Through the Looking Glass - When AI Stops Assisting and Starts Working
In Lewis Carroll’s Through the Looking-Glass, Alice steps through a mirror into a world that looks vaguely familiar.
The people, objects and language are recognisable—but the rules no longer operate as she expects.
I suspect organisations have just passed through a similar mirror.
For the past few years, we have largely treated artificial intelligence as a tool.
Write this.
Summarise that.
Analyse this spreadsheet.
Research this company.
Create that presentation.
The AI waited for us to give it a task. We specified what we wanted and reviewed what it produced.
That model is beginning to change.
OpenAI recently released GPT-6 Astra, a new generation of AI capable of navigating software, selecting tools, maintaining context across long-running assignments and making many of the intermediate decisions required to complete complex work.
Professor Ethan Mollick provided an early demonstration of what this means.
He reportedly gave Astra access to tens of thousands of emails, years of writing, calendar appointments, contacts and unfinished work. He did not prescribe a detailed methodology. Astra selected the software it needed, devised its own approach and spent five days assembling a personal knowledge system covering his research, relationships, ideas and commitments.
It now updates that system twice a day.
Astra did not simply answer a prompt.
It assumed responsibility for an outcome.
Something fundamental has changed
Whether Astra satisfies anyone’s preferred definition of Artificial General Intelligence is almost beside the point.
Executives do not need philosophers and computer scientists to agree on three letters before responding to what the technology can already do.
Until now, we have principally delegated tasks to AI:
Review this contract. Reconcile this account. Prepare this report.
We are beginning to delegate ongoing areas of responsibility:
Keep me informed about the risks I am likely to miss. Keep the books accurate and current. Maintain the integrity of this customer account. Alert me when the evidence supporting this decision changes.
Those assignments have no natural finishing point.
The agent must remember what has happened, monitor changing circumstances, decide whether new information matters and return to the work without waiting for another prompt.
That is not simply a better chatbot.
It is the beginning of a different relationship between people, technology and work.
The organisation on the other side of the mirror
The organisation still looks familiar, but some of its fundamental assumptions are reversing.
People once operated software. Software can increasingly operate software.
AI once waited for instructions. It can increasingly maintain a standing responsibility.
Managers once allocated tasks and checked whether they had been completed. They will increasingly define outcomes, boundaries and authority for teams that include both humans and agents.
Technology once supported the workforce. It is increasingly becoming part of the workforce.
This is why AI is not merely changing what organisations do.
It is changing what an organisation is.
An organisation has traditionally been understood as a coordinated group of people using systems, processes and resources to pursue a shared purpose.
What is it when some of those responsibilities are discharged by persistent intelligent agents that can remember, analyse, act and communicate?
It remains an organisation—but its workforce, management practices and accountability mechanisms have changed.
This is not principally an IT implementation.
It is an operating-model transformation.
Curiouser and curiouser
There is, however, a complication.
As these systems become more capable and autonomous, we may not always be able to observe every internal step through which they reach a conclusion.
That creates an uncomfortable paradox:
We are preparing to delegate more consequential work to machines at precisely the moment their internal reasoning may be becoming harder to inspect.
This does not mean that intelligent agents are secretly plotting against us.
It does mean that governance cannot depend on asking the agent to explain itself and accepting the resulting explanation as evidence.
A confidently expressed rationale is not a control.
We do not govern human organisations by asking everyone whether they acted appropriately and trusting them when they say yes. We establish delegated authorities, approvals, reconciliations, audit trails, exception reporting and independent assurance.
The same disciplines will be required for intelligent agents.
We may not be able to observe every computational step, but we can still determine:
what the agent was authorised to do;
which data and systems it accessed;
what evidence it used;
what actions it took;
what information it changed;
which decisions it escalated;
whether its output reconciled to authoritative sources;
which human remained accountable.
The objective is not to make every machine thought visible. It is to make the work observable, testable and controllable.
The question is therefore not:
Can we trust this agent completely?
People, processes and systems have never been completely reliable.
