日本語

RESEARCH NOTE · AGENTIC AI

The barrier is the
operating model.

Read side by side, eight recent reports on agentic AI reach the same diagnosis. The models are capable enough. What stalls deployment is the organisation around them: who may decide what, how actions are overseen, and whether the evidence exists to show it.

KODA Sōken · Research note · Tokyo, 16 October 2026

01 — WHAT THE REPORTS CONVERGE ON

Capability is moving. The organisation is not.

The capability curve is not in doubt. The Stanford AI Index reports organisational adoption of AI at 88%, and agents rising from 12% to roughly 66% task success on the OSWorld benchmark, while still failing about one attempt in three.[3] Expectations have moved with it: in BCG’s survey, around 90% of CEOs believe AI agents will enable measurable ROI in 2026, and CEOs have committed more than 30% of their 2026 AI investment to agents.[1]

Governance has not kept pace. Deloitte finds that only 21% of organisations report a mature governance model for agentic AI, and defines maturity concretely: clear boundaries on which decisions agents may take alone and which need human approval, real-time monitoring of agent behaviour, and audit trails of the full chain of agent actions.[2] The AI Index reaches the same conclusion from the other side: responsible AI “is not keeping pace with AI capability”, and documented AI incidents rose to 362, from 233 in 2024.[3]

Several reports name the operating model directly. BCG writes that AI requires “creating an operating model that incorporates machines and humans”, and that leading companies are about twice as likely as followers to deploy agents end to end across a workstream or process.[1] IBM states that “running AI in the enterprise requires a new operating model”, and that as organisations move from a handful of agents to thousands, the challenge becomes keeping them “governed and auditable in near real time”.[5] McKinsey places the main constraint in weak data foundations rather than in the technology, and recommends governance embedded in systems, lineage, auditability and redesigned operating models.[4]

The practitioners say the same. In Anthropic’s survey of technical leaders, the most cited challenges are integration with existing systems (46%), data access and quality (42%) and change management (39%).[8] Accenture observes that “too many pilots don’t scale because AI, business and platform strategies remain siloed”.[6] And Microsoft finds that culture, manager support and talent practices account for more than twice the reported AI impact of individual factors (67% against 32%), while only 26% of AI users say their leadership is clearly aligned on AI.[7]

A note on method. These are different surveys, with different samples, questions and definitions. We quote each figure against its own source and do not combine them into a single statistic. Where the reports agree, it is on the shape of the problem, not on a number.

02 — SIX DIMENSIONS

Where the operating model breaks.

Grouped together, the findings fall into six dimensions. None of them is about model quality.

01 — Decision rights and accountability

Which decisions an agent may take alone, which require a person, and who answers for the outcome. BCG reports that leading organisations expect AI to change governance and decision rights; Deloitte defines mature agent governance first by exactly these boundaries.

02 — Governance and oversight

Senior leadership shaping governance rather than delegating it to technical teams alone (Deloitte), and agents kept "governed and auditable in near real time" as their number grows (IBM).

03 — Data and architecture foundations

McKinsey locates the main constraint in weak data foundations, and recommends governance embedded in systems, with lineage and auditability. Anthropic’s respondents rank data access and quality second among challenges.

04 — Platform and integration

Integration with existing systems is the most-cited challenge in Anthropic’s survey; Accenture finds pilots fail to scale when AI, business and platform strategies remain siloed.

05 — Monitoring and safeguards

Real-time monitoring and audit trails of the full chain of agent actions (Deloitte), against a backdrop of documented AI incidents rising from 233 to 362 in a year (Stanford AI Index).

06 — Workforce readiness

Microsoft finds organisational conditions account for more than twice the reported AI impact of individual factors; Deloitte names insufficient worker skills as the biggest barrier to integrating AI into workflows; BCG’s leading CEOs invest more in upskilling.

03 — OUR RESEARCH VIEW

Governance has to be enforced, not documented.

Most organisations answer these findings with documents: a responsible-AI policy, a RACI chart, a review board. Those are necessary, but an agent acting at three in the morning does not read them. The dimensions the reports identify (decision rights, oversight, data boundaries, monitoring) are exactly the places an agent runtime touches on every step. That is where we think they belong.

Our working hypothesis is that an operating model for agents is only as real as the mechanisms that enforce it at run time. Six mechanisms, each documented in the Kurahi runtime we develop and study, correspond to the six dimensions:

Policy as code

Decision rights are written as a signed, versioned policy that the runtime loads and enforces at the points where an agent acts: tool dispatch, model requests, outbound messages. Each decision records the clause that produced it. Policy documentation ↗

Evidence per action

Every model turn, tool call and permission decision is appended to a hash-chained event log, closed with a signed manifest and verifiable offline by a third party. Evidence documentation ↗

Permissioned data access

Each agent carries its own read, write and execute rules; writes and execution are refused unless a rule allows them, so the data boundary is a property of the system rather than of a prompt. Agent permissions documentation ↗

