AI in the News: Agents Outnumber Employees, a Model You May Not Be Allowed to Use and a Five-Dollar Attack

Price is no longer the hard part of model choice. Provenance is. A tool your vendor built on a model you cannot use becomes your problem the day a regulator or a large customer asks. The question for every AI vendor: which models sit underneath your product, where are they hosted, and can you switch without a rebuild?

Short post or talk clip: "Do you know which model is inside your software?" Offer a one-page vendor provenance checklist: underlying model, hosting country, swap-out clause, notice period.

Attackers are riding the same productivity curve you are. If an attempt costs five dollars, volume is the strategy, and small online and operating businesses are the target set. The readiness question is not whether you use AI. It is whether your basics (patching, access, payment pages) survive a thousand cheap tries.

Talk segment or post: "The five-dollar attack." Close with three controls a 50-person company can verify this month, and an offer to walk a leadership team through them in thirty minutes.

The scarce skill is supervision. The company with the best tools reports that its people increasingly spend their time managing agents, and that Claude sessions spend around 80% of their time idle, waiting on a human. That is a staffing and attention question before it is a technology question. One company's data, and the authors do not claim causation, but it is the clearest picture yet of the job after the pilot.

LinkedIn post and workshop exercise: "Name the reviewer." Pick one task an AI now does, name the human who reviews it, and write down what that person would catch that the AI would not.

Accounts, memory and triggers are the three things that make an agent behave like a colleague rather than a feature. For any Microsoft-based company, each one needs a named owner before the first agent goes live: who creates the account, what it remembers, and what is allowed to wake it up.

One-page "agent onboarding sheet" with three fields (account owner, memory scope, trigger list). Post it as a downloadable and use it as the opening exercise in a Microsoft-shop workshop.

Copy the design, not the percentage. A revision cap and a human release gate are what keep a 75% success rate from becoming a 25% cleanup bill. Seven percent of work is a more honest adoption number than most vendor decks will give you.

Case-study card: "Three revisions, then a human." Turn the cap-and-gate pattern into a template any team can apply to its own agent pilots.

Usage is up eight points on the daily measure and training access is down eight. Employers are buying the tool and cutting the teacher. The confidence gap between daily and occasional users suggests the fix is cheap: more reps, with support, for the people not yet using it.

Post: "Your people are using AI. Is anyone teaching them?" Offer a two-question pulse for leaders: how many hours of AI coaching did each team get this quarter, and who delivered it?

One company is a signal, not a trend. But it points at the planning question every leader will face: if the entry-level work is automated, where does your next generation of senior people learn the job? The answer is to redesign the first role around reviewing AI output, not to wait and see.

Workshop exercise: take one entry-level role and rewrite it so the junior's job is to review, test and question AI output. Share the before and after as a short case study.

Note-taking, onboarding and research are table stakes now. The gap is leadership: who owns the AI agenda, what is the measure of success, and what moves a pilot into production. Panelists described "bearing with it" through the awkward middle rather than restarting every quarter.

LinkedIn post for advisors: "63% have AI. 10% have a strategy." Offer a five-question firm self-check: owner, use cases in production, review step, client-data rule, next quarter's one bet.

Platform vendors are laying plumbing. What they cannot ship is the firm's own answer to three questions: who owns the client data, who reviews what an agent drafts, and what gets logged for the examiner. Be wary of "agentic" as a label when the features are still on the roadmap.

Short explainer for advisory firm owners: "What to ask your platform vendor about agents." Four questions, one page.

"Up to 90% faster" is a speed claim. The buyer's questions are accuracy and accountability: what is the error rate, who signs off on a missed endorsement, and does the carrier or the agency absorb the miss? Time saved is only value if the review step survives.

Post for agency principals: "Three questions before you turn on a policy-check agent." Offer a one-page pilot scorecard covering time, error rate and the named reviewer.

The AI that touches pricing and eligibility often looks like a data feed. That makes governance a data question: which inputs, validated by whom, and can you explain a decision built on them to a regulator or a policyholder?

Explainer: "The most important AI in underwriting has no chat window." Walk through input, model, decision, explanation and where a human checks each.

Expect compliance automation pitches in your inbox within the year. The evaluation question is the evidence trail: can the system show an examiner what it checked, when, and who approved? A tool that cannot reproduce its own work is a liability.

Post or talk clip: "Buying compliance automation? Ask to see the audit trail first." Pair with a five-line vendor evaluation card.

The product does not exist yet, so the control has to be procedural today. Any firm that moves money on a voice or message instruction needs a second-channel callback rule, written down and rehearsed. It costs nothing and covers the cheapest deepfake attack.

Short video: "The two-channel rule." Ninety seconds on why a familiar voice no longer authorizes a wire, with a one-page script for staff.

Established vendors are shipping agents into policy checks this week; YC-backed startups are going after claims, agencies and underwriting from outside. Incumbents will move faster, and customers will have more choices to evaluate. For any insurance leader, the live question is which step of your chain you want to be first to redesign.

One-slide "insurance value chain, AI edition": where each company sits, and the human decision that stays at each step. Use as a conversation starter at an insurance roundtable.

The conjurer is the Anthropic six-to-one number in a human voice. When a leader's job becomes assigning and evaluating work done by agents, the training gap is judgment: what to hand over, what to verify, and what you will not delegate. Capability charts make a good case for starting now. They do not tell you who is accountable.

Talk clip or post: "Centaur, cyborg, conjurer: which one is your team?" Ask each leader to place three recurring tasks on the map and name the review step for each.