There is a line item in your AI budget you haven’t given a name to yet.
You know it is there. You can feel it in the gap between what your AI provider promised and what your monthly invoice actually shows. You can see it in the gradient between“fractions of a cent per token” and “we are spending six figures a month on this thing.”
But you have never named it. So you have never targeted it. So you have never cut it.
Here is the name.
It is the recomputation tax. And on most enterprise AI deployments, it accounts for somewhere between fifty and seventy percent of total spend.
Every probabilistic AI system in production has the same constitutional problem. It cannot remember what it did yesterday. It cannot remember what it did an hour ago. It cannot remember what it did in the previous query in the same session unless you manually reconstruct the context every time.
So that is what it does.
Reconstructs the context. Every time. From scratch. From zero.
You are not paying for the answer. You are paying for the system to remind itself who it is talking to, what it has been asked before, what the company sells, what the policy is, what last quarter’s numbers were, what the current strategy is, before it generates anything new.
The output token is the visible cost. The reconstruction is the hidden one. And the reconstruction layer is between four and ten times larger than the output layer in most enterprise deployments.
That is where the seventy percent comes from.
Deterministic cognitive architecture eliminates this entire category of spend.
Not reduces. Eliminates.
When intelligence is pre-compiled into structured artefacts and stored in a deterministic registry, the system does not reconstruct context on every query. It retrieves the artefact. The retrieval cost is two orders of magnitude lower than the regeneration cost.
When state persists across sessions through deterministic ledgers, the system does not ask itself who it is talking to. It already knows.
When reasoning has been performed once and the output structured into a reusable form, the system does not reason through it again. It executes from structure.
Three eliminations. One cost curve. Down.
Take a real enterprise deployment. Anonymous on the client. Specific on the math.
Annual API spend: £2.4 million.
Annual reconstruction tax — context rebuilds, redundant reasoning chains, drift correction overhead — estimated at sixty-seven percent of total spend, or £1.6 million.
Cost of a deterministic cognitive layer placed in front of the foundation model, scoring intent, retrieving pre-compiled artefacts, routing to verb-based execution: £25,000 per month all-in, or £300,000 annually.
Net annual saving on API spend alone: £1.3 million.
That is before productivity gains from removing review cycles caused by probabilistic drift. That is before the elimination of human moderation overhead that exists solely to catch hallucinations. That is before what auditability does to your compliance posture.
The seventy percent number is not aspirational. It is conservative.
Why your foundation model provider won’t tell you this
They sell tokens. You buying fewer tokens is not their growth strategy.
The pricing model that exists today is engineered to make the recomputation tax invisible by pricing it per unit so small the unit cost feels negligible. Multiply small numbers by enormous volume and the bill stops being negligible. But the unit cost framing protects the revenue model.
This is not a moral observation. It is a structural one.
A foundation model provider is in the business of selling the maximum quantity of probabilistic generation per dollar of customer infrastructure. That is the business model. It works for them.
The question for you is whether it works for you. The honest answer in 2026 is that for structured enterprise tasks — internal cognition, compliance, planning, knowledge management, strategy execution — it does not.
The CTO conversation that’s coming
Every CTO is about to have a conversation with their CFO that goes roughly like this.
CFO: “Why is our AI bill triple last year’s projection?”
CTO: “Usage scaled.”
CFO: “Did the value scale with it?”
The honest answer in most enterprises is that value scaled linearly while cost scaled non-linearly. Because the architecture is constitutionally non-linear in cost. Every new use case adds reconstruction overhead that does not decay with familiarity.
The fix is not a different model. It is a different architecture sitting in front of the model. That is the deterministic cognitive layer. That is what removes seventy percent of the bill.
You are paying a tax that is invisible by design.
You are paying it monthly. You are paying it on every query. You are paying it because the architecture you adopted was built for language generation, not for enterprise cognition.
You can name it. You can measure it. You can cut it.
Or you can keep paying it.
The companies that figure this out in the next eighteen months will outspend their competitors on transformation by the margin they save on infrastructure.
The ones that don’t will keep funding probabilistic exploration of things their systems already knew the answer to.
