Beyond the AI Decision Matrix: Where Should Humans Continue to Matter?
A third question for the age of AI
One of the most important questions facing organisations adopting artificial intelligence is deceptively simple:
What should we allow AI to decide?
Researchers at the MIT Center for Information Systems Research have offered a compelling way to answer it. Their AI Decision Matrix proposes that organisations allocate decision rights between humans and AI according to two characteristics of a decision: its ambiguity and its risk.
The model is grounded in 30 interviews with 27 executives across nine global organisations conducted during 2025 and 2026. It is illustrated particularly effectively through the experience of telecommunications provider One New Zealand, which had more than 50 AI agents operating by early 2026. Among its applications, a customer-segmentation agent translates marketers’ requirements into SQL queries, reducing audience creation time by 60 per cent. Other agents support customer service, financial reconciliation and network management.
One NZ's approach is notable not simply for the number of agents deployed, but for the organisational disciplines around them. Its responsible-AI arrangements include training, lifecycle controls and gradual deployment. No agent is deployed without a named business owner accountable for its outcomes and learning.
This is serious work addressing a real management problem. The MIT framework deserves to be treated as such.
But its argument also exposes a complementary question.
The Decision Matrix helps organisations ask:
When can AI safely and appropriately be entrusted with greater decision-making authority?
There is another question we may increasingly need to ask alongside it:
Where should humans continue to matter?
The distinction is more significant than it first appears.
The strength of the AI Decision Matrix
MIT CISR begins with an important observation: not all decisions are alike.
The first dimension of its matrix is ambiguity. Some decisions are repeatable and predictable, with available data pointing reasonably clearly towards an answer. Others allow multiple legitimate interpretations and depend much more heavily upon context and judgement.
The second dimension is risk: the consequences of getting the decision wrong. Some errors have limited effects and are readily reversible; others carry significant financial, operational or reputational consequences and may be difficult to undo.
Combining the two produces four categories: routine, consequential, exploratory and strategic decisions. The point is not merely to classify the decision, but to determine how humans and increasingly autonomous AI systems should participate in it.
There is an intuitive strength to this. Practitioners familiar with ideas such as the Stacey Matrix or the Cynefin framework will recognise part of the underlying logic. Where cause and effect are clear, rules and established responses become more useful. As ambiguity rises, contextual interpretation and judgement become increasingly important.
MIT applies that reasoning directly to AI decision rights.
It also makes another valuable move. A decision is not treated as a single moment. The researchers separate decision-making into framing, acting and learning. Framing establishes the problem, assumptions, stakeholders, constraints and measures of success. Acting covers information gathering, evaluation, recommendation, authorisation and execution. Learning involves interpreting outcomes, maintaining accountability and adjusting the decision system over time.
The result is a considerably richer model than a simple question of whether a human presses an approval button.
That matters because nominal human involvement is not the same as meaningful human authority.
A necessary question, but not the only one
The MIT matrix is principally concerned with the conditions under which decision rights can move between humans and AI.
That is a necessary question.
But ambiguity and risk cannot tell us everything we may want to know about the appropriate human role.
Consider an activity that is highly predictable. There is little ambiguity about what needs to happen. Errors would have modest consequences and could easily be reversed.
Under the Decision Matrix, that appears to be an excellent candidate for automation.
Often it will be.
Processing a reconciliation, validating standard documentation, routing an enquiry or transferring structured information may gain little simply because a person performs the activity. If technology can do the job faster, more accurately and with less effort, insisting upon manual work merely to preserve a human role would be difficult to justify.
Now consider another interaction.
A customer contacts an organisation following the death of their spouse.
The administrative process may also be predictable. The evidence required could be clear. The permitted actions might be tightly prescribed. The financial consequences of an error could be relatively limited and reversible.
Viewed only through ambiguity and risk, much of the interaction may again appear highly automatable.
Yet something important has changed.
The value of the interaction does not lie solely in reaching the administratively correct outcome.
How the person experiences it matters.
Recognition matters. Care matters. Tone matters. The capacity to respond to a particular person rather than simply process a category of case matters.
The distinction is not primarily about whether AI can safely perform the task.
It is about whether human participation is itself part of the value.
That requires another lens.
From capability to human value
This question sits at the centre of an emerging framework called Lúwāra.
Lúwāra starts from a deliberately different position. Before deciding where AI should be used, it asks what is valuable about human participation in the first place.
Its core proposition is simple:
AI should be used where it adds value; human contribution should be protected where human contribution is the value.
This is an AI-positive position.
It does not assume that more human involvement is always better. It does not seek to preserve work merely because people used to perform it. AI can remove drudgery, accelerate analysis, improve consistency, widen access and extend human capability.
The question is one of appropriate allocation.
Where is the recipient primarily buying or valuing an outcome?
And where is the judgement, relationship, authorship, care, craft or expression of a person part of what makes that outcome valuable?
This produces a third spectrum:
Functional ↔ Expressive
At the functional end, value lies predominantly in the result.
Was the calculation correct? Was the payment processed? Did the information reach the right place? Was the service completed quickly and reliably?
At the expressive end, the human quality of the activity increasingly forms part of the result.