The better question is:
For which decisions, within which boundaries, under which controls and at what level of residual risk can this agent be trusted to act?
The constraint may not be the technology
Many organisations are approaching agentic AI as another technology implementation:
Identify use cases. Select tools. Run pilots. Scale those that succeed.
That may generate useful efficiencies, but it leaves a more important question unanswered.
Is the organisation itself ready to delegate responsibility to intelligent agents?
An agent cannot safely maintain a customer account, manage a financial process or monitor enterprise risk unless the organisation has defined:
the outcome it is expected to achieve;
the information it can trust;
the authority it has been given;
the decisions it must escalate;
the evidence its work must leave behind;
the human who remains accountable;
how its performance will be assessed and improved.
These are not principally questions about the AI model.
They are questions about the operating model surrounding it.
An agent will struggle to “keep the books accurate” if processes are inconsistent, systems do not integrate, responsibilities are unclear, approvals occur outside controlled workflows and important knowledge remains inside individual heads.
AI does not make those organisational weaknesses disappear.
It encounters them faster.
The organisations best positioned to benefit will not necessarily be those with first access to the most capable model. Most organisations will eventually have access to broadly comparable intelligence.
The differentiator will be the environment into which that intelligence is introduced.
The Red Queen’s race
Carroll offers another useful warning.
In the world Alice enters, the Red Queen tells her that it takes all the running she can do to remain in the same place.
That may soon describe significant parts of the professional workforce.
As people use agents to produce more work, faster, expectations will rise. Work that once required a team may be delivered by one capable professional supported by several persistent agents.
The productivity potential is extraordinary.
But junior accountants, analysts, lawyers and engineers traditionally develop judgement by doing detailed work. They find the reconciliation that does not balance. They follow the transaction back to its source. They discover the assumption that does not survive contact with reality.
Over time, those experiences become the pattern recognition we call expertise.
If intelligent agents perform all the detailed work, where will the next generation acquire the judgement required to supervise them?
Teaching a graduate to “manage an agent” is not enough. Someone who has never learned what good work looks like cannot reliably determine whether an agent has produced it.
An organisation that automates today’s apprenticeship without redesigning how tomorrow’s judgement will be developed may improve productivity while quietly destroying its succession pipeline.
Artificial Intelligence is not Organisational Intelligence
Preparing for this world does not begin with selecting an AI model.
It begins by examining the organisation through a connected set of lenses:
purpose, strategy and outcomes;
customers and stakeholders;
work and processes;
people and capability;
data and organisational knowledge;
technology and integration;
governance, authority and control;
performance, adaptation and learning.
These elements already exist in every organisation. The question is whether they operate together intelligently enough to support a workforce increasingly composed of both people and autonomous systems.
Artificial Intelligence determines what the machine is capable of doing.
Organisational Intelligence determines what it should do, what it is permitted to do, how its work is verified, who remains accountable and whether the organisation becomes better as a result.
Buying access to the most capable AI will not create an intelligent organisation—just as employing intelligent people has never guaranteed that an organisation will behave intelligently.
Sustainable advantage will come from combining machine intelligence with human judgement, institutional knowledge, reliable data, effective governance and the ability to convert insight into coordinated action.
We are definitely through the looking glass now Alice!
The familiar structures of work remain visible, but the rules beneath them are changing.
Most organisations are still asking where they can use AI to improve an existing task.
The more important question is what happens when intelligent agents can assume continuing responsibility, operate across systems and increasingly improve how work itself is performed.
Leaders might begin by asking:
Are our outcomes and accountabilities clear?
Is our information reliable enough to act upon?
Can authority be delegated within explicit boundaries?
Can actions and outcomes be independently verified?
Are we developing the human judgement required to remain in control?
If those questions are difficult to answer, the constraint may not be the intelligence of the machine.
It may be the intelligence of the organisation around it.
AI may provide the capability.
It will still require an intelligent organisation to turn that capability into trusted, accountable and sustainable performance.
We are through the looking glass.
The question now is whether our organisations are ready for what is waiting on the other side.
Comments