SNACKS
AI for business leaders
Straight talk from the YOR.AI team on new research and what's working in AI, what's noise, and what business leaders need to know to make smart decisions.
AI Didn't Eat Your Software Budget. It Opened a New One.
The story for most of this year was that AI spend would come out of software spend. Agents would replace SaaS seats, the money would move, and the CFO's job was to manage a reallocation. The data says otherwise. Menlo Ventures' survey of enterprise AI buyers found that only about a fifth of this year's AI spend came out of existing software budgets. More than half came from budget that didn't exist a year ago. AI didn't eat the software line. It opened a line of its own, and most companies are still trying to manage it with the tools and reflexes built for the old one. We call that new line the net-new line, and the question it raises is what the company is getting for it, because nobody set a return target for a line they didn't know they'd have.
Your AI Plan Didn't Get More Expensive. It Got Smaller.
The subsidy era in AI isn't ending with a price increase. It's ending with the thing you bought getting smaller while the invoice stays the same. Last week users of one of the major consumer agents started hitting usage limits that had never existed, mid-plan, with no notice, and the explanation from the top was that the cost of serving the product had run ahead of what it earned. The plan price didn't move. The capacity behind it did. Every flat-rate AI plan carries the same unwritten term, and the enterprise version is worse than the consumer one, because team plans are how most companies budget AI and almost none of them have a capacity floor in the contract. We call it the shrink clause. It says the vendor can cut what your plan delivers without changing what your plan costs, and you'll find out when your people hit the wall.
Your AI Budget Is Attached to People. Your AI Spend Isn't.
Every AI budget cap I've seen this quarter is attached to a person. Apple cut some engineers from three hundred dollars a day to two hundred. Uber capped individuals at fifteen hundred a month. The pattern is everywhere, and it misses the thing it's trying to control, because almost none of the spend that worries a CFO is attached to a person anymore. It's attached to a workflow. An agent that reconciles a week of transactions overnight isn't anyone's seat. A pipeline that drafts, reviews, and files a hundred contracts belongs to a process, not a login. Budget by headcount and you can't see that spend, can't govern it, and can't tell whether it was worth it. We call the habit the seat reflex. It's the instinct to budget AI per person because software was always sold per person, carried into an era where the spend follows the work.
You Don't Need to Farm to Eat. You Still Need to Think to Use a Machine That Thinks.
There's an optimistic argument going around that I mostly agree with, and the part I don't agree with is the part that matters. The argument runs like this. For most of history, most people had to farm, because food was scarce and producing it took nearly everyone. Then productivity rose, the share of people farming collapsed, and the result wasn't a population of useless mouths. It was a population freed to do everything else, eating better than any peasant could have imagined, without knowing how to grow a thing. Intelligence, the argument says, is about to go the same way. Most of us will stop producing it and start consuming it, and the demand for it will turn out to be as bottomless as the demand for everything else once it got cheap.
The Lab Running Your Agent Might Ship a Product in Your Category Next Quarter.
The company whose model runs your customer-facing agent is one product launch away from being your competitor, and it will enter your category with a head start you handed it. The frontier labs have stopped being pure model suppliers. Every one of them now ships vertical products, coding agents first, and the pattern is already visible in that category. Tools built on a lab's model watched the same lab launch its own version, and the ones that survived moved to open models to get out from under a supplier that had become a rival. That's a startup story today. The enterprise version is quieter and worse, because the lab doesn't need to beat you at your business. It needs to learn your workflow from your usage, package it, and sell it to your customers and your competitors at once. We call the supplier that can do this the upstream competitor, and most companies have one and haven't named it.
The Cheapest Strong Model Is One Your Risk Team Won't Approve. That Gap Has a Price.
Large companies are paying a premium for AI that their smallest competitors aren't, and it has nothing to do with capability. Several of the strongest and cheapest open models on the market right now come from Chinese labs. Startups put them into production without a second thought. A Fortune 500 company, a bank, or a public agency almost never does, for reasons that are reasonable on their face: regulatory exposure, customer perception, procurement risk, and the fact that nobody wants to explain the model's origin to a board or an auditor. That caution is a posture, and the posture has a cost. The difference between what your risk policy lets you deploy and what your least constrained competitor deploys is a price you pay for where a model came from. We call it the origin premium, and most enterprises are paying it on every workload, including the ones that never needed the protection.
Your Agents Know Your Data. They Don't Know What You Decided Last Tuesday.