Who created it may matter. The relationship may matter. Individual judgement, lived experience or authorship may matter. There may be a reasonable expectation that what appears to be human communication or expression genuinely involved a person.
“Expressive” should not be mistaken for artistic alone. A counselling conversation, an act of care, a manager giving difficult feedback or a teacher responding to a struggling pupil may all possess expressive human value without being works of art.
Equally, “functional” does not mean unimportant. Some of the most critical processes in an organisation are highly functional.
The distinction asks a narrower question:
How much does the human quality of the activity itself matter?
Three independent dials
Seen this way, Lúwāra's practical triage uses three spectra.
The first is Functional ↔ Expressive: how much does meaningful human participation form part of the value?
The second is Determinate ↔ Ambiguous: how knowable is the right answer, and how much contextual interpretation or practical judgement does it require?
The third is Low ↔ High Criticality: what happens if this goes wrong, and how readily can the consequences be reversed?
These are deliberately independent dimensions.
They should not simply be added together to create an “automation score”.
A highly functional process might also be highly critical. That does not necessarily mean a person should perform the work. It may mean that automation requires extremely strong assurance, clear accountability and effective intervention mechanisms.
An expressive activity might be low risk and highly determinate while still warranting substantial human contribution because the identity, care or authorship of the person participating is part of the value.
Similarly, a high-ambiguity activity may demand human judgement without necessarily demanding direct human execution of every resulting action.
The three dials therefore expose three different reasons for retaining human involvement.
Human involvement may matter because the consequences of failure are serious.
It may matter because the situation requires contextual judgement.
Or it may matter because human contribution is intrinsically part of what is being created.
Those reasons should not be treated as proxies for one another.
Three forms of human involvement
Once these distinctions are made, “human in the loop” begins to look rather imprecise.
Lúwāra instead asks three linked questions of any AI-enabled activity. It calls them tests: Human Experience, Human Contribution and Human Authority.
Human Experience asks what it is actually like to be a person inside the system: whether the workers, customers and others affected keep agency, dignity, trust and a sense that they matter, and whether the organisation’s account of the change matches what people actually live with. It is a test in its own right, not a satisfaction score added at the end.
Human Contribution asks what people actually add to the value being created. This may include judgement, care, relationship, authorship, craft, expression and the continuing development of these capacities. It also covers truthful claims about who, or what, did the work: provenance is part of contribution.
Human Authority concerns where legitimate and effective authority resides: who defines the question, who decides, who judges the outcome, and whether people can in practice challenge, intervene, reverse and stop AI-enabled activity when it goes wrong.
The three are connected, but they are not interchangeable.
A well-controlled AI system does not demonstrate human authorship.
A human signature does not prove that meaningful authority has been exercised.
And the presence of a pleasant customer experience does not establish either.
Assurance, in Lúwāra’s terms, is not a fourth test. It is the evidence, processes and controls that show an organisation takes all three seriously, and it is what any future assessment would examine.
This distinction becomes important as organisations move from isolated AI tools towards increasingly autonomous systems. The relevant question is no longer simply whether a human is somewhere in the process, but what that human is actually there to do.
Beyond the approval button: Define, Decide, Evaluate
Here Lúwāra builds directly on an insight within the MIT research.
MIT's framing, acting and learning model shows that control can move away from humans at several different points within a decision. Lúwāra adapts that logic into three questions of authority:
Define. Decide. Evaluate.
Who defines?
Who defines, or frames, the problem being addressed? Who determines its purpose and boundaries? Which evidence counts? Which stakeholders matter? What trade-offs are acceptable? What constitutes success?
These choices shape the decision before an option is ever selected.
Who decides?
Who actually resolves the trade-offs and makes the choice? Who accepts or rejects a recommendation, approves an exception or changes course?
Importantly, deciding is not the same as executing.
Once a legitimate decision has been made, there may be no particular value in requiring a person to carry out every resulting action. AI may execute an authorised choice perfectly appropriately.
Controls over execution, including intervention, escalation, reversal and the ability to stop a system, belong to Human Authority. Authority that cannot stop the system is not authority.
Who evaluates?
Who subsequently judges whether the outcome was good?
Not merely whether a target was achieved, but whether the target was the right one. For whom was the outcome good? What unintended consequences occurred? Were the original assumptions sound? Should the criteria or process now change?
Learning follows evaluation.
An automated system can optimise against a feedback signal. But somebody still has to determine what deserves to count as success.
This exposes a weakness in some apparently “human-controlled” systems.
A person may retain the final approval while AI has already defined the problem, selected the evidence, generated the alternatives and established the criteria by which the outcome will later be judged.
The human technically makes the decision.
The substance of human authority may nevertheless be very thin.
MIT CISR similarly stresses that decision rights should shift according to risk and ambiguity while maintaining clear human accountability. In One NZ, named human owners remain responsible for deployed agents and their outcomes.
The question is therefore not whether AI participates in Define, Decide or Evaluate. It can usefully assist at all three.
The question is whether human owners retain the competence, information and practical authority to challenge and change what the system is doing.
The hidden cost of removing judgement
There is a further consequence that becomes visible when human contribution is considered over time.
