AI workflows look cheap today. CIOs need a test for when they don't.

A common thread has run through quite a few conversations I have had lately. The AI business cases people are signing off rest on a quiet assumption: that the price of the underlying models will stay where it is or keep falling. It is an understandable assumption. Per-token prices have dropped sharply for two years. But it is the wrong basis for a five-year commitment. Today's AI economics are subsidised. Providers are pricing below the cost of serving to capture market share, funded by huge amounts of investor capital. So the question I keep putting to anyone approving a critical workflow is not what it costs today. It is whether they are confident it will still make commercial sense on the day the discount ends.

That reframing turns AI adoption from a procurement decision into a commercial risk assessment. The suggestion I offer in these conversations is a way to run that assessment.

Separate the bill, then stress-test the risky half

Start by splitting AI spend into two unrelated halves. Seat licences for staff are fixed and predictable. They scale with headcount and are easy to forecast. The workflows you build on the API are different. They are variable, they scale with usage, and that is where the commercial risk sits. A workflow that costs a fraction of a penny per transaction today can become a material line item once volume grows and pricing normalises, or when a new model is introduced that utilises more tokens.

The instinct is to forecast that variable cost by extending a trend. Resist it. Every available trend misleads. Borrowing the gentle, once-a-decade price rises of a mature SaaS suite anchors you to the wrong model. Those products have near-zero marginal cost and raise prices through lock-in. AI carries a real compute cost on every query and is being sold below it. Extrapolating the last two years of AI price falls is just as unsafe. The steep falls have been concentrated in older models, while the frontier reasoning models that critical workflows tend to need have held firm. And "thinking" models burn far more tokens per task, so bills can climb even as headline rates drop.

So do not forecast a single number. Model a range. Three scenarios bracket the plausible future well enough for a board paper. The first is continued commoditisation, where competition and efficiency keep effective costs flat or falling. The second is subsidy withdrawal, where prices step up by perhaps a third to a half as providers are pushed towards sustainable economics. That pressure is now arriving. The Financial Times reports that OpenAI and Anthropic are both moving towards public listings, and that xAI has folded into SpaceX, which listed in June 2026. For the frontier providers, the discipline of public markets is shifting from prospect to fact, and with it the pressure to turn a profit. The third scenario is consumption-led escalation, where unit prices ease but agentic and reasoning workloads grow faster, so total spend rises anyway.

The feasibility test is then simple. An AI-enabled workflow is only worth building if it stays value-positive under the adverse scenario. If it works only while today's subsidy persists, you are betting the business case on someone else's funding round.

The second axis: can you actually leave?

Cost is only half of commercial feasibility. The other half is exit. Competition disciplines pricing only for work that is genuinely contestable, meaning new or portable workloads you could move tomorrow. For workloads welded to a single provider, the ceiling on what you can be charged is not the rival's price. It is the rival's price plus your switching cost. A provider can keep acquisition pricing keen to win new business while quietly raising effective prices on the customers who cannot leave. That is ordinary enterprise-software economics, and public-market pressure will only sharpen it.

This gives the framework its second axis. Plot each workflow by how critical it is against how costly it would be to reverse. Where a workflow is both critical and hard to exit, you are in the danger zone. Invest deliberately in reversibility: a model-agnostic architecture, a credible fallback provider, and retained human capability. Where a workflow is non-critical, accept the lock-in consciously and spend the saved effort elsewhere. The aim is not to make everything reversible. Reversibility has a real cost, and over-hedging destroys value. The aim is to choose, per workflow, which traps are worth a premium to avoid.

Engineer the exit before you need it

Two practical realities make this urgent. First, models drift. A new version can change output formats, reasoning and refusal behaviour, and quietly break a workflow tuned to the old one. Providers also retire old versions on their own schedule, not yours. Second, leaving is far harder than it looks after a few years. The staff who once did the work manually have moved on. The surrounding process has been redesigned around machine throughput. The cost to rebuild can exceed the escalated bill you were trying to escape.

One control addresses both problems at once. A maintained evaluation harness is a representative set of inputs with defined acceptance criteria. It turns a forced model upgrade from a gamble into a measured event, and the same harness tells you whether a rival's model clears your bar. The discipline that protects you from drift is the one that lowers your switching cost. Build the harness once and it does both jobs.

Make it a gate, not an afterthought

None of this works retrofitted. Commercial risk, lock-in and drift control belong at the feasibility gate, alongside the conventional value case. Estimate the value per unit of work. Estimate the cost. Stress-test it against the adverse pricing scenario. Place it on the criticality-versus-reversibility grid. Only then decide whether to adopt, adopt with controls, defer or reject.

None of this is an argument against AI or against lock-in. What I suggest is a way to decide, workflow by workflow, which commercial traps are worth risking and what to pay to avoid the rest. Do it while the pricing is still generous and you keep control of your AI economics. Leave it until the subsidy ends and the provider takes that control instead.