Gartner Forecasts $2.7 Trillion in 2026 AI Spending: How Much Is New Budget vs. Rebranded IT
Gartner forecasts AI spending will rise 49.5% to $2.7 trillion in 2026, but much is relabeled IT. A breakdown of the agents, security, and platform lines, with lessons from Microsoft, Capital One, and SAP for building an AI budget.
Gartner forecasts that worldwide AI spending will rise 49.5% in 2026 to $2.7 trillion. The headline number is enormous, but Gartner’s own analyst cautions that much of the growth is relabeling rather than new money: existing technology purchases increasingly count as AI spending because products now include AI features. For CIOs, the more useful parts of the forecast are the category breakdowns, which show where budgets and risks are actually moving.
This article summarizes the forecast, explains the difference between new and rebranded AI spending, and draws lessons from Microsoft, Capital One, and SAP for building an AI budget that holds up.
The forecast in numbers
CIO.com’s Lynn Greiner reported on September 23 on Gartner’s updated forecast:
| Category | Forecast change |
|---|---|
| Total worldwide AI spending, 2026 | +49.5%, to $2.7 trillion |
| Total, 2027 | +36.2% |
| Agents and assistants (separated from other software) | +77.3% |
| AI software overall | +60.2% |
| AI infrastructure | +51.2% |
| AI security | Almost doubles |
| Development platforms | 39% growth, raised from 28% in May |
Gartner raised the development-platform forecast because enterprises and vendors are building more custom applications. These categories involve different teams and different risks, so they should be planned separately rather than averaged into one AI budget line.
Rebranding versus new budget
John-David Lovelock, Distinguished VP Analyst at Gartner, explained that CIOs received net-new AI budgets in 2024 and a little in 2025, but that more of today’s spending is rebranding than diversion. A laptop with a neural processing unit is now an “AI laptop”; an existing software contract becomes an AI purchase when the vendor adds AI features. Lovelock expects that by 2030, every technology dollar will be an AI dollar one way or another.
He also noted that hyperscalers have not diverted spending from cloud to fund AI. Amazon, Google, Microsoft, and Meta continue to build cloud capacity while carrying out an AI buildout he described as the largest infrastructure project humanity has undertaken.
For an individual organization, this has two implications:
- The $2.7 trillion is not a benchmark for your budget. It is a worldwide total that includes relabeled purchases.
- Diversion still happens inside IT teams. Even if macro spending is not being diverted, engineers may be moved from foundational work (for example, a data catalog) to AI features. That trade-off should be explicit.
Why the agents line matters most
The fastest-growing category, agents and assistants at 77.3%, is also the riskiest. An assistant that drafts text is low risk; an agent that issues refunds or changes records can cause real harm. Gartner’s forecast does not separate those, but internal policies must.
Microsoft’s changes show how buying patterns are shifting. CIO Dive reported on September 25 that Microsoft is reorganizing its Copilot app into three areas (Home for tasks in Word, Excel, and PowerPoint; Code built on GitHub technology; and Autopilot as a productivity assistant) and moving charging toward usage rather than seats. Microsoft has reported more than 30 million paid Microsoft 365 Copilot seats, including customers with more than 50,000 seats.
Usage-based pricing can help governance if someone monitors consumption, but it can also create surprises. An assistant running tasks overnight shows up as a spike in usage, not a fixed seat count. Setting spending caps before enabling new features by default is a simple safeguard.
The near-doubling of AI security spending reflects the same issue. Agents with access to company systems create new risks, and the most valuable security controls are those that can block an unwanted action, not only report it afterwards.
Capital One: the foundation comes first
CIO Dive’s September 28 profile of Capital One offers a useful counterpoint to fast deployment. Chief executive Richard Fairbank has described spending 14 years transforming the bank’s technology from the bottom of the stack up. Its agentic systems build on that foundation. Chat Concierge, a multi-agent car-buying assistant that helps customers with dealers, test drives, and digital purchases, has been in production for more than two years.
Capital One’s governance approach, as described in the article, is that agentic and generative AI applications should be managed, governed, secure, and standardized, both for regulatory needs and to create a dependable path to deployment.
