Why AI risk is outpacing value and how to fix that: Lessons from Gartner IT Symposium/Xpo 2026

Agents generate more web traffic than humans. They also contribute more new code. And while nearly every organization is on a journey to adopt AI, Gartner finds that only 58% are getting value (opens in new tab) from it. At the same time, news of escaping models and other incidents of AI misalignment have become common. Risk is outpacing value.

This was the common thread at Gartner IT Symposium/Xpo 2026. AI risk is growing faster than value, and the reason is governance. Based on Gartner research, most organizations do not have AI governance in place today. Those that do tend to bolt it on after the fact using policy binders, committee sign-offs and reviews that run long after the data has moved. 

That was already a strained system. AI broke it. Agents act on their own, at machine speed, and sometimes in ways no one designed. Standard perimeter defenses were not built for that. The durable fix is to shift security left and protect the data itself, where it lives. Protect sensitive values at the source, so that protection travels through every pipeline, model and downstream system. Use a method that maintains referential integrity so the team can still join, analyze and train models on that data. This doesn't prevent an agent from acting. But when it does, it acts on tokens instead of sensitive values, so the potential blast radius shrinks. Read on to learn how we do this at Capital One. 

AI investment isn't delivering expected returns

In the opening keynote (opens in new tab), Gartner reported that 86% of CIOs say AI risk is growing faster than the value it returns. Spend is not the problem. Organizations are paying roughly 18x more to get value from AI than they did two years ago. But, only 58% are seeing positive returns. The gap between investment in AI and return is widening.

Gartner put hard economics behind it. By 2030, AI vendors are forecast to spend $7 trillion on infrastructure while only earning $6.3 trillion. And the productivity gains are smaller than advertised. While vendors proclaim market productivity gains of 50%+ from their AI, Gartner puts that closer to 16%.

So what do we do with rising costs and risk outpacing value? We maximize value. We focus our effort on what’s working and scale those use cases, rather than adding new ones that introduce more risk to the mix. AT&T is a great example: the telecommunications giant cut AI costs by roughly 90% (opens in new tab) by managing usage and spending deliberately. 

The executive question has essentially shifted from how fast we can ship AI to whether we can govern it well enough to keep shipping. 

AI cost became a governance surface

Cost used to be a finance conversation. This year it became a governance one. Gartner expects the inference cost per agentic workflow to rise more than fivefold through 2028, because an agent can consume 5-30x the LLM tokens of a simple chatbot for the same task. Gartner calls this the inference paradox: costs per token falls, yet costs per task rises. 

Unmanaged, that math can break budgets. Uber (opens in new tab) exhausted its annual AI tooling budget in four months. Microsoft (opens in new tab) pulled most of its Claude Code licenses over cost. These were early indicators of the shift from tokenmaxxing to valuemaxxing. 

Gartner's forward view (opens in new tab) treats runaway spend as a security problem too: by 2030, 80% of organizations with public-facing AI will have faced a cost-exhaustion attack. AI cost observability is becoming a required control.

AI agents illuminate a governance gap

Governance is emerging as the way to balance AI value with risk. Yet Gartner found that only 21% of organizations have agentic governance in place today, even as 78% plan to deploy agentic AI within the next 12 months. Where agentic governance does exist, it’s often added as an afterthought that can be bypassed. In fact, 47% of CIOs admit they have bypassed their own governance process (opens in new tab) to push an urgent deployment out the door. Governance that gets skipped under pressure is just paperwork.

Agents didn’t invent the governance gap, they just raise the stakes because they act. They use real credentials, call tools and external systems, spawn subagents and execute tasks at machine speed. Deloitte warns (opens in new tab) that, left unmonitored and without central control, agents can make unseen mistakes, reveal sensitive information, and invite a cyberattack.

Overall, agents introduce two new risks:

• Shadow AI: Models and agents no one is tracking. From employees using their personal AI accounts for business tasks to system admins enabling 3rd-party AI features without consulting cyber and everything in between. You cannot govern what you do not observe. 

