Blog Posts
AI systems process every classification level simultaneously. Polygraf wrote a guide to building a data classification policy that actually governs your data.
In December 2025, Uber gave Claude to 5,000 engineers. Four months later in April 2026, the whole of the annual AI budget was spent and the CTO said he was "back to the drawing board". The rollout worked. That's the point. Enterprise AI is not cheap to scale because it doesn't fail, but because its production economics is nothing like its pilot economics. Here's the line by line.
There is one moment that repeats in every enterprise AI program, and it always sounds the same in the retro. The pilot was a success. Adoption was high. The demo impressed the board. Then the deployment went to production, the bill came and it had nothing to do with the number in the business case. The most cited example of 2026 is Uber – not because Uber did something reckless, but because it did everything right and blew through its budget in a third of the year. When your successful deployments are the ones breaking your budget, the problem is not execution. It is that the cost model was wrong from the start.
This is a line-item breakdown of why. Not "AI is expensive" as a feeling, but the cost structure: where the money goes, which lines scale linearly with usage and which scale super-linearly, and – most importantly – the costs that don't appear in the pilot budget and dominate the production one. We write this as a security company, so we will be honest about where security and governance sit in the stack: they are real costs and they are also the costs that, done right, prevent the much larger ones.
Inference and infrastructure are the visible costs. The one that appears as a single big catastrophic number is a breach or a compliance failure. Polygraf's AI Risk Calculator models that exposure across breach, regulatory and litigation risk – the line item you can't see until it hits.
The most important thing to know about the economics of enterprise AI is that pilots and production are in different universes. It's not that production is "more" – it's that the cost structure changes shape.
The majority of enterprise AI budgets were based on per-seat or per-subscription logic – the world of fixed-price SaaS where you multiply a license fee by headcount and you have a flat line finance can plan around. API-based AI broke that. Cost now scales with consumption – how many times and at what depth a model is called – which is architecturally volatile and forecast-resistant. FinOps teams spent a decade learning to govern reserved cloud instances; they are now asked to forecast token consumption for workloads that didn't exist a fiscal year ago. That is why the share of FinOps teams managing AI spend went to almost universal and why AI cost management became the most in-demand finance skill of 2026.
Here is the full cost stack for a production enterprise AI system. Click any line to see what it is driven by and why it is the way it is. The percentages are an example of a typical API based deployment – your mix will change with architecture – but the order is very consistent.
Inference is the cost of running the model every time someone uses it. It is not a build cost you pay once, it is an operating cost you pay forever and it scales directly with usage. Estimates have always been in the range of 80–90% of ongoing AI spend. And the counterintuitive part is that per-token prices have dropped dramatically – one analysis of 2.4 billion enterprise API calls found blended cost dropping about 67% year over year – but bills went up. The price per unit went down; the number of units went up.
A chatbot query is one API call. An agentic task is a sequence: the agent plans, loads context, calls tools, checks its work, and retries when it fails. Gartner's March 2026 analysis puts this at 5 to 30× the tokens per task; EY measured a single customer-service interaction going from about $0.04 in 2023 to $1.20 in 2026 – about 30-fold – as orchestrated multi-tool workflows took over linear exchanges. That's why the Uber budget disappeared even though the per-token price was going down: the organization moved from chatbot-shaped use to agent-shaped use and every task quietly became an order of magnitude more expensive. Budgets set in 2025 before agents were at production scale were structurally dead on arrival.
The one architectural decision that dominates the inference line. The same analysis of 2.4 billion API calls showed that organizations routing a tiered model architecture – small models for simple tasks, frontier models only when necessary – paid a median blended cost of $2.31 per million tokens, while organizations routing everything to frontier models paid $18.40. That is ~87% gap from a single decision that is usually made at the beginning of a deployment and never revisited. Sending every request to the biggest model is the AI-era equivalent of leaving every cloud instance at its largest size.
Inference is the king, but the others are where projects leak and the ones that skip them are the ones that fail.
GPUs (owned or rented), serving stack, autoscaling, networking, storage. For self-hosted models it is the cost-per-GPU-hour and how many requests you serve before latency breaks. AI-first products are spending 40–50% of their revenue on hosting and compute vs 15–20% COGS for traditional SaaS.
