GlobalToken prices fall month after month, but many companies’ AI bills don’t drop at the same pace: the cause is almost always that they budget by license instead of by process.
You approve an AI tool for the sales team with a clear quote: a flat fee per seat, billed monthly. Six months later, the bill has gone up, even though the model provider just announced a price cut. No one lied to you: the license was simply never the real cost. The real cost includes CRM integration, the time someone spends reviewing what the model generates, the corrections needed when it gets something wrong, and the usage other departments have quietly started without IT or finance knowing about it. That’s what never shows up in the initial quote.
This article isn’t about whether AI is expensive or cheap in the abstract. It’s about a specific, measurable confusion: budgeting by the wrong unit. While token prices follow one trajectory, the cost of the processes you build on top of those tokens can follow a completely different one, moving in the opposite direction.
The short answer: why token price isn’t the cost of AI
The cost of AI in a business isn’t the token price or the monthly license fee: it’s what it costs to complete a business process from start to finish, including human oversight, corrections, and integration with the rest of the sales system. Two companies can pay the same price for the same model and end up with completely different real costs, because that cost depends on the process built around the model, not on the model itself. Budgeting by license means looking at the price tag; budgeting by process means looking at the full bill.
Token prices are falling. Why isn’t your bill falling too?
There’s a data point that often circulates as proof that “AI keeps getting cheaper.” According to the September 2026 Ramp AI Index, among Ramp’s corporate customers the effective price paid per million tokens fell 41%, to $0.68, down from a peak of $1.15 in March 2026. It’s a real data point, but with a very specific scope: it measures token and API spending for the customer base of a US expense-management platform, skewed toward startups and smaller tech companies. It isn’t a figure for the Spanish or European market, and it doesn’t include contracts, seats, or licenses.
The problem arises when you assume that this drop in unit price automatically translates into lower total spending. Gartner predicts the opposite for one specific type of AI use: inference cost per agentic workflow will increase more than fivefold by 2028, a phenomenon the firm calls “the inference paradox.” Gartner’s explanation is straightforward: each successive generation of AI capability requires more tokens, often more expensive ones, so product leaders can’t count on more efficient token economics alone to rationalize total cost.
The mechanism is easy to understand once you separate two types of use. A chatbot processes a request and delivers a response: one call, one cost. An agent, on the other hand, may need several rounds of chained reasoning—calling a tool, verifying a data point, drafting an output, reviewing its own response—before completing a single task. According to Gartner, using an agentic reasoning model can multiply inference cost by at least five compared to a basic chatbot interaction, and the multiplier grows the more demanding the workflow is. An assistant that drafts an email isn’t comparable, in cost, to an agent that researches a client, cross-references CRM data, drafts a proposal, and adjusts it after an initial review.
The mistake of budgeting by license
Most AI budgets that reach the executive team are built on a single number: the price per seat or the monthly subscription fee. It’s an understandable mistake—it’s the only figure the vendor makes easy to find—and it’s also the most common reason an AI project loses credibility with the CFO once the real spend starts showing up in other line items.
What the tool’s bill doesn’t show
A total cost of ownership (TCO) framework applied to any IT asset includes, beyond the purchase price, line items such as setup, training, maintenance, and the downtime tied to actual use, as the general total cost of ownership framework from Tractian lays out. It’s a generic framework, built for industrial and IT assets in general, not AI specifically, but the structure is useful for looking past the sticker price. Applied to an AI project, there are usually at least four line items missing from the initial quote:
- Integration: connecting the tool to the CRM, the ERP, or the proposal system isn’t free or instant, and usually requires hours from a technical role or an outside vendor.
- Human oversight: someone has to review what the model produces before it reaches a customer, especially for proposals, sales communications, or technical content.
- Correction: when the output isn’t usable, the time needed to rewrite or redo it doesn’t disappear—it just shifts to someone else on the team.
- Governance: defining who can use which data, with what permissions and under what review, carries a management cost that’s rarely accounted for in the initial project.
