Key takeaways
- The constraint in 2026 is not model capability — it is governance, ownership, and evidence.
- Spending and ambition are massive; scaling and production readiness are rare. The gap between them is where projects die.
- Every failure mode in the data — cost overruns, stalled pilots, rollbacks, audit failures — maps to a missing control: identity, approval gates, stop conditions, or an evidence trail.
- Regulators are converging on the same expectations: agents must be identifiable, their actions attributable, and their stop controls tested.
The one-sentence summary of 2026
The world is spending record amounts on agentic AI, most of it is stuck between pilot and production, and the reason is governance — not model quality. The eleven stats below back that sentence. Where a figure comes from us rather than published research, we say so.
The 11 stats
1. More than 40% of agentic AI projects will be canceled by 2027
Gartner’s projection is the most-quoted number in the category, and the reasons matter more than the percentage: escalating costs, unclear business value, and inadequate risk controls. Note what is not on the list — model quality. Agents are canceled because nobody tied them to an accountable owner and a measurable outcome, or because governance review killed them before production. This is the failure mode Moring exists to fix, and it is why we anchor every engagement to one named workflow with an owner and a metric.
2. Global agentic AI spend is roughly $50B — and growing
KPMG estimates about $50 billion of global agentic AI spend in 2025. Combined with Gartner’s estimate of $2.5 trillion in total AI spend for 2026, the money is demonstrably there. The bottleneck is not budget.
3. Financial services agentic AI: $7.78B in 2026, heading to $43.5B by 2031
Mordor Intelligence (January 2026) sizes agentic AI in financial services at $7.78B in 2026 with a 41% CAGR to $43.5B by 2031. BFSI is the largest vertical, driven by exactly the workflows Moring targets: KYC and onboarding evidence, fraud, claims, compliance. Regulated industries are not AI laggards — they are the biggest spenders, and the most exposed when governance is missing.
4. Only 15% of enterprises have scaled multi-agent adoption
Deloitte’s August 2026 survey of 501 US leaders found just 15% have scaled multi-agent AI. Nearly everyone is piloting; almost no one has made it operational at enterprise scale. The gap between pilots and production is not a capability gap — it is a governance gap.
5. Only 16% say their business processes are ready for agents
The companion Deloitte finding: 16% describe their business processes as prepared for agentic AI. An agent dropped into an ungoverned process inherits the process’s ambiguity — unclear ownership, undocumented handoffs, no defined “completed.” This is why Moring starts from the workflow, not the model.
6. 60% of agentic AI decisions sit with CIOs and CTOs
BCG’s finding on decision rights: 60% of agentic tech decisions are made by CIOs/CTOs rather than business owners. That concentration speeds up tooling decisions and slows down process decisions — and it is why every Moring account maps two to four buyer routes, including the workflow owner who lives with the queue size and SLA, not just the technology buyer.
7. 70% of AI transformation is people and process change
BCG’s wider transformation research keeps landing on the same split: roughly 70% of the effort is people and process, not technology. Governance design — who approves what, who is accountable, what evidence exists — is process work. Buying another tool does not touch it. Forward-deployed engineering does.
8. Companies are rolling back AI agents at scale
Sinch’s The AI Production Paradox report (May 2026, independent survey of 2,527 senior decision-makers across 10 countries) found that 74% of enterprises have rolled back or shut down a live AI agent after deployment due to governance failure. Leading causes: customer data exposure (31%), hallucination or brand risk (22%), and inability to diagnose what went wrong (16%). Notably, the rollback rate is highest — 81% — among organizations with the most mature guardrails: better monitoring catches more failures. The operational lesson is the point: teams that cannot stop an agent cleanly, list what it did, and roll back specific actions end up removing the whole agent instead of governing it.
9. US banking regulators have agents on the agenda, not yet in scope
SR 11-7 was superseded by SR 26-2 (April 2026). Agentic AI is not yet in scope — but the Fed has signaled it plans to issue an RFI, which means banks that instrument governance now will answer the RFI with evidence rather than with a project. We track this in detail in our SR 26-2 explainer.
10. DORA makes agent activity part of operational resilience
The EU’s Digital Operational Resilience Act has applied since January 2025, and it regulates the systems agents act in, even though it does not name AI. An agent that changes records or moves payments is an ICT-supported function: it needs logging, risk assessment, tested stop controls, and incident evidence. We map the requirements in DORA and AI agents.
