The consequence
We are fixing AI rather than redesigning the organisation it has exposed.
When execution becomes cheap and fast, the constraint moves to deciding, authorising, coordinating and proving.
Who can act?
Capability is not authority.
Who decides?
Decisions, not defaults.
Who is accountable?
A named person, not a system.
Can you prove it?
Evidence, not assertion.
Agents can act in seconds while authority, accountability and evidence still span documents, roles and systems designed for human-speed organisations.
The proposition
Don't redesign the organisation around agents. Design it around the work.
The job should not depend on whether the work is performed by a person, an AI agent, a service or software. The organisation should know what job is being done, who or what is performing it, under whose authority, for what value, and with what evidence.
The job is invariant. The participant is a governed selection: it still decides skills, controls, risk, cost and accountability.
See it proven
Ask the organisation.
An illustrative bank on synthetic data answers what an organisation chart cannot.
Who performs this customer job?
Follow the work into the organisation that performs it.
Who is accountable for it?
Where the accountability chain arrives, and where none can be established.
What authority permits this action?
What may be done, by whom, within what bounds.
What decision put this arrangement in place?
Apply a decision and see the organisation it produces.
What value does it claim?
Which value claims the work bears out, and which it does not.
What evidence establishes what happened?
Every step records what the bank will hold as evidence.
What changes if a different participant performs it?
Some consequences move, some hold, some the bank cannot establish.