Ingestion, cleaning, embedding, vector storage and the retrieval overhead of RAG (which adds tokens to every query as retrieved context is fed back into the prompt) – retrieval is working as expected, the cost model just doesn't account for it.
Model-risk management, monitoring, access control, DLP and compliance frameworks. Running on going governance and safety costs $30k–$100k+ per year per system. The line most cut – and the most cut that correlates with failure.
The eval harness that tells you when a model silently degrades, plus drift detection and quality monitoring. Without it you can't see decay until users do. Most stalled projects skipped this step.
The biggest hidden cost: engineers, ML talent and the change-management work of putting AI into real workflows. MIT discovered the tipping point between the 5% who win and the 95% who lose is not model quality – it's integration and learning.
Models are deprecated, prompts change, integrations break, and providers change pricing or terms mid-cycle. Every migration is re-testing, re-tuning and re-validation. The AI stack is moving faster than any software stack ever.
Two of the biggest enterprise AI costs never show up on the planning line items because neither looks like a cost until it hits. Both are where a security company has something to say.
The MIT NANDA GenAI Divide study examined 300+ initiatives and found a brutal funnel: 80% of organizations are AI-ing, 60% are evaluating enterprise tools, 20% are running pilots and only 5% are getting to production with measurable impact (95% are getting nothing back against $30–40B total spend). Every dead pilot is real money spent on infrastructure, integration and people that is getting nothing back. Data-privacy and security are among the most-cited reasons why pilots are killed before production (i.e. governance is not only a cost, but a driver of whether the rest of the spend is recovered).
Every line so far is a flow – a bill you pay over time. Breach and compliance exposure is a stock, a one-off that can be bigger than a year of inference. A single bad dataset, a leaked prompt with regulated data, an agent that took an unauthorised action – they don't show up monthly. They show up once, with legal, regulatory, remediation and reputational costs on top of them, and often a regulator's fine schedule. This is the line the AI Risk Calculator above is for, to make visible before it hits, because it is the one that turns a good cost program into a bad year.
The costs that blow enterprise AI budgets are not on the pricing page. They are the multiplier you did not model, the pilot that never shipped, and the breach that shows up as a single number. The teams that scale AI affordably are not spending less. They are making the invisible costs visible early enough to govern them.
— Polygraf AI, on the economics of scaling AIThe organizations that scale AI without the budget shock are not lucky and they are not spending less per unit. They do a particular set of things.
They route by task, not by default. Tiered model architectures – cheap models for easy work, frontier models only where they earn it – are the ~87% lever. They model consumption, not seats. Finance forecasts token volume and sets spend alerts and caps, treating AI like cloud, not like SaaS. They instrument before they scale. Eval harnesses and monitoring catch silent degradation before it becomes rework. They govern from day one. Because the pilots that die are often killed by security and privacy gaps and because a breach is the most expensive line of all, they treat governance as cost-avoidance, not overhead. And they make every cost visible early – the recurring finding across 2026's cost analyses is that the winners simply saw the hidden lines before the board did.
Polygraf is on the governance-and-security line and its economic argument is that the right control on that line reduces the two most dangerous costs in the whole stack. It is on-premise, CPU-only, no GPU needed and sub-100ms latency, which is a direct concern of the cost conversation: the enforcement layer doesn't add a GPU bill or a per-token toll of its own and it doesn't sit in a cloud vendor's meter. It is inspecting and controlling what data is flowing in and out of AI at the boundary, which is exactly the gap that gets pilots killed on privacy grounds and the gap that becomes a breach.
Cost wise: Polygraf AI is a bounded, mostly-fixed line item that lowers the probability of the two unbounded ones – the pilot that never ships because it couldn't pass security review and the breach that arrives as one single catastrophic number – for teams moving AI from pilot to production in a regulated environment that is not overhead on top of the AI budget but the line that protects it.
Polygraf AI is on-premise and CPU-only - no GPU bill, no per-token toll, no cloud meter - checking and controlling data at the AI edge. The bounded cost that reduces your two unbounded ones: killed pilots and breaches. Sub-100ms, zero data egress.
At Polygraf, we envision a future where AI augments human capabilities without compromising safety, privacy, or ethical standards. Trust in our commitment to building this future with you.
© 2026 Polygraf AI. All rights reserved.
Your download will start now.
Please provide information below and we will send you a link to download the white paper.
Data Privacy 
Data Provenance 
Content Generation