YJ

one reason AI pricing feels weird companies sell the unit the infrastructure understands tokens. credits. runs. actions. seats. the customer does not wake up wanting 40,000 additional AI actions. they want: every qualified lead answered every call captured every proposal followed up every customer issue routed every important exception surfaced sell the thing the customer can recognize in their business. not the thing your API recognizes on the invoice.

Google's Gemini 4 Argon can autonomously find, validate and patch software vulnerabilities in its cyber defense setting. OpenAI is building persistent agents. Google is turning reusable instructions into Skills. the direction is becoming pretty obvious. we spent the first phase of generative AI asking: “what can this model tell me?” the next phase is: “what responsibility can this system reliably own?” that is a much bigger question. because once software can own a recurring piece of work, you're no longer just selling intelligence. you're redesigning the company around it.

automation without observability is just faster mystery. this is going to become painfully obvious as companies deploy more agents. if an agent sends the wrong email at 2:14 AM, I want to know: what triggered it what context it saw what decision it made what tool it called what changed and how to undo it not: “huh. that’s weird.” people obsess over making agents autonomous. I’m increasingly interested in making them easy to investigate when autonomy goes sideways.

AI agents are going to make bad lead forms look increasingly absurd business: “please enter your company name” agent already knows domain. “what industry are you in?” agent can research it. “company size?” researchable. “what are you interested in?” you have 4 previous emails from me. “tell us about your business” good heavens brother, you have a CRM. the future of personalization isn't generating a creepily personalized first sentence. it's stopping customers from repeatedly telling the business things it already knows.

the easiest AI service to sell might be a job somebody already hates doing. not just “AI transformation.” not just “agentic infrastructure.” find the sentence: “every morning Sarah has to…” there’s your starting point. map what Sarah checks. map the decisions she makes. map the exceptions. map what she updates afterward. automate the predictable middle. keep Sarah for judgment. this is significantly less glamorous than building AGI. it also has the wonderful property of having a customer.

cheap ai model vs expensive ai model is going to become basic business architecture you don't need frontier reasoning to extract an email address categorize a ticket summarize a call check a field route a lead save the expensive intelligence for: ambiguous requests complex research negotiation prep high value accounts hard decisions cheap model for volume. strong model for ambiguity. human for consequence. that little hierarchy can save an absurd amount of compute.

the early OpenAI Dots reviews contain a lesson every AI agency should tattoo somewhere visible. Dots reportedly did well on structured projects involving user data and controlled environments. random consumer errands across the public web? much messier. security checks. login friction. failed purchasing flows. human intervention. this pattern keeps showing up across agents: the narrower the job the cleaner the context the clearer the permissions the better the system becomes. “do everything” makes a good demo. “you own this responsibility” makes a good product.

the AI agent is probably not the product. the boring stuff around it is. hosting it. monitoring it. fixing integrations. handling edge cases. updating rules. watching costs. checking logs. recovering failures. being there when the client says: “uhhh why did it do that?” this is why “clients can just use ChatGPT themselves” misses the point. technically I can also host my own email server. I’d rather lick a battery.

one of the weirdest new questions in business is: what should your software be allowed to know? not just what it can do. what it can see. agents can potentially ingest: Slack conversations customer history internal documents emails meeting transcripts financial information employee discussions then reason across all of it. that’s extraordinarily powerful. it’s also why “connect everything” is terrible agent design. give the system the minimum context required to perform the job. context is capability too.

new AI releases have become so frequent that following them can become procrastination disguised as research omg new model !!!! rewrite the stack !!! new agent!!!! rewrite the offer!!! new benchmark! rethink the business!!! stop. your customer still wants roughly the same thing they wanted Wednesday. more revenue less cost less waiting less repetitive work fewer mistakes better service focus on the thing that matters.

“trust the AI company” is not an enterprise security policy. doesn’t matter whether the model underneath your workflow comes from OpenAI, Anthropic, Google or somebody else. the model provider doesn’t know: which customer is sensitive which action is irreversible which discount requires approval which database cannot be touched which mistake costs you $50 and which mistake costs you $50,000 that’s your job. the lab provides intelligence. you provide the boundaries.

California's attorney general has now subpoenaed OpenAI as part of an investigation into cybersecurity risks involving its AI models this is where “agent safety” stops sounding like something only research labs need to care about because when YOUR agent sends the email changes the record touches customer data updates the account or makes the promise the customer sees your logo not the model provider's benchmark score

an agent with unlimited API spend is basically a corporate credit card that can operate at machine speed. please enjoy that sentence for a moment. then add: spend ceilings scoped credentials rate limits usage alerts action logs credential rotation kill switches AI security gets described like some exotic futuristic discipline. half of it is still: don't give the robot your fvcking master key.