AI in the News: Banks Block the Agents, an Assistant That Starts the Conversation and a 22-Year-Old Flaw

The shift is from tools you prompt to tools that prompt you. Once an assistant acts in the background, the question stops being 'what can it do' and becomes 'what is it allowed to do without asking.' That is a policy question, not a software question.

LinkedIn post or 3-minute talk clip: 'The approval gate is the product.' Walk through Duet's pause-before-payments design and ask leaders to list the five actions in their business an agent should never take without a human yes.

An AI model found a bug that survived 22 years of human review. Attackers have the same class of tool. Assume the time between patch and exploit is now measured in hours, and that every WordPress, Drupal or vendor-run site you pay for is in scope.

Checklist post or workshop exercise: 'The 30-minute vendor inventory.' Ask leaders to list every website and portal run by a third party, then ask each vendor three questions: are you patched, how do you detect hostile AI agents, and who tells me if you're not.

The banks chose to block first and write the rules second. That works for a quarter. The firms that decide now which agents may connect, what they may see, and who is liable will be the ones clients choose when the agents show up anyway.

Anchor LinkedIn post and a 'Know Your Agent' discussion card for wealth and insurance leaders: three questions every firm should answer about client-side agents (which connect, what they see, who owns the outcome).

Exclusive cloud partnerships are dissolving. For a mid-market buyer that is good news: the model you choose no longer dictates the cloud you are locked into. Ask your vendors where your data lives and how portable your prompts and workflows are.

Short explainer graphic: 'Model, cloud, data: three decisions that used to be one.' Use it to reframe vendor selection for a leadership team.

Regulation is arriving as contract language, not legislation. If you sell to government, or to anyone who sells to government, your AI data handling will be asked about in procurement before any law forces it.

Explainer post: 'Three questions a procurement team will now ask about your AI,' with a downloadable one-page vendor-readiness checklist.

The skill shortage is not prompting. It is framing, supervision, verification and governance: the human work around the agent. Organizations are buying the technology and skipping the job description.

Workshop exercise: write the job description for 'agent supervisor' in your own department. Which of the five orchestration skills does someone on your team already have?

Same tools, same budgets, a 43-point swing in success. The variable is the people. If you are choosing between another license and a training cohort, the data favors the cohort.

One-chart LinkedIn post: 71 vs 28 with the single sentence 'Same technology. Different readiness.' Follow with the three habits of the 71.

The data vendors are now shipping agents. When the research layer your advisors already trust adds an agent, adoption stops being a firm decision and becomes a default. The governance question is what the firm records and supervises.

Post or short video: 'Your research vendor just became an AI vendor.' Three questions for a wealth firm: what do the agents see, what do they log, who reviews the output before a client does.

Insurers are not rejecting general AI; they are insisting it be governed and specific to policy wording and appetite. The same logic applies to any firm: the closer the decision is to money or liability, the more context and oversight the model needs.

Post: 'Six percent.' Why almost nobody will trust a general model alone on a claim, and what a 'governed hybrid' means in plain language.

Notice what is for sale: not a model, but governance across the handoffs. In regulated work the agent is the easy part. The sign-off chain around it is the product.

Carousel: 'Where agents sit in an insurance workflow,' mapping the seven use cases to the human approval each one needs.

Agentic payments are moving from demo to pilot. The finance question isn't whether an agent can pay an invoice, it's what mandate sits underneath it. Practitioners say the sensible order is agents repairing payments first and paying second.

Explainer: 'Mandate before money.' A three-step policy for letting an agent touch accounts payable (read, repair, pay) with the approvals for each.

When YC asks for it, expect a wave of vendors in 12 to 18 months. Firms that clean up their compliance workflows now will be buying from a better market than firms that wait.

Post: 'What YC thinks compliance will look like in 2028,' with three tasks in your own compliance process an agent could already do.

Seeing a face or hearing a voice is no longer proof. Any approval process that depends on 'I recognized them' needs a second factor today, not after the vendors catch up.

Training moment: a five-minute 'call-back protocol' exercise for payment approvals. Practice it in a leadership meeting.

Notice the framing: not a feature, an operating system. The vertical-OS pattern is how YC-backed founders are going after insurance, and it is the same pattern executives are starting to ask for internally.

Compare-and-contrast post: 'Three ways an agency could build an operating system versus buy one.'

Thirty months is short enough that most leaders remember the first ChatGPT pilot. Looking back at it in charts is the fastest way to reset an executive team's assumptions about what 'six months from now' means.

Workshop exercise: have each leader draw their own chart of what their team could do with AI 30 months ago versus today, then 12 months ahead. Compare the shapes.