11. Oversight must live outside the agent’s execution loop
Forrester’s framing (Leslie Joseph) is the clearest one-line description of the category forming: agent oversight has to live outside the agent’s own execution loop — not as logging the agent emits about itself, and not as trust in its self-restraint. That is precisely the architecture of an AI control plane: identity, policy gates, approvals, and evidence enforced by infrastructure the agent cannot rewrite. We define the category in what an AI control plane is.
What the numbers add up to
Read together, the eleven stats describe one situation:
- The money is committed (stats 2, 3).
- The failure rate is known and high (stats 1, 8).
- Almost nobody has crossed to production (stats 4, 5).
- The decision-makers and the effort are in technology and process, not models (stats 6, 7).
- Regulators are converging on identical expectations: identifiable agents, attributable actions, tested stops, auditable evidence (stats 9, 10, 11).
The gap between spending and scaling is not a technology gap. It is the absence of a control layer — and that layer is what we build. If these numbers describe your situation, start with a workflow diagnostic: one named workflow, one accountable owner, one measurable outcome, and a governance path that survives the review that kills most pilots. Our risk and control framework page shows the six control layers examiners and auditors test first, and the how it works page shows how an engagement runs from discovery to handover.
A note on how we treat numbers
We separate proposed, modeled, measured, and realized outcomes in everything we publish, including our own customer work. The figures above are published by the named research firms. We apply the same standard to ourselves — no anonymous customer statistics, no marketplace claims without evidence.
Sources
- Gartner, agentic AI project cancellation projection, 2026 — reported widely; see Gartner’s press releases on agentic AI.
- KPMG, global agentic AI spend estimate (~$50B, 2025), KPMG research, 2026.
- Mordor Intelligence, Agentic AI in Financial Services market report ($7.78B 2026 → $43.5B 2031, 41% CAGR), January 2026.
- Deloitte, survey of 501 US leaders on multi-agent AI (15% scaled; 16% processes prepared), August 2026.
- BCG, agentic AI decision rights (60% CIO/CTO) and transformation split (70% people/process), 2026.
- Sinch, The AI Production Paradox (74% rollback rate; 81% among organizations with mature guardrails), May 2026 — press release. Survey of 2,527 senior decision-makers across 10 countries.
- Federal Reserve, SR 26-2 (April 2026), superseding SR 11-7.
- EU Regulation 2022/2554 (DORA), applicable January 2025 — EUR-Lex.
- Forrester, Leslie Joseph, on oversight outside the agent’s execution loop, 2026.
Frequently asked questions
What percentage of agentic AI projects get canceled?
Gartner projects that more than 40% of agentic AI projects will be canceled by 2027, citing escalating costs, unclear business value, and inadequate risk controls. The cancellations cluster around pilots that never defined an accountable owner, a measurable outcome, or a governance path to production.
What share of enterprises have scaled multi-agent AI?
Deloitte's August 2026 survey of 501 US leaders found only 15% have scaled multi-agent AI adoption. A companion finding is just as telling: only 16% say their business processes are prepared for agentic AI. The bottleneck is process readiness and governance, not model capability.
Who decides on agentic AI purchases in enterprises?
BCG found that 60% of agentic AI technology decisions sit with CIOs and CTOs, not business unit heads. And BCG's wider finding applies: about 70% of AI transformation effort is people and process change, which is why governance design — who approves, who is accountable, what evidence exists — matters more than model selection.
Why do AI agents fail in production?
Production failures cluster on governance, not capability: agents acting with shared credentials, no record of who approved what, no tested way to stop them, and no evidence trail an auditor can accept. The fix is a control layer outside the agent — identity, policy gates, human approvals, and tamper-evident records — what Forrester calls oversight living outside the agent's execution loop.
What is the state of agent governance in 2026?
The pattern across 2026 research: spending is enormous (about $50B globally on agentic AI in 2025), cancellation risk is high (Gartner's 40%+ by 2027), scaling is rare (Deloitte's 15%), and the constraint is governance. Regulators are moving — SR 26-2 in US banking, DORA in the EU, the EU AI Act. Organizations that instrument governance now are the ones that scale.
Governing Autonomous AI Agents in Production
Moring provides the runtime control layer, identity verification, human-in-the-loop approvals, and audit records needed to deploy production-grade AI agents safely.
About the author
Balaji Nagaraj KumarChief AI Officer, Moring AI, Moring AISweden
Leads platform and delivery engineering at Moring AI, turning each engagement into a substrate the customer’s own team operates at handoff. His background spans developer-platform leadership at Volvo Cars and engineering at CERN. He writes about governance for agents that take actions in production systems.
LinkedIn profile