The world will spend $2.67 trillion on AI in 2026. To protect that AI: about 0.1% of it.

One dollar of protection for every nine hundred and forty

In September, Gartner published how much the world is expected to spend on artificial intelligence in 2026: $2.67 trillion, 49.5% more than the year before. Three weeks earlier, the same firm had released a forecast that drew far less attention: the market for securing AI, which covers security tools for AI applications, usage control, governance and gateways. For 2026, $2.8 billion.

I did the division, and the math is mine, not Gartner's. Protecting AI amounts to roughly 0.1% of what is spent on it. About $1 for every $940.

This comparison has a flaw, and I would rather say so up front. More than half of AI spending is infrastructure (servers, chips, data centers), and that does not call for the same kind of protection as an application. So I took infrastructure out of the calculation to see what changed. Protecting AI still comes to less than 0.3% of the rest. The order of magnitude does not move.

What the 95% doesn't tell you

In May, Vicente Conde wrote in his newsletter about the MIT study that has become a mandatory line in every AI meeting: despite $30 to $40 billion invested by companies in generative AI, 95% of organizations saw no measurable return. His argument is that the problem lies in the approach, not the model. He lists what sets apart the minority that gets it right: data ready before the use case, an outcome metric from the first week, and processes redesigned together with the people who run them.

I need to be honest here. MIT does not name security as a cause of these failures, and I will not force that reading. My point is a different one. The study describes the path of projects that succeed, and that path runs through leaving the pilot, entering real systems and touching customer data. That is where the protection nobody put in the budget starts to be missed.

Where the bill shows up later

Every year IBM publishes, with the Ponemon Institute, the "Cost of a Data Breach". In the 2025 edition, covering 600 organizations, 63% of the companies that suffered a breach had no AI governance policy or were still writing one. Among those that had incidents involving AI models or applications, 97% lacked proper access controls for those tools. And unauthorized AI use by employees themselves, so-called shadow AI, showed up in 20% of breaches and added about $670,000 to the average cost.

The 2026 edition came out in July and went further. One in four malicious breaches now involves AI on the attacker's side, and those incidents cost on average around $6 million, above the overall average of $4.99 million.

And in Brazil? According to the 2025 survey, a breach cost on average R$ 7.19 million, and 87% of Brazilian organizations said they had no AI governance policy.

Eighty-seven percent.

Companies know the risk. They just haven't changed the process

The World Economic Forum gives a measure of the gap. In the "Global Cybersecurity Outlook 2025", 66% of organizations expected AI to have the greatest impact on security in the year ahead, and only 37% had any process to assess the security of an AI tool before putting it to use.

In the 2026 edition, with 804 respondents in 92 countries, the share that assesses beforehand rose to 64%. That is real progress. But 36% still have no process at all, only 40% review these tools periodically, and 24% assess once and never look again. The most telling figure is the gap between groups: at companies with high cyber resilience, 83% assess AI before using it. At those with insufficient resilience, 39%.

Accenture arrived at a number that, for me, sums up the first-week problem. In the "State of Cybersecurity Resilience 2025", with 2,286 executives from 17 countries, only 28% of organizations build security into AI projects from the start. For 77%, basic data and AI security practices are missing.

The detail that should interest the CFO

I understand the financial instinct to wait for AI to prove its return before spending on protecting it. Brazilian banks, which operate under heavy regulation, run the numbers differently. The Febraban Banking Technology Survey 2026, conducted with Deloitte, projects R$ 50.4 billion in technology investment by the sector this year, with close to 10% (about R$ 5 billion) going to cybersecurity. For AI, the projection is around R$ 3 billion.

The same survey shows that 84% of banks rank generative AI among their priorities, but only 28% consider themselves at medium or high maturity. High priority with low maturity is a combination that calls for caution.

There is also the other side of the ledger. IBM's 2025 report showed that in Brazil, companies making extensive use of AI and automation in their own security had an average breach cost of R$ 6.48 million. Those that did not came in at R$ 8.78 million. AI applied well to defense saves money; AI without defense becomes an expense.

What changes when security comes in during week one

1. Protection goes into the AI budget. The security line sits in the project spreadsheet, next to data and infrastructure, instead of waiting for next year's security team allocation.

2. Assess before, and keep assessing. At go-live, of course, but also with every change of model, data or permissions. That is the distance between the 64% who assess and the 40% who review with some regularity.

3. Start with the inventory. Before protecting new AI, go after the AI already running without authorization. That is where the shadow AI breaches IBM measures every year come from.

The question for your next budget meeting

When the next AI project comes up for approval, open the spreadsheet and look for the line that pays for its protection. If you can't find it, ask in which month it comes in and which budget it will come out of.

$1 for every $940. It looks like enough until the day someone needs that line.