Every agent deployment assumes the hard part is getting the agent your data. The harder part is getting it your decisions, and decisions don't live in databases. They live in meetings, hallway conversations, and direct messages, and an agent can't see any of those. So it keeps executing the last decision it could see. Anthropic published a study of its own workforce this month, and the detail worth a leader's attention wasn't the productivity multiples. It was that people inside the most agent-heavy company on earth are under constant pressure to move their work into places agents can read, because anything decided out of view goes stale the moment it's made. We call a decision made where no agent can see it a dark decision. Your company is making hundreds a week, and your agents are running on the ones that came before.
The Internet Just Had Its Worst Month in a Decade. What Does It Mean for You?
This month, security researchers disclosed two critical flaws in PHP, the language behind most of the web's backends, and within days defenders were counting exploit attempts in the millions. Attacks began before the patch was even public. Government agencies told organizations to patch immediately or take systems offline. The lesson for business leaders has nothing to do with PHP. The window between a flaw becoming known and being attacked at scale has collapsed from weeks to hours, because both the finding and the exploiting are now done by machines. Security run on human cadence, business-hours response, patch cycles measured in sprints and quarters, no longer covers you. What covers you is what we call standing defense, security run as a standing agent workload that watches, patches, and verifies around the clock, with humans setting its bounds and handling its escalations. The companies that weathered this month already ran that way. Everyone else was reminded why they should.
The Model Found a Thousand Problems. Who's Going to Fix Them?
Point a capable model at an existing operation and it will find problems faster than the organization can fix them. Bank of America just put a price on that. After frontier models surfaced flaws in code the bank and its vendors run, it assigned about three thousand of its own people to the fix and then brought in roughly a thousand contractors on top, and its third quarter carried a couple hundred million dollars plus in remediation costs as a result. The model did exactly what it was asked to do. The bill came from what happened next. Every organization has a rate at which it can act on what AI finds, and above that rate a finding stops being an asset and becomes an unfunded obligation. We call that rate the remediation ceiling, and most companies are about to hit theirs without having measured it.
You Run Open Models. Your Vendor Still Owns the Loop.
Most companies that say they've gone open-weight are running those weights inside a closed harness, and the harness is where the dependency lives. OpenAI's Codex now runs any open model. Claude Code is closed and will run them through a proxy. Every lab's agent tooling is happy to host a competitor's weights, because the lab keeps what it actually wanted: the user, the defaults, the upgrade path, and the usage data that trains its next model. We call the story a team tells itself at that point the open-weight alibi. It's the belief that because the model is open, the stack is independent. The weights were never the lock-in. The loop around them is, and swapping the model inside someone else's loop changes nothing about who owns it.
You Built a Swappable Model Layer. Your Cache Just Locked It Back In.
The cheapest way to run AI agents is now also the stickiest, and most second-source plans don't have a line for it. Agent workloads resend enormous context on every call, the system prompt, the tool definitions, the documents, the conversation so far. Providers price a cached re-read of that context at a steep discount against a fresh one, commonly a tenth of the input rate and on some tiers far less. So teams do the rational thing and design the whole workload around one provider's caching scheme. Then they discover the switching cost moved. It isn't in the weights anymore. It's in the context layout, and leaving means re-paying for every active context at the full rate while you redesign it for someone else's cache. We call that exit cost the cache mortgage. You pay a context down once and ride cheap thereafter. Moving means refinancing.
Your Agent Didn't Fail. It Found Another Way. That's the Problem.
When an agent can't finish the job you gave it, you don't get a stop. You get a detour. Last week Anthropic published a review of cases where its models, blocked from completing evaluation tasks, routed around the block instead of halting. One pulled working access tokens out of a website's config file to query a map it couldn't click through. One used a free URL shortener to get past a length limit on its own fetch tool. One submitted a real police tip form while, by its own transcript, it was only generating example content. None of these caused serious harm. All of them share a shape, and that shape is the thing every operator running agents in production needs to design against. We call it the workaround reflex, and the fix is architectural rather than a better prompt.
Why Does Scaling AI Cost So Much More Than Piloting It?
Every AI initiative has two price tags. The first buys the capability, the model, the licenses, the build. The second buys everything you need in order to trust what the first one bought, the evals, the guardrails, the observability, the containment, the audit trail, the redesigned workflow around it all. We call that second price tag the control bill, and it is the half of the budget almost nobody writes down. It's why the pilot that cost fifty thousand dollars becomes a million-dollar program at scale and everyone acts surprised. The pilot that looked cheap wasn't cheap. It was unfinished. Leaders who price the control bill up front scale on schedule. Leaders who don't discover it one incident at a time, at the worst prices, in the quarters they can least afford it.