The lesson is about sequencing. Agents depend on clean data, clear identity and access controls, and reliable platforms. Organizations without those foundations should invest in them before scaling agents. When a vendor or consultant presents Capital One as an example, it is worth asking which stage of that 14-year path they are actually offering.
SAP: linking agents to processes
Futurum’s Keith Kirkpatrick covered SAP’s Transformation Excellence Summit in Atlanta on September 22, where SAP updated Signavio, LeanIX, WalkMe, and Cloud ALM in its Business Transformation Management portfolio. The message was that agents should be deployed, governed, and monitored with knowledge of processes, ownership, risk, and adoption.
Each tool covers a different question:
| Tool | Question it answers |
|---|---|
| Signavio | How does the process actually run? |
| LeanIX | Which systems does the organization own? |
| WalkMe | Will employees actually use the new tool? |
| Cloud ALM | Can operations monitor what is running? |
The same article cites Futurum’s 2H 2026 Enterprise Software Decision Maker Survey (833 respondents, August), which found that expected and actual returns on recent software purchases were both around 14%, and that 41.5% said validated ROI studies would increase their confidence in budgets. Buyers are skeptical, and vendors are responding by tying AI to measurable processes.
A practical takeaway is to require every agent to be linked to a specific, documented process and owner. If an agent cannot be tied to a named process, it is difficult to govern or measure. Organizations should also choose one place where all agent actions are logged, whether agents come from SAP, Microsoft, or internal development.
Sovereignty and exit rights
Writing in CIO.com, Martin De Saulles notes that the 2026 Stanford HAI AI Index places open models only 3.3% behind frontier models. That makes open models a realistic option, but fine-tuning can create lock-in if model weights and training configurations cannot be exported. He also flags reliance on third-party fine-tuning help as a potential source of lock-in.
Three questions help turn “AI sovereignty” into concrete requirements:
- Where does inference run?
- Who owns fine-tuned models, prompts, and evaluation data?
- Which jurisdiction’s laws apply to logs and data?
Contracts for custom AI development should state that weights, prompts, evaluation sets, and traces belong to the customer and can be taken elsewhere.
Building an AI budget that holds up
- Split the budget into three lines: agents and assistants, AI security, and development platforms, each with a named owner.
- Make the agent owner someone who can deny actions, not only the person managing the vendor relationship.
- Monitor usage-based billing and set caps before enabling new features widely.
- Fund foundations first: data catalogs, identity, and observability.
- Require process links and ROI evidence for agent proposals.
- Protect high-risk systems: payroll, the general ledger, and regulated data should only be touched by agents under explicit, logged permissions.
Related coverage
See our analysis of Gartner’s finding that only 13% of CIOs report significant AI value, why only 4% of CIOs have redesigned roles for AI agents, the enterprise AI ROI gap, and CIO guardrails for agents.
Bottom line
Gartner’s $2.7 trillion forecast reflects both genuine investment and relabeling of existing technology. The categories to watch are agents and assistants (+77.3%), AI security (nearly doubling), and development platforms (39%). Capital One’s long foundation-building and SAP’s process-linked approach point to the same conclusion: AI spending produces value when it rests on solid data and governance, is tied to specific processes, and is measured against clear returns.
Sources
- CIO.com, Lynn Greiner, “AI spend will jump 49.5% in 2026, says Gartner” (September 23): https://www.cio.com/article/4225785/ai-spend-will-jump-49-5-in-2026-says-gartner.html
- CIO Dive, Microsoft Copilot enterprise changes (September 25): https://www.ciodive.com/news/Microsoft-copilot-enterprise-AI-tools/831423
- CIO Dive, Capital One agentic AI strategy (September 28): https://www.ciodive.com/news/capital-one-agentic-ai-strategy/831435
- Futurum, “SAP links AI agents to business processes and oversight” (September 24): https://futurumgroup.com/insights/sap-links-ai-agents-to-business-processes-and-oversight
- CIO.com, Martin De Saulles, “6 steps to weigh compromise and priorities for AI sovereignty” (September 28): https://www.cio.com/article/4226650/6-steps-to-weigh-compromise-and-priorities-for-ai-sovereignty.html