Human approval in existing channels

When a policy requires a person, the request is forwarded to a chat channel the approver already uses and answered there, and the permission decision is recorded with the run. Approvals documentation ↗

Provable continuity

An agent’s role, memory, permissions and pending work survive a restart or a model change, and that survival is checked against signed invariants instead of assumed. Continuity documentation ↗

Sovereign deployment

The runtime runs on infrastructure the organisation controls, with models treated as interchangeable components, so governance does not depend on any single vendor’s console. Local models documentation ↗

Six operating-model dimensions mapped to runtime mechanismsOPERATING-MODEL DIMENSIONENFORCED BY THE RUNTIME AS01Decision rights and accountabilityPolicy as codeSigned; every decision logs its clause02Governance and oversightEvidence per actionHash-chained log of each turn and tool call03Data and architecture foundationsPermissioned data accessPer-agent rules; write and exec deny by default04Platform and integrationSovereign deploymentOn infrastructure the organisation controls05Monitoring and safeguardsProvable continuityVerified after restarts; budgets, kill switch06Workforce readinessApproval in existing channelsPeople approve where they already workSix operating-model dimensions mapped to runtime mechanismsDIMENSION → RUNTIME MECHANISM01Decision rights and accountability→ Policy as codeSigned; every decision logs its clause02Governance and oversight→ Evidence per actionHash-chained log of each turn and tool call03Data and architecture foundations→ Permissioned data accessPer-agent rules; write and exec deny by default04Platform and integration→ Sovereign deploymentOn infrastructure the organisation controls05Monitoring and safeguards→ Provable continuityVerified after restarts; budgets, kill switch06Workforce readiness→ Approval in existing channelsPeople approve where they already work
Figure 1 — The six operating-model dimensions the 2026 reports converge on, and the runtime mechanism that enforces each one rather than documenting it. KODA Sōken synthesis.

The test we apply is simple. If an auditor asks why an agent took a particular action, the answer should come from a record the system produced at the time, citing the policy clause and the person who approved it, rather than from a reconstruction after the fact. When governance is enforced in this way, the operating model stops being a slide and becomes a property that can be checked.

This is a research position, not a finding. The reports above establish the problem; they do not test our proposed answer, and we have not yet published an external evaluation of it.

04 — IF YOU ARE WORKING ON THIS

If your organisation is designing decision rights, oversight or evidence for agents, we would be glad to compare notes, including where this framing does not fit your case. Write to research@kodasoken.com, or see how to work with us.

REFERENCES

  1. BCG AI Radar 2026: As AI Investments Surge, CEOs Take the Lead. Boston Consulting Group, 15 January 2026. Survey of 2,360 executives, including 640 CEOs, in 16 markets. www.bcg.com/publications/2026/as-ai-investments-surge-ceos-take-the-lead
  2. State of AI in the Enterprise (2026 edition). Deloitte, 2026. Survey of 3,235 business and IT leaders in 24 countries, August–September 2025. Governance figures as reported in “Business and IT leaders report AI agents are scaling faster than their guardrails”, Deloitte Insights, 24 April 2026. www.deloitte.com/global/en/issues/generative-ai/state-of-ai-in-enterprise.html · www.deloitte.com/global/en/insights/topics/emerging-technologies/ai-agents-scaling-faster.html
  3. The 2026 AI Index Report. Stanford Institute for Human-Centered Artificial Intelligence (HAI), April 2026. Top takeaways. hai.stanford.edu/ai-index/2026-ai-index-report
  4. Building the foundations for agentic AI at scale. McKinsey & Company, 2 April 2026. www.mckinsey.com/capabilities/mckinsey-technology/our-insights/building-the-foundations-for-agentic-ai-at-scale
  5. Think 2026: IBM Delivers the Blueprint for the AI Operating Model as the AI Divide Widens. IBM Newsroom, 5 May 2026. newsroom.ibm.com/2026-05-05-think-2026-ibm-delivers-the-blueprint-for-the-ai-operating-model-as-the-ai-divide-widens
  6. The New Rules of Platform Strategy in the Age of Agentic AI. Accenture, research report, 18 December 2025. www.accenture.com/en/insights/strategy/new-rules-platform-strategy-agentic-ai
  7. 2026 Work Trend Index Annual Report: Agents, human agency, and the opportunity for every organization. Microsoft WorkLab, 5 May 2026. Survey of 20,000 knowledge workers who use AI, in 10 markets. www.microsoft.com/en-us/worklab/work-trend-index/agents-human-agency-and-the-opportunity-for-every-organization
  8. The 2026 State of AI Agents Report. Anthropic, with research firm Material. Survey of more than 500 technical leaders in the United States. resources.anthropic.com/2026-state-of-ai-agents

Figures are quoted as published by each organisation. Accenture’s report is dated December 2025 and is included as its most recent publication on agentic platform strategy. Sources accessed 15 October 2026.

All publications →日本語 →
The Barrier Is the Operating Model — Research Note | KodaSōken