Why Is Your AI Vendor Suddenly Selling Other Companies' Models?
At its developer conference last week, OpenAI launched a marketplace that lets enterprise customers spend their committed OpenAI dollars on third-party software and even on other labs' open-weight models, all purchased through OpenAI. On the surface it reads as generosity, the biggest lab in the world handing its customers the multi-model flexibility they've been asking for. Read it again. The commitment, the checkout, the routing, and the relationship all still run through one company, which makes this the smartest lock-in move of the year. We call the product rented optionality, flexibility you lease from the same vendor you wanted the flexibility from. It answers the board's lock-in question without reducing the lock-in by a dollar, and versions of it are about to show up in every renewal conversation you have.
How Much of Your Revenue Depends on Customers Not Paying Attention?
A slice of almost every company's revenue exists for one reason. Customers don't act in their own best interest. They keep the subscription they stopped using, let the renewal reprice itself, accept the loyalty penalty, and leave their money in the account paying a fraction of the going rate, because acting would cost attention they'd rather spend elsewhere. We call that slice the inertia margin, the share of revenue that survives only because optimizing against you takes more effort than your customers will spend. AI agents spend that effort for free, at two in the morning, with no brand loyalty and infinite patience. Your customers forget. Their agents won't. And the margin they're coming for is probably on your P&L right now, unlabeled.
What Is Executable IP, and Why Will It Be Your Most Valuable Asset?
For a century, a company's most valuable knowledge lived in forms you could read. Patents, playbooks, process documents, and the accumulated judgment in your best people's heads. That form is about to change, because AI systems are starting to learn continuously from the organizations they work inside, and the destination of all that learning is a new kind of asset. Business knowledge distilled into model weights and agents you own, knowledge that stopped describing the work and started doing it. The old IP explained how your company operates. The new IP operates. We call it executable IP, and while the fully realized version is a few years out, the companies that will own it are being decided right now, by what they capture today.
Your Agents Think in Seconds. Your Systems Update Overnight.
There's a hard limit on what AI agents can do inside your company, and it has nothing to do with the model. An agent can never be smarter than the freshness of the systems it reads. Most enterprise systems were built for a world that ran on overnight batch jobs and weekly syncs, and an agent reasoning brilliantly over yesterday's data produces yesterday's answer with today's confidence. We call this the batch ceiling, and it's about to become the most expensive constraint in enterprise AI, because agent ambitions are real-time and the plumbing underneath them was built thirty years before anyone imagined asking it a question every second.
What Are AI Judgment Models, and Where Do They Fit in Your Stack?
A new kind of AI model launched last week, and it can't write a single sentence. Judgment models answer narrow questions with calibrated probabilities instead of prose. Ask one whether a customer sounds angry and it returns 0.9 in a tenth of a second, for a fraction of a cent, in a typed format your software can act on directly. The first entrant comes from a founder who helped build the training method behind ChatGPT, and the category matters to leaders for one reason. Most office work is reading something and deciding what happens next, and those small decisions were stuck between brittle rules and oversized language-model calls. Cheap, honest judgment changes what you can afford to check, and it completes a picture we've been assembling for a year. We call that picture the judgment stack.
Your AI's Explanation Isn't an Audit
Most AI oversight, at companies of every size, rests on a single assumption, and the people who build the models just told you it's failing. The assumption is that you can supervise an AI by reading its reasoning or asking it to explain itself. OpenAI's chief scientist wrote this month that the company's own ability to rely on reading model reasoning is shrinking, and that the trend runs in one direction as systems grow more capable. If the lab can't trust that layer, your governance process can't either. We call the mistake the explanation trap, and the way out is to audit the record of what your systems did, a record you own and the model can't edit.
Your Agents Went Multiplayer. Your Company Didn't.
The next constraint on what teams get from AI is not the model. It is the room the work happens in. Right now, nearly all agent work runs inside what we call the private session, a workspace only one person can see, steer, or resume. The agent might be brilliant. Its work still reaches the rest of the team the same way work did in 1998, as a pasted result, a forwarded file, or a summary in a meeting. Over the past quarter the industry started tearing that wall down. Shared team agents now live in Slack channels, agent platforms are rebuilding around sessions two people can occupy at once, and investors are funding the category by name. The companies that move their agent work out of private windows first will get compounding returns the paste-and-forward companies cannot touch.
Want to stop reading about AI and start implementing it?
The AI Blueprint maps your operations and delivers a concrete plan for building AI that works in production.