AI costs: in 2026, AI goes from pilot to bill
On October 21, at MEOCONNECT in Paris, MEOGROUP unveiled the 1st AI Cost Barometer before an audience of decision-makers from Procurement, Finance, and Information Systems. The first finding: 62% of companies do not compare their AI budget with their actual spending. Use cases are accelerating, while AI cost management is moving forward more slowly.
“My view is that AI, this year, is moving from pilot to invoice.” Steve Gordon, Partner at MASAÏ, MEOGROUP’s consulting division, opened the presentation of the Barometer this way. Four years after the arrival of consumer assistants, AI has moved beyond pilot projects. It now appears in budgets, contracts, and invoices.
Three figures from the Barometer summarize the gap between usage and management. 62% of companies do not compare their AI budget with their actual spending. 81% of AI projects in production do not have a quantified return on investment. 72% of companies have never assessed the exit cost of their main AI provider.
Asma Mhalla’s keynote, as a political scientist specializing in the geopolitics of AI, gave these figures their strategic interpretation. The roundtable hosted by Stéphane Soumier, business journalist, confronted them with on-the-ground realities. Around the table: Florence Baiget, Group Chief Procurement Officer at Veolia and member of the Executive Committee, Nabil El Khedri, Chief Financial Officer of Société des Grands Projets, and Grégoire Mentré, Performance & Transformation Director at Thales.
The AI bill is changing in nature
The first figure reflects a shift in the model. AI costs used to be based on per-user licenses. They now depend on consumption, and agents multiply it: where an assistant answers a question, an agent executes an entire process to handle it.
Florence Baiget summed it up with a Procurement reflex: “This is a concept that is completely familiar in Procurement: there is price and there is volume. We lowered the token price by 75%, but volumes with AI agents have exploded. In the end, we are rather looking at a 250-fold increase in price.”
Grégoire Mentré pinpoints the blind spot precisely: “Our project costs are something we are able to see. What we see much more difficultly is this notion of agent consumption.” At Thales, the response involves tight governance, regular monitoring, and spending caps set for the agents themselves.
Veolia has made the same choice of discipline. The group distinguishes between a tool for day-to-day productivity and a secure environment reserved for complex agents, whose costs differ significantly. It has also created token budgets, with caps, modeled on euro and carbon budgets. Such a budget assumes knowing the actual spending—the very spending that 62% of companies do not reconcile with their budget.
Value is visible, but its measurement still needs to be built
The decision-makers present agree on one point: AI generates gains. And yet, the Barometer shows that 81% of AI projects in production have no quantified return on investment.
Nabil El Khedri illustrated the difficulty of measurement: “Before, I used to do five tasks. Today, thanks to AI, I am going to do a sixth task that I did not even do before.”
And yet the results are tangible. SGP’s finance department built an interactive map of the costs of the Grand Paris Express, with its 68 stations and 200 km of lines. The project had been estimated at between €1 million and €1.5 million over six months. The team developed it in-house with an off-the-shelf AI assistant, for less than €10,000 in tool costs, and put it into service at the end of August. The Green Bond report, which used to cost €30,000, is now produced without any external budget. Nabil El Khedri draws one conviction from this: stakeholders will become more rigorous, and unprepared organizations will be the first exposed.
Grégoire Mentré advocates selectivity. He is keeping four use cases for next year, each with a productivity target, and holds regular steering committees to quickly stop what is not working.
Florence Baiget points to two items often missing from the calculation. First, data quality—a costly prerequisite and one that is rarely included in a project’s total cost. Then trade-offs: an AI tool justifies cutting subscriptions or modules that have become redundant, and those savings are part of the value calculation.
Dependency, the blind spot of AI cost management
The third figure is the most strategic: 72% of companies have never assessed what it would cost them to exit their main AI provider.
Asma Mhalla explained the significance of this figure. According to her, the major AI players have become geostrategic bottlenecks: a unilateral political decision can be enough to cut off access to a service. She invites executives to frame the question in economic terms: “Tomorrow, you no longer have access: how much does that cost you, how long are you autonomous, how long can you survive? These are business model questions, and all of a sudden, the technology question becomes an Executive Committee question.”
On the ground, the response starts with contracts. At Veolia, no tool is brought in without an approval process co-designed with the legal department, which includes data reversibility requirements. “Of course price negotiation is important, but data reversibility is often one of the most difficult issues to negotiate. But we do not let go,” emphasizes Florence Baiget. To do this, the group relies on a team of buyers specialized in IT, continuously trained.
Grégoire Mentré points out that an information system without American technology remains extremely difficult to build today. Thales therefore classifies its data by sensitivity level and chooses hosting accordingly, from internal servers for the most sensitive data to trusted cloud offerings. At Société des Grands Projets, a service interruption experienced this summer convinced Nabil El Khedri to test open-source models locally, to ensure the continuity of his applications. His rule can be summed up in one word: stay agile.
Three questions to ask before finalizing the 2027 budget
These three figures translate into three concrete questions for Procurement, Finance, and Information Systems departments.
- Do we know our actual AI spending, including consumption and integration, and do we reconcile it each quarter with the approved budget?
- For each use case in production, which value indicator are we tracking, and who is responsible for it?
- How much would it cost us to exit our main AI provider, and how long could we operate without it?
Position your organization against the panel
The Barometer goes beyond these three figures. These results highlight the importance of better managing spending, objectifying productivity gains, and anticipating technological dependencies.
The challenge now is to know how much AI costs, what it delivers, and what dependencies it creates.
If you would like to explore these results in greater depth or discuss how they apply to your organization, the MEOGROUP teams are available to continue the conversation.
MEOGROUP can present its full results directly to your teams, to position your organization against the panel, identify your blind spots, and prioritize your trade-offs.