THE ENTERPRISE AI SIGNAL · BANKING SERIES 2 OF 5 · 2026-11-10
By Roman Bodnarchuk, Founder, WisdomTwin, Inc.
+ advisers versus executors, classifying on what an agent may do and touch, and why accountability never moves to the model
Cold open. Most banks still approve AI use cases one committee meeting at a time. That approach slows low-risk work to a crawl and still misses the agent that can write into a payments system.
Signal 1. Two kinds of coworker, two control sets
McKinsey places agentic coworkers on an autonomy spectrum. Advisers gather evidence and recommend, and a person approves before anything executes. Executors act inside an approved scope, and people review the outcomes.
Its illustration pairs a credit memo adviser with read-only scoped identity, source validation and a human approval gate. An executor closing low-risk AML alerts gets a write scope with a blast-radius limit, evidence before action and a tested rollback path.
The examples matter less than the method. Controls should come from a standard library chosen by classification, not be designed fresh for every project.
Signal 2. Classify on two axes: do and touch
Autonomy is half the classification. Sensitivity is the other half. An adviser summarizing public filings and an adviser drafting from borrower financials carry different risk.
Score every agent on what it may do (read, recommend, write a draft, execute, delegate) and what it may touch (public data, internal data, customer data, regulated systems of record such as payments, credit decisioning and AML case management).
The cell an agent lands in sets its controls automatically. A change of scope means a new classification and a fresh approval, never a quiet permission upgrade.
SOURCE: McKinsey & Company, 2026-10-26. Grid levels in the companion toolkit are 10XAI.News's own build on that method.
Signal 3. The law follows the work, not the worker
There is no separate agentic AI rulebook, and McKinsey's advice is not to wait for one. Fair lending, Bank Secrecy Act and AML standards, and consumer protection rules apply whoever, or whatever, performs the work.
In Canada, E-23 expects institutions to manage model risk in proportion to each model's risk rating. A do-and-touch classification gives that rating a defensible basis for agents.
Decisions that matter to supervisors need a decision chain: the data used, the policy applied, the action taken and the person who approved it.
EXECUTIVE NOTE. If you cannot say what an agent may do and what it may touch, it is not ready to do anything.
Control Signals
▲ 3 controls per role in McKinsey's illustration. Advisers: scoped read-only identity, source validation, approval gate. Executors: write and blast-radius limit, evidence before action, tested rollback. SOURCE: McKinsey
■ 172 days from this issue until OSFI E-23 takes effect (2027-05-01). SOURCE: OSFI
● 4 elements in a defensible decision chain: data used, policy applied, action taken, human approver. SOURCE: McKinsey
The Enterprise Move
Place every live agent on a five-by-four do-and-touch grid this month.
Write one standard control set per zone and retire case-by-case approval for everything outside payments, credit and AML.
Require a decision record from every executor before it goes live.
Go further
WisdomTwin, Inc. AI Judgment Twins that turn the judgment trapped in critical enterprise roles into governed, auditable agents. https://www.wisdomtwin.ai
Watch on YouTube: https://www.youtube.com/@WisdomTwin
Forwarded this? Get The Enterprise AI Signal every Tuesday: www.10xai.news/subscribe
Know a CRO, CIO or head of model risk who should read this? Forward it. That is how this briefing grows.