Every business case I am shown for this technology has the same shape. Here is what the work costs today, here is what it costs with the tool, and here is the difference.
The arithmetic is usually correct. What it omits is that all three of the numbers on the right hand side depend on things the business does not control, and in several cases on things nobody currently controls.
Is AI in business sustainable is a question that gets answered as though it were about the technology. The capability is not the fragile part. The commercial arrangements around it are, and there are three dependencies in every case I see, none of which appear anywhere in the model.
01Is AI in business sustainable, or is your case resting on somebody else's decisions?
A price somebody else sets, capacity somebody else builds and a model somebody else can withdraw.
That is not an argument against using any of it. It is an argument for knowing which of your operations would stop working if any one of the three moved, because at the moment most businesses could not tell you. The saving has been booked, the process has been redesigned around it, and the dependency has been treated as though it were as stable as electricity.
The useful frame is the one you would apply to any other single-supplier arrangement. If a supplier provided fifteen per cent of your delivered value, on annually renewable terms, at a price they had recently trebled for other customers, with the right to discontinue the specific product you use, you would have a plan. Here, most businesses do not, because it does not feel like a supplier relationship. It feels like software.
02The first dependency: the price is not the price
Current pricing is being set in the middle of a competition for adoption, which is a specific and temporary condition.
There is a long pattern in technology of capabilities being sold below their eventual sustainable price while market position is being established, and of that changing once positions are settled. I am not going to predict which direction this particular market moves, because anyone claiming to know is guessing. What I will say is that a plan whose value disappears at three times the current unit cost is a plan with an undisclosed assumption in it.
The test takes five minutes. Take your actual usage, multiply the unit cost by three, and re-run the case. In most of the businesses I do this with, the answer is reassuring: the saving is large, the usage is modest, and the case survives comfortably. That is worth knowing. In a minority, particularly where something is being run at high volume across every transaction, the case only works at today's price, and the founder has been unaware that they built a margin on a promotional rate.
03The second dependency: this runs on physical things
Compute is not abstract. It is buildings, chips, water and a great deal of electricity, and all of those are constrained in ways software has not been for two decades.
The practical consequence for a mid-sized business is not that you will be unable to buy capacity. It is that the pricing and availability of capacity now respond to physical constraints, including energy costs, grid connections and construction timetables. That is a different risk profile from the one everybody in technology grew up with, where capacity was effectively infinite and cheaper every year.
There is a second consequence, and it arrives through your customers rather than your suppliers. Larger organisations increasingly carry reporting obligations on emissions across their supply chain, and the energy intensity of what you run on their behalf is a question that has started appearing in procurement documents. If you sell to businesses with those obligations, the honest position is that you will be asked, and that not knowing is a poorer answer than a large number.
A capability you rent, at a price set during a landgrab, running on physical capacity nobody can quickly expand, is not an advantage yet. It is an arrangement.
04The third dependency: the model you built on will be withdrawn
This is the one founders find most surprising and it is the most concrete of the three.
Models are deprecated. The version your process was tuned against gets replaced, the replacement behaves differently, and the differences are frequently subtle rather than obvious: it is slightly more verbose, it refuses a category of request it used to handle, it formats output differently, it is better at one thing and worse at another. Anything downstream that depended on the old behaviour now needs revisiting, on a timetable set by the vendor.
For a demonstration this is an inconvenience. For a process running unattended across thousands of transactions, it is an outage waiting for a date. And the businesses that feel this most acutely are the ones that did the most careful work, because a process finely tuned to one model's quirks is precisely the process that breaks when the quirks change.
05Is this not the same argument people made about the cloud?
It is very nearly the same argument, and the cloud turned out fine, so this objection deserves a proper answer rather than a deflection.
The parallels are real. Vendor lock-in, unpredictable pricing, loss of control and dependency on somebody else's infrastructure were all raised, loudly, and the businesses that refused to move on those grounds were mostly wrong. Most of them ended up paying more to run worse systems themselves.
Two differences matter. The first is that cloud pricing was, from early on, a margin business rather than a customer acquisition exercise, which made it far more predictable. The second is that cloud infrastructure was substitutable in a way models are not yet: a virtual machine is a virtual machine, and the discipline of portability was well understood and widely practised.
But the honest reading of the cloud experience supports the same conclusion I am arguing for, which is not abstinence. The businesses that did well out of cloud were the ones that adopted enthusiastically and kept their data portable, avoided the proprietary services that were hardest to leave unless the benefit was overwhelming, and knew what their exit would cost. That is exactly the posture worth taking here.
06How do you build so that a change is survivable?
Five things, none of which slow you down noticeably.
Own your evaluations. A set of real cases with known good answers, which you can run against any model. This is the single most valuable asset you will build, it costs a day, and it converts changing vendors from a judgement call into a measurement.
Own your data and your definitions. The prompts, the examples, the rules about what a good output looks like. If those live only inside a vendor's product, you have rented your own knowledge back.
Keep the manual path alive where it matters. For anything that touches a customer commitment, know what happens if the capability is unavailable for a week. In most cases the answer should be slower and more expensive, not stopped.
Price at three times. Make it a standing rule in any business case. If it only works at today's cost, say so explicitly in the paper rather than discovering it later.
Measure value in outcomes, not usage. Hours saved, cost per delivered unit, cycle time. Businesses that measure adoption end up defending a number that means nothing. Businesses that measure outcomes can tell instantly whether a price change matters.
07The version of this that matters
The capability is not going away and neither is the value. The arrangements around it are unsettled, and treating an unsettled arrangement as though it were infrastructure is how you end up with a margin that belongs to somebody else.
So adopt, quickly, and build as though your supplier will change. Own the evaluations, own the data, know the manual fallback, and check the case at three times the price before you sign anything. That is a fortnight of discipline, and it is the difference between having built a capability and having rented one you cannot leave.
