When an agent answers incorrectly, you correct a sentence.
When it acts wrongly, however, you must remedy the consequences of its actions in the real world.
When an AI system quotes a company’s price incorrectly, you can update the text. If that same system makes a purchase on a customer’s behalf using the incorrect price, the problem is no longer confined to inaccurate information. Money has been spent, a commitment has been made and other people’s behaviour has been affected.
An agent may mistake someone for a company executive.
This is an identity error.
But if the agent makes a payment on that person’s instructions, an identity error becomes a financial action.
An agent may misunderstand the scope of a service.
This is a representation error.
But if it sends the customer an offer that states the wrong scope of service, it creates a commercial commitment.
An agent may fail to recognise when human approval is required.
This is an authorisation error.
But if the agent publishes the content, the organisation’s reputation and legal liability are affected.
The central problem in the age of agents is not simply that machines make mistakes. People, organisations and software make mistakes too. The distinguishing risk is that some agent systems can carry a single error at high speed across a wide sphere of impact and through multiple tools.
A human may send an incorrect email to one person; with the necessary access and automation, an agent can personalise and send the same error to many people.
An employee can use the wrong price in a single document; an agent with necessary links can spread this price across multiple languages, catalogs, structured data, bid template and sales flow.
A human may edit the wrong file; an agent with broad write access can propagate that error consistently through a large production chain.
The danger here is not that AI is always bad or careless.
An agent can, in fact, do the wrong thing with remarkable order, consistency and apparent success.
The wrong price can be translated seamlessly in six languages.
Unauthorised message can be written highly professional.
An unsuitable provider may still appear to be supported by strong evidence.
Unreal success can be presented convincingly on a beautiful dashboard.
When error is coupled with technical quality, it becomes difficult to spot.
Therefore, errors made at GBO are only:
"The model answered incorrectly."
We can't handle it at the level.
The error can start in any of the following layers:
The organisation has published false or incomplete information.
Identities are not connected correctly.
Service limits are unclear.
Authority is confused with technical access.
The agent has broadened the user’s intent.
No limits have been passed on to sub-agents.
The measurement system has rewarded misbehaviour.
The demand for a human stop only stopped the main agent.
The rollback plan covered only technical files and overlooked harm to people.
In other words:
The agent error is often not just born inside the agent.
Misbehaviour can arise in the gap between the reality of the organisation, the order of authority, tool architecture, human instruction and the measurement system.
This book is a map of these gaps.
In Volume I, we explained what GBO is and the conditions required for correct behaviour. We built a system linking verified identity, genuine capability, validated suitability, valid authority, behavioural integrity, evidence-based action, independent verification, responsible recovery and human sovereignty.
In this second volume, we will not retell the ideal system.
We'll show you where they broke.
An approach from this book's point of view is a think tank only when it describes correct principles; when it calls error forms, the common language of study becomes an auditable system draft when it establishes testable controls.
The 99 records that follow are therefore not a randomly assembled ‘do not’ list.
Each record:
a separate form of failure,
A separate marker of detection,
a separate code of conduct,
a separate audit question
represents.
Some mistakes will be similar.
Authority can be intertwined with identity.
Representation error can become a choice error.
The measurement error can direct the agent to the scope overlay.
A stop error can lead to a compensation problem.
But the first door where each record is broken is different.
This distinction is important.
Because if we fix the result alone, the error will come back in another form.
Deleting the wrong message does not correct its ability to send.
Updating the wrong price does not solve uncertainty in the price source.
Shutting down the sub-agent does not eliminate the way to clear authority.
Apologizing to a user does not set up a withdrawal mechanism.
The purpose of GBO is not to blame the agent.
And human.
The goal is to make the system visible, which makes misbehaviour possible.
Throughout this book, we will repeatedly ask:
Why did the agent act like that?
Then we'll go one step further:
What contract, registration, authority or design was lacking that made this behaviour possible?
And finally:
Which machine-enforced rule would make this failure less likely to recur?
99 errors made at GBO do not claim to cover all future agent issues.
New tools, markets, attacks and forms of behaviour will emerge.
But the universe of error established here will provide a common starting language for assessing behaviour systems.
An organisation is now alone:
"The agent did it wrong."
It shouldn't be enough to say.
They should be able to say:
"The action began before the Identity Gate was passed."
"Research authority was used as external communication authority."
"The machine-readable price contradicted the visible price."
"The lower agent exercised authority that the main agent did not have."
"The restraining order did not spread to the queue."
"The metric of success rewarded unauthorised action."
Once an error has a name, it becomes easier to search for, detect and measure; its chain of responsibility can be examined, and controls can be built to make the same failure less likely to recur.
The first chapter of this book begins from the most basic basis of behaviour:
Identity
Because when an agent chooses the wrong person, the wrong company, the wrong account, or the wrong role, all subsequent accuracy can become meaningless.
The perfect operation on the wrong person is again the wrong one.
The right offer sent to the wrong organisation is again misbehaviour.
An action based on expired authority remains unauthorised, even if it succeeds.
The sentence a realistic synthetic face utters may not be a statement of the real person.
An agent may retain the same persona name even though the technical system, tools and authorisation version behind it have changed.
Therefore, the first nine errors are centered around the following problem:
Does the agent really know who they are dealing with?