Five connected readings on the AI-stock sell-off, AMD, Palantir, the U.S. election results and Canada's budget. What changed, why it matters and practical actions for the business.
Five connected articles on AI liability, the cost of capital, digital identity, insurance claims and agent agreement. Expanded November 1 with Ray Kurzweil's digital twin and Codex Micro, the custom keypad from OpenAI and Work Louder.
October 23, 2026: The work between capability and value
Five connected articles on investment rules, productivity, software security, agent authority and practical AI in education.
Dated signals on insurance, AI and enterprise systems. The earlier notebook analyses remain below with their original sources.
October 12 to 16, 2026: AI meets the world outside the model
October 5 to 11, 2026: shared meaning, named authority and insurable boundaries
SIGNAL NOTE • SEPTEMBER 29, 2026
Not every AI service needs to hold a conversation.
Generative models are useful when the work involves language, exploration or explanation. Many operational questions are narrower: Is this submission complete? Does it fit appetite? Which queue should receive it? Is confidence high enough to proceed? A smaller decision service can answer quickly, cheaply and in a typed form that software knows how to use.
Specialization does not automatically make the answer correct. A probability still depends on the question, evidence, calibration and population on which it was tested. But the constrained output can make evaluation and integration clearer than asking a large model to write a persuasive paragraph.
Design the workflow as four components:
extract facts from messy material;
make a bounded decision or score;
explain the evidence and uncertainty;
send consequential or ambiguous cases to a person.
Use the right model or rule for each component. Test them independently. This reduces cost, but more importantly it makes failure easier to locate.
The future enterprise stack will not contain one intelligence. It will contain several forms of intelligence connected by explicit responsibilities.
SIGNAL NOTE • SEPTEMBER 5, 2026
A large insurance coverage gap makes an attractive AI headline. It also hides several different problems inside one number.
People can be underinsured because they do not understand the risk, cannot afford the product, distrust the seller, find the process difficult or have been offered coverage that does not fit their lives. AI may help with some of those conditions. It can explain options, identify likely gaps, reduce application friction and help an advisor prepare for a more useful conversation. It cannot make an unaffordable product affordable or repair trust through a better chatbot.
There is also a danger in treating the gap only as unclaimed revenue. Insurance exists to create resilience. A recommendation system should be judged by whether a person ends up with suitable, understood and sustainable coverage, not merely whether a policy was sold.
Choose one customer segment and separate the coverage gap into awareness, affordability, eligibility, product fit, trust and process friction. Apply AI only where the obstacle is informational or operational. Track completed coverage, customer understanding, early lapse and complaints, not just leads and conversion.
The useful signal is not that AI can close the gap on its own. It is that we can now investigate and address specific causes at a much more personal scale. The design must remain as careful as the promise.
SIGNAL NOTE • JUNE 12, 2026
Straight-through underwriting is often described as if autonomy were a switch: the person is either in the loop or out of it. In practice, autonomy should be a ladder.
An agent can begin by assembling evidence. Then it can recommend appetite, identify missing information, produce a quote-ready package and suggest terms. Only after the carrier understands performance by segment should it consider allowing narrowly defined cases to proceed further without intervention.
E&S is an important test because speed matters and the work is heterogeneous. It is also where the exception can be the whole point. A system that performs the majority of analysis may still need a person to recognize the one exposure that changes the decision.
Set explicit levels for each class of business:
summarize only;
prepare and route;
recommend with evidence;
produce quote-ready output;
proceed within a tightly bounded authority.
Name the evidence required to move up a level, the conditions that force a step down and the underwriter accountable for the design. Audit by class, broker, exposure and confidence band.
“Straight through” should describe a governed path, not the absence of people.
SIGNAL NOTE • JUNE 11, 2026
A large technology provider integrating a frontier model into systems used by banks, airlines and insurers is more than another partnership announcement. It is a new dependency entering old, consequential machinery.
The attraction is clear. An established provider understands legacy integration, contracts, operations and regulated customers. The model company supplies capability that would be difficult for each institution to reproduce. Together they can make adoption faster.
The risk is that layers of convenience make the real architecture harder to see. Which organization controls the prompt, retrieval, model version, logs and safety settings? Can the customer identify when behaviour changed? What continues to work during a model, network or vendor outage? Who can investigate an incident across organizational boundaries?
Before treating the integration as core, run three exercises:
the model is withdrawn or materially changed;
the integration is unavailable during a critical business period;
a regulator or customer asks for the evidence behind one output six months later.
Document the fallback, data boundary, version control, logs, support path and contractual responsibility. A strong partner can reduce implementation risk. It cannot eliminate the institution's responsibility to understand the system it has placed in the flow of work.