AI policy is becoming product constraint.

What the model is allowed to do is now a licence-to-operate question. The control plane is the product.

Written executive summary

4 min listen · 8 min read

For Board Preparation

Policy has left the legal appendix. McKinsey’s 2026 trust survey puts security and risk as the main brake on scaling agents, and only about 30% of organisations sit at higher maturity across strategy, governance, and agentic controls (McKinsey). For the board, that is a licence-to-operate issue: multi-step agents that can move money, change records, or speak to customers need an enterprise control plane — identity, tool access, logging, evaluation, kill-switch — before more capital is released. Capital already committed last cycle (BCG) now meets a constraint the CIO cannot waive. Residual risk on agents belongs next to cyber on the committee agenda, not under ‘innovation updates’.

For ELT Attention

The ELT question is no longer which model. It is which workflows are allowed to act, under whose identity, and what residual risk the line will carry. Ownership of the control plane — platform versus risk versus the P&L — is the operating-model decision. Production paths that cannot show supervision should not receive the next increment of capital.

Must-know AI headlines

Enterprise control planes for multi-step agents are becoming the buyer’s language for ‘trusted AI’.
Vendor demos without identity, tool-gating, and a kill-switch will fail the same way unlogged admin access would.
(McKinsey)
EU AI Act obligations and US frontier-testing norms continue to set extra-territorial buyer expectations.
APAC product roadmaps that ignore those packs will stall in European procurement and in insurer questionnaires.
(EU)

Audio briefing

4 min listen

Three numbers

~30%

Organisations at higher maturity across strategy, governance, and agentic controls

McKinsey

2/3

Organisations citing security and risk as the top barrier to scaling agents

McKinsey

1.7%

AI spend still planned as a share of revenue — capital is not waiting on policy

BCG

Weekly triage

Board prep

ELT attention

Delegate

From the Library

Deep dive · Foundational · ~40 min

NIST AI Risk Management Framework 1.0

Govern, Map, Measure, Manage. The shared language for residual risk.

Must-know AI headlines

1
McKinsey’s 2026 trust survey: security and risk are the top brake on agent scale; only about 30% are mature across strategy, governance, and agentic controls.
This is the evidence pack for holding capital against a control-plane bar rather than against another model bake-off.
McKinsey — State of AI trust 2026
2
EU AI Act high-risk documentary duties continue to propagate through European buyers and D&O / cyber insurers.
Product teams that treat Brussels as out-of-scope will still meet the Act in contracts.
EU AI Act (Regulation 2024/1689)
3
US frontier-testing practice (AISI) is the de-facto language of ‘how was this model evaluated’ in enterprise questionnaires.
A vendor that cannot point to independent evaluation is asking the CEO to carry undocumented residual risk.
NIST AISI
4
Data and model access across Japan, ASEAN, and China remain the binding constraint on a single regional AI product.
Policy is now a product constraint in APAC as well: the same agent cannot be assumed to run everywhere.
METI AI policy (Japan)

Watch list

OECD AI Principles remain the diplomatic backbone behind national rule-making.
Still the cleanest common reference when the board asks which standard we claim.
OECD AI Principles
BCG’s capital numbers from the Radar have not been withdrawn — spend is still planned up, not down.
Policy is constraining use, not reversing the budget. That combination raises residual risk if unsupervised.
BCG AI Radar 2026

Asia Pacific

Singapore

Reference governance, now a product input

IMDA and PDPC materials are no longer only a compliance narrative. Product and risk can use the Model Framework as the checklist for what an agent is allowed to do in the regional hub — testing, human oversight, data, and incident handling — then mark where Japan, ASEAN, and China must diverge.

IMDA GenAI Model Framework

Japan

Guidance with procurement teeth

Japan’s promotional statute still sits on METI and sector guidance that public buyers and large corporates treat as binding. An agent that is fine in Singapore may fail a Japanese transparency or training-data expectation. Treat Japan as a product SKU, not a language pack.

METI AI policy

ASEAN

Fragmented permission to act

Cross-border inference and sector licences still decide whether an agent may act in a given ASEAN market. The control plane has to know which jurisdiction it is in — not only which model it is calling.

OECD AI Principles

China

Separate stack, separate residual risk

China model access, algorithm filing, and content rules remain a separate product and a separate residual-risk note. A regional control plane that cannot see the China stack is not a regional control plane.

CAC Generative AI measures