The tether: why your most important AI system should be measured by what it can survive losing
Picture two systems doing the same job. This is a thought experiment, not a case study.
System One runs on rented cloud capacity. The model, the data store, the workflow engine, and the human approval step are all reached over the network, each from a different vendor, each on its own invoice. System Two was designed on the assumption that one day the network, the vendor, or the budget would be gone, and it was told in advance what to do about each.
On a good day you cannot tell them apart. On a bad day, when a quota is throttled, a vendor reprices, or finance cuts the line, System One stops and System Two degrades. Nobody measured the difference, because nobody asked the question that exposes it: how long is the tether?
The number that should bother every owner
In IBM's 2024 survey of AI leaders, 100% reported canceling or postponing at least one AI initiative over compute cost (IBM Institute for Business Value, 2024). Yes, 100% of those surveyed. Every one. And the same research found that cloud infrastructure already costs about twice as much as the AI model licenses themselves (IBM Institute for Business Value, 2024).
What that means for you: for every dollar you spend on the model that makes the decision, roughly two ride on everything the decision depends on. The expensive part of an AI system is not the intelligence. It is the tether.
Position A: flexibility is the point. Do not engineer for scarcity.
The strongest version of this view is not naive. Designing a system to keep working when a dependency disappears means duplicate capability, local fallbacks that lag the best models, and slower iteration. Meanwhile the economics favor renting: the cost of a standard AI query fell about 280 times in 18 months, from $20 to $0.07 per million units of text (Stanford HAI, AI Index Report 2025). If prices fall like that, why spend engineering effort on a disconnected day that may never come? Wait, and the tether gets cheaper by itself.
That is a rational position, and on most days it wins.
Position B: cost per unit is not cost per decision.
The counter-position starts from the 100%. Unit prices fell and budgets were still cut, because the bill that gets reviewed is not the price of a unit. It is the price of a decision: the model call plus the data retrieval, the workflow, the network, the standby capacity, and the approval step. Each of those is a place where price, availability, policy, or authority can change without your consent. Operators in high-stakes settings treat those as design inputs because they cannot negotiate them. Most businesses treat them as free until someone prices them.
The synthesis: keep the cloud, know the tether
Both camps are right about different things. Position A is right that rented, flexible capacity is usually the cheapest way to run. Position B is right that "usually" is not a design requirement. The resolution is an "and": use flexible capacity, and decide in advance, in writing, what each system does when a dependency goes away. Stop, degrade, or fail safe.
That is what "born in production" means. The failure behavior is specified before launch, not discovered during an outage or a budget review. Retrofitting it onto a system people already depend on is the retrofit tax in its sharpest form: the same work, done later, with less room to maneuver and a live system standing on it.
Where this discipline already exists, and the objection you are forming
The objection: this is fine for an organization with a safety culture and a dedicated engineering staff. I have a quarterly plan and a cloud contract.
Fair. But the reason some operators handle tethers well is not that they have more rules. It is that the questions are asked at design time. In my operating career I have led production platforms where a system did not get turned on until someone could say what it would do when its dependencies failed. Hospitals ask it of a monitoring system. Logistics companies ask it of a routing engine. Plants ask it of a control loop. None of them call it AI governance. They call it operating.
Not the same rigidity. The same honesty about what production requires.
What this means for leaders: three moves
1. Draw the tether map. For each AI system your business relies on, list every dependency it needs to make one decision: the model vendor, the data source, the network path, the compute, and the approval authority. For each, write down who owns it, what it costs, and what the system does when it is gone. Most teams find at least one dependency nobody owns.
2. Track cost per decision. Divide a system's monthly compute and infrastructure spend by the number of decisions it makes. Watch the trend, not a point estimate. It gives finance a number that survives a budget review, and it shows you the tether getting longer or costlier before someone cuts it.
3. Run the cord-cut drill. Once a quarter, sever one tether on purpose, in a test environment first, and record whether the system fails safe, degrades, or stops, and how long recovery takes. A failure mode you have rehearsed is a design decision. One you have not is a surprise.
The CFO perspective
An AI system you cannot price per decision and have never run through a cord-cut is a forecast, not an asset. Capital committed to an initiative that stops when a vendor changes its terms is not durable capital. It is an expense that has not been recognized yet. The question for the next budget review is not "what does the model cost?" It is "what is our cost per decision, which way is it trending, and what is stranded if our cheapest dependency disappears?"
The window
IDC projects 1.3 billion AI agents in operation by 2028 (IDC, cited in Microsoft's CIO Playbook for AI, 2025). That is a projection, not a count, but the direction is clear. Every one of them will have a tether. The businesses that map theirs now will choose its length. The rest will find out in a budget review.
If you want a second set of eyes: send me the top three dependencies of your most important AI system and I will mark where I would expect the first cut to land. No cost, and I will do it regardless of fit.