Why spending slips past IT
There’s a pattern that explains much of this cost overrun: the purchase decision no longer runs through a single department. According to the Capgemini Research Institute, 59% of generative AI spending now originates in business units, without direct IT involvement, and 68% of the organizations surveyed exceeded their generative AI budget. 98% of executives admit to having acquired technology without going through IT at least once. This is a survey of 1,000 senior executives across 14 countries, mostly at large organizations, so it shouldn’t be read as a statistic about Spanish SMEs; but the pattern it describes is recognizable at any company size: the sales team subscribes to an assistant for proposals, customer service subscribes to another one to handle inquiries, and no one adds up both bills until finance asks why software spending has grown without a formal decision behind it.
If your company doesn’t have any AI-powered process running yet and you want to avoid this pattern from the start, it helps to define the process before the tool: you can check fit before approving any budget. And if you already have standalone tools approved by different departments, the starting point is the same one we describe in AI for Sales Processes: Where to Start Without Buying Another Tool: map the process before adding one more tool.
How to define the real cost of an AI process (not a tool)
The shift in approach we propose at E2M is simple to state and demanding to apply: stop asking “how much does the license cost?” and start asking “how much does it cost to complete this process, start to finish, with AI involved?” The unit of measurement stops being the seat or the token and becomes the completed process transaction: a proposal sent, a customer inquiry resolved, a lead qualified and handed off for follow-up.
That unit requires adding up five components, not one:
- Model inference cost (the token or the API call).
- Integration and technical maintenance cost of the workflow.
- Human oversight time for each output generated.
- Correction time when the output isn’t usable.
- Governance cost: permissions, quality review, compliance.
A B2B services company using AI to generate first drafts of technical proposals shouldn’t ask how much the assistant’s subscription costs, but rather how much—in dollars and in hours—each proposal coming out of that process costs, compared to producing it without AI. If the generated draft needs half an hour of technical review before it can be sent, that half hour is a process cost, not a free side effect.
Swipe or scroll the table ↔ to see every column.
| Unit of measurement | What it measures | What it hides | Question to ask before approving budget |
|---|---|---|---|
| Price per million tokens | The model’s unit cost at a given moment | Says nothing about how many calls your process needs to complete | How many model calls does it take to complete one real task, not a test? |
| Price per license or seat per month | Cost of access to the tool | Integration, oversight, correction, and governance are left out | What additional cost does each person using this license generate, beyond the fee? |
| Cost per agentic workflow | Cost of a task with multiple chained reasoning steps | Can multiply with the number of reasoning rounds without the token price reflecting it | How many internal steps does the agent run to complete a single task? |
| Cost per complete business process | Everything above, plus oversight, correction, and governance | Nothing: it’s the unit that actually determines return | How much does this process cost start to finish, including human review? |
Why adoption is rising while returns aren’t keeping pace
The confusion between unit price and process cost also helps explain a gap that shows up in enterprise adoption data. According to the McKinsey global survey from August 2026 (1,719 participants, collected between May 4 and June 8, 2026), 37% of respondents attribute some impact on their organization’s EBIT to AI, a figure nearly unchanged from the previous year, while almost nine in ten organizations use AI regularly in at least one business function. The same survey shows a meaningful gap by scale: among organizations with $1 billion or more in revenue, 54% report having scaled AI company-wide, compared to roughly a third of smaller organizations. It’s a survey of executive perception, not an independent financial audit, but the pattern is consistent with what this article describes: almost everyone uses AI, very few have rigorously measured what it costs them and what it returns.
There’s a finding worth citing with all its caveats, not as a headline. A preliminary report from MIT NANDA, published in July 2025 and based on 52 interviews and 153 surveys collected at industry conferences—not peer-reviewed or representative of businesses as a whole—found that a very high proportion of the generative AI pilots analyzed hadn’t generated the return expected by whoever championed them. It isn’t a definitive figure, nor can it be extrapolated to “most companies”; it’s a preliminary finding with a small sample and a collection method skewed toward attendees of industry events. What it does offer, with those limitations made explicit, is a signal: scaling a pilot without first measuring its real cost per process is a concrete risk, not a theoretical hypothesis.
