. A callback to a known number and a second approver defeats a convincing video. Compliance cost growing faster than revenue is the reason regulated firms will buy AI for it, and the reason they will need supervision of it.

LinkedIn post: 'The $25 million video call.' Give a three-step wire-request verification (known callback number, second approver, delay for new instructions) that any advisory firm or agency can adopt this week.

Mollick lists four reasons an agent should ask a human: approval (spending money, contacting outsiders, touching sensitive material), expertise, variance of ideas, and interest, meaning people should keep the engaging decisions. That is a ready-made agent permission policy in four lines.

Workshop exercise: 'The four reasons to look up.' For each agent or AI tool in the room, participants write which of the four triggers should force a human check, then compare. Pair with the Anthropic containment story.