Publication date: Friday, 16 October 2026
Author: Malik Carter, Chief Intelligence Architect
Deverout and Associates | Strategic Intelligence for Transformation Leaders
AI is no longer arriving as a tool at the edge of the organisation. It is moving into the flow of work. The next competitive divide is whether leaders can make authority, identity, cost, and accountability visible while that flow accelerates.
The Leverage Weekly: Core Pillars — 60-second overview
The Signal This Week
The market is moving from individual AI assistance to distributed agentic work. Google reports organisations deploying thousands of custom agents across operations, knowledge work, customer service, finance, education, and infrastructure. Orange Spain is reported to be deploying more than 1,000 custom Gemini Enterprise agents. SOMPO is reported to have built more than 10,000 agents across 34,000 employees. DBS is reported to be deploying deterministic chains of 70 to 80 specialised agents for corporate credit memos while keeping humans in the loop. These are Google-reported customer examples, not independent benchmarks, but they show the direction of travel: the agent is becoming part of the operating system of work.
KPMG’s Q1 2026 AI Pulse analysis puts the control question plainly. It reports that 54% of organisations are actively deploying AI agents, compared with 12% in 2024. It also reports that 73% use agents across multiple functions, 53% route information and decisions between teams, and 63% require human validation of outputs. The conclusion is direct: governance cannot remain a policy document beside the work. It must be built into the work.
At the infrastructure level, the State of AI 2026 policy review reports selected sovereign AI programmes with approximately $138 billion in pledges toward compute, infrastructure, and domestic development. Yet it also warns that a domestic data centre does not itself confer control over the models inside it. Capacity, energy, researchers, model access, operational continuity, and actual usage matter together.
The common signal is this:
The control plane is no longer a technical layer around the work. It is becoming the work.
When Agents Become a Network
A single assistant can be treated as a bounded application. A network of agents creates a different management problem.
One agent may retrieve evidence. Another may transform it. A third may make a recommendation. A fourth may route the recommendation to a person or another system. If authority is unclear at any point, the organisation can lose the ability to answer basic questions: Who initiated the action? What evidence did the system use? Which policy applied? Who could intervene? What did the system cost? What happens when a supplier, model, data source, or jurisdiction changes?
Google’s examples show the upside of this networked model. Customer and employee workflows can be redesigned around specialised agents. But the operating capability depends on more than the model. It depends on identity, permissions, hand-offs, evaluation, observability, and a human owner who understands the outcome.
The more agents coordinate, the more important the chain becomes. Leaders should stop asking only whether an agent is accurate. They should ask whether the chain of authority is legible.
Sovereignty Is Continuity, Not a Building
The State of AI report separates a useful idea from a common shortcut. A domestic data centre may improve capacity or resilience, but it does not automatically provide control over models, supply chains, energy, software, or access conditions.
The report identifies approximately $138 billion in selected sovereign AI programme pledges, while CNAS tracks 184 government-backed AI projects in 67 countries outside the US and China with disclosed budgets of approximately $84 billion. These sets are different and should not be added together or treated as competing estimates. The strategic lesson is more important than the headline totals: sovereignty has multiple layers.
A credible sovereignty position asks:
- Can essential work continue if access to a model or cloud provider is interrupted?
- Which data, models, chips, energy, and skills must remain available locally?
- What capability can the organisation inspect, negotiate, replace, or rebuild?
- Is public compute being measured by actual hours delivered and useful outcomes, rather than only by announced capacity?
- Which communities and jurisdictions carry the risk if infrastructure is concentrated elsewhere?
Sovereignty is therefore a continuity strategy. It is the ability to keep important work moving and to retain bargaining power when the external environment changes.
Accountability Must Be Designed Into the Chain
KPMG reports that 63% of organisations now require human validation of AI-agent outputs, up from 22% in Q1 2025. The increase signals a change in organisational expectations. Human validation is not an admission that AI has failed. It is a recognition that consequential work requires a visible decision owner.
Validation must be specific. “A human is involved” is not enough. The organisation must define what the human checks, what evidence is available, what threshold triggers escalation, and what happens when the human disagrees with the system.
The practical distinction is between approval theatre and operational control. Approval theatre adds a person to the end of a chain without giving them enough context, time, authority, or evidence to intervene. Operational control makes intervention possible before the consequence becomes irreversible.
The Distributed Work Control Loop
| Control layer | Question to answer | Minimum operating evidence |
|---|---|---|
| Identity | Which person, agent, service, or vendor acted? | Unique identity, role, and timestamp |
| Authority | What was the actor allowed to do? | Permission boundary, policy, and escalation threshold |
| Evidence | What information supported the action? | Source record, model version, retrieval path, and confidence or uncertainty |
| Handoff | Who received the work next? | Traceable transfer, queue, or system event |
| Intervention | Who can pause, reverse, or override it? | Named owner, stop control, and response time |
| Economics | What did the chain consume? | Usage, review, rework, incident, and infrastructure cost |
| Learning | What changes after an error or exception? | Incident record, evaluation update, and operating adjustment |
This control loop should be tested at the workflow level, not only at the model level. A model can perform well in isolation while the overall chain still produces an unsafe or uneconomic outcome.
A 30-Day Control-Plane Readiness Sprint
- Days 1–5: Draw the chain. Choose one consequential workflow. Map every agent, person, system, data source, hand-off, and decision point.
- Days 6–10: Name authority. Record what each actor may read, recommend, change, or trigger. Assign a named human owner to the outcome.
- Days 11–15: Make evidence visible. Capture model version, sources, prompts or instructions, confidence signals, exceptions, and the reason for each escalation.
- Days 16–20: Test intervention. Pause an agent, revoke a permission, route an exception, and reverse a low-risk action. Measure the time and information required.
- Days 21–25: Measure the economics. Calculate usage, infrastructure, human review, rework, incidents, and the cost of delay. Compare the chain with the original process.
- Days 26–30: Decide the boundary. Scale, constrain, redesign, or stop the workflow. Record what must remain local, what may be outsourced, and what continuity plan applies if access changes.
The Bottom Line
AI agents are becoming a networked layer of organisational work. The organisations that benefit will not be those that simply deploy the most agents. They will be those that can see the chain clearly enough to govern it, improve it, and keep it running when conditions change.
If authority is invisible, speed becomes exposure. If authority is designed into the work, speed can become capability.
Continue the conversation and return to deuerout.com for the next issue, updated intelligence, and practical transformation frameworks.
Sources and attribution
- Google Cloud, “Welcome to Gemini at Work 2026,” 8 October 2026: source.
- State of AI, “Washington is asserting control over frontier AI,” 8 October 2026: source.
- KPMG, “As AI Agents Scale Across the Enterprise, Who Is Accountable for Their Decisions?”, 24 April 2026: source.
Editorial note: Google and KPMG figures are attributed organisation-reported findings, not independent universal benchmarks. State of AI programme totals use different populations and should not be combined as a single spending estimate.