Measure before you scale: the minimum metrics before expanding a pilot
The principle we apply at E2M is the same for any commercial AI project: process first, then approved information sources, then clear ownership, then review, then metrics, and only then the decision to scale. Before expanding an AI pilot to more teams or more volume, there’s a minimum set of metrics that should be on the table, not in a rough estimate:
- Average cost per completed process transaction (not per license or per user).
- Human oversight time per output generated, in minutes, not in perception.
- Percentage of outputs that require substantial correction before they can be used.
- Who, outside IT or finance, is already paying for similar tools.
- What happens to cost if volume triples: does it grow linearly, or does it spike because of the type of workflow (agentic versus one-off use)?
If your company hasn’t defined these metrics yet for the pilot you have running, the risk isn’t just financial: it’s that the decision to scale or not gets made with the wrong unit of measurement. We go deeper into the criteria for knowing when it’s worth waiting in When an SME Should Not Automate Yet. And if you’re considering public funding for part of this spend, it’s worth being clear on what it covers and doesn’t cover before you apply: we cover that in Spain’s AI Voucher (Plan IA360): What We Know Before You Apply.
What E2M wouldn’t do
We wouldn’t budget an AI project by looking only at the price per token or per monthly seat: it’s the most visible and least representative part of the real cost. We wouldn’t approve scaling a pilot without first measuring the cost and human correction time for the complete process, because that’s precisely where the cost overrun hides before it reaches leadership as a surprise. And we wouldn’t promise that falling model prices will automatically translate into lower total spending: Gartner’s own data on agentic workflows shows the opposite can happen, and any budget that doesn’t account for that scenario starts from an assumption unsupported by the available evidence.
From insight to action
What would AI cost in your business process?
Start by identifying the use case, its dependencies and the evidence you need before making an investment decision.
Sources and limits of this guide
- Ramp, “September 2026 Ramp AI Index”. Shows a 41% drop in the effective price per million tokens among Ramp’s corporate customers between March and September 2026. Limit: it’s proprietary spending data from a self-selected population, predominantly American, skewed toward startups; it doesn’t represent the Spanish or European market.
- Gartner, “Gartner Predicts AI Inference Costs Per Agentic Workflow Will Increase More Than Fivefold Through 2028”. Shows Gartner’s reasoning and forecast on rising inference costs for agentic workflows. Limit: it’s a forecast from the consulting firm, with methodology not publicly detailed, not a closed retrospective measurement.
- Stanford HAI, “Artificial Intelligence Index Report 2025,” Chapter 1. Shows the drop in inference price for a fixed performance level (MMLU benchmark) between 2022 and 2024. Limit: it measures the price of a given model capability level, not the cost of running a complete business process; it shouldn’t be read as “AI has gotten cheaper for businesses.”
- McKinsey & Company, “The State of AI in 2026: On the Road to ROI”. Shows the gap between AI adoption (nearly universal) and self-reported EBIT impact (37%, flat year over year). Limit: a self-reported survey of executive perception, not an independent financial audit.
- Capgemini Research Institute, press release on “On-Demand Tech”. Shows that generative AI spending is mostly decided outside IT and that a majority of organizations exceed their AI budget. Limit: a survey of 1,000 senior executives at large organizations across 14 countries; it doesn’t represent Spanish SMEs.
- Tractian, “Total Cost of Ownership”. Provides the generic conceptual framework for total cost of ownership (beyond purchase price). Limit: it’s a vendor glossary entry about IT/industrial assets in general, not specific to AI; none of its numerical figures are used in this article.
- MIT NANDA, preliminary report on generative AI pilots (July 2025). Offers a signal that a high proportion of pilots fail to meet the return expected by whoever championed them. Explicit limit: a preliminary report, not peer-reviewed, based on 52 interviews and 153 surveys collected at industry conferences; not representative of businesses as a whole.