• Dark AI: Systems causing harm, regardless of whether anyone intended it. This threat also covers the “unknown unknowns” that come from agents behaving in ways no one designed. Frontier models escaping sandboxed environments and hacking external services are prime examples of what can happen with autonomous software that goes rogue.

Accountability is moving with the risk. Gartner expects (opens in new tab) 80% of the global Fortune 500 to designate a senior executive, often the CIO, as an AI custodian by 2030. The practical advice is to invest in an AI governance system of record, a centralized location detailing all the AI you run and who is accountable for each agent or app. 

Regulation is accelerating data security changes

If governance feels urgent this year, regulation is a big reason why. Last year, North American privacy fines passed $3.4 billion, more than the previous five years combined (opens in new tab). Mandatory frameworks like the EU AI Act are setting a global bar for high-risk models. Gartner actually points to regulatory change as a primary catalyst for governance review, reporting that 71% of security leaders reassess their data security approach when regulatory requirements change.

The clock extends past privacy. The National Institute of Standards and Technology (NIST) is beginning to deprecate traditional public-key cryptography like RSA and ECDSA by 2030, and plans to disallow it completely by 2035. Data protected by those algorithms is already on a countdown. 

The real shift: Protect the data at the source

Here is what I took away from the week. The governance that is failing is the governance that lives in documents and last-minute reviews – static policies, committee sign-offs, manual escalations. None of it keeps pace with systems that make decisions in milliseconds and can copy data across a dozen environments before anyone signs off.

Gartner's guidance was to embed runtime controls (opens in new tab) in the pipelines and platforms rather than relying on manual escalation. I would shift it further left. The most durable place to apply that control is the source. If you tokenize sensitive data the moment it is created or ingested, it travels protected through every downstream pipeline, model and agent. It also shrinks the exposure window to the moment of ingestion. Without it, exposure persists across every downstream copy and every later read. 

Protect data at the source so governance stops being a tax on speed. It becomes the thing that lets you keep moving. That is what we’ve done at Capital One. 

How Capital One built to this principle early

Capital One Databolt was built out of our own need to secure sensitive data at scale while keeping it useful. It's based on an internal tool that protects more than 900 internal apps and handles +5 billion operations a day on average, with sub-millisecond latency.

Securing data for AI and analytics only pays off if teams can still use it. Databolt tokenizes sensitive data at the source, enabling teams to join, analyze, and train models on it without exposing the raw sensitive values. A sensitive value, like a Social Security number or account number, is replaced with a non-sensitive, format-preserving token before it reaches downstream workflows or systems. Every downstream copy carries tokens and models are trained on tokens instead of sensitive values.

Utility survives because the tokens are format-preserving. A tokenized account number still has the same format as a real account number, and the same input always maps to the same token. That gives you referential integrity, so tokenized values can be joined across sources exactly like the original did, with no schema changes. Your analysts keep their joins. Your models keep their features. 

The accuracy numbers back this up. A PwC study found that data protected with Databolt’s tokenization preserved 99.7% of model accuracy compared with plaintext, while masking the same data retained only 50.6% of accuracy. 

Protection is half of governance. The other half is control. Databolt applies fine-grained access controls and field-level governance, so who can detokenize what is an explicit policy decision. Any model, agent or human without detokenization rights only ever sees tokens. And every request and operation is logged automatically, producing the data-access audit trail an AI governance system of record depends on.

None of this works as technology alone. Our approach aligns people, policy and technology. Governance experts sit inside engineering pods and shape the work as it happens, instead of reviewing it after the fact. This is what helped us take governance from a bottleneck into a facilitator of innovation. 

Conclusion

The message at Gartner IT Symposium/Xpo 2026 was clear. AI without governance is dangerous, and governance built for documents does not work for systems that act on their own at machine speeds. The fix is to govern where your data lives. Secure your sensitive data at the source, so every downstream copy, model and agent works with tokens instead of sensitive values.

When people, policy and technology are aligned and data is protected before it moves, governance is what lets your AI strategy go faster. We learned that early, at enterprise scale, and built those learnings into Databolt so that organizations are able to use sensitive data in AI and analytics workloads without exposing raw values. Interested in learning more? Talk to the team about testing tokenization on your own pipelines today.