AI can support human judgement extraordinarily well. It can find patterns, retrieve evidence, generate alternatives and expose inconsistencies at a speed no person could match.
But organisations also need to ask where future human judgement comes from.
Practical wisdom is not acquired simply by being designated the accountable human.
It develops through experience.
People build judgement by confronting varied situations, seeing decisions succeed and fail, learning from more experienced colleagues, testing their interpretations and taking responsibility for consequences.
The classical term phronesis, usually translated as practical wisdom, captures something important here. Good judgement is not merely reasoning applied to information. It is the capacity to interpret context, weigh competing considerations and decide what should be done.
If AI progressively performs the formative work through which people acquire that capability, an organisation may retain human authority on paper while quietly eroding the human competence required to exercise it.
A junior professional who only approves highly developed AI recommendations may never acquire the judgement of the experienced professional they are eventually expected to replace.
This creates a longer-term question for AI-enabled organisations:
Are we using AI to extend human judgement, or unintentionally removing the experiences from which future judgement develops?
The issue is not an argument for artificial inefficiency. Organisations should not manufacture unnecessary manual work simply as training.
But human capability is itself an organisational asset.
Automation decisions should consider not only today's performance but the knowledge, judgement and practical wisdom the organisation will need tomorrow.
The experience of being human in the system
There is another dimension that does not reduce neatly to a decision right.
Two organisations could allocate authority identically and still create radically different human experiences.
One might leave customers and employees feeling respected, capable and able to exercise agency.
Another might leave them feeling powerless, monitored, alienated or irrelevant.
Lúwāra therefore treats Human Experience as the first of its three tests, examined in its own right, rather than as a lens applied afterwards or another score to be averaged.
The underlying question is deliberately simple:
What is it actually like to be a human being here?
Do people experience agency?
Do they understand when and how AI is affecting them?
Can they challenge what happens?
Do they feel that their contribution matters?
Does technology extend their capability, or progressively reduce them to attendants of systems they no longer meaningfully influence?
Customer satisfaction measures and attractive interfaces can contribute evidence, but they cannot settle those questions alone.
This matters because responsible AI is not simply a property of the technology. It is a property of the human system created around it.
Human control is necessary, but human value goes further
There is already an important body of work around meaningful human control.
Research in this field has emphasised that humans need both the ability and the authority to exercise responsibility over autonomous systems, including genuine mechanisms for intervention and correction. That work is especially important in high-consequence contexts.
Lúwāra is compatible with that argument.
But it asks a different question as well.
Human beings are not valuable only as supervisors of machines.
Sometimes a human should remain involved because AI cannot safely be trusted alone.
Sometimes a human should remain because the circumstances demand contextual judgement.
And sometimes a human should remain even where AI is reliable, safe and capable, because the person's involvement is part of the value being promised.
That third case is easy to overlook.
As AI capability improves, it may also become the most important.
Historically, technical inability protected many human roles automatically. Machines simply could not perform them.
That boundary is disappearing.
Capability therefore cannot be allowed to become the sole criterion for substitution.
Starting from the other end
MIT CISR concludes that leaders should move beyond asking where AI can be used and instead manage the business decisions AI shapes. That is an important progression.
Lúwāra proposes another shift alongside it.
Much AI strategy understandably begins with the technology.
What can we automate?
What can we make faster?
Where can autonomous agents operate?
Where does risk mean that a human must remain?
Lúwāra begins with the human role worth preserving.
Where does human judgement matter?
Where does authorship matter?
Where do care, relationship or expression form part of the value?
Where must people retain genuine authority?
And what does the resulting system feel like to the humans who live with it?
Only then does it ask how AI should participate.
This does not lead automatically to greater human involvement.
In some settings, the appropriate answer will be extensive automation.
There is nothing inherently humanising about requiring people to perform repetitive transactional work that technology can do better.
Elsewhere, AI will augment people: removing administration, providing information, challenging assumptions or generating alternatives while humans retain the contributions that matter.
And in some places organisations may consciously preserve human participation even when substitution becomes technically straightforward.
That is not necessarily inefficiency.
Efficiency is only one form of value.
We already understand this elsewhere in human life. We value live performance when recordings are more convenient. We value handmade objects despite the precision of industrial production. We hold ceremonies whose practical function could be completed far more quickly without the ceremony. We value a personal message differently from an automatically generated one, even where the words are equally polished.
In each case, the process, provenance or human act contributes something to the value of the outcome.
AI does not remove that distinction.
It makes choosing deliberately more important.
From automation strategy to human design
The organisational challenge is therefore broader than identifying AI use cases.
It is one of intentional design.
For each material process, interaction or decision, leaders can consider three independent questions.
How functional or expressive is it?
How determinate or ambiguous is it?
How low or high is its criticality?
The resulting profile can then inform what Human Contribution and Human Authority are appropriate, while Human Experience shows whether those arrangements are real for the people who live with them. Assurance is the evidence that all three hold.
This is not a formula that automatically allocates work to people or machines.
Context matters.
Customer expectations matter.
Provenance matters.
And preserving human contribution should not become an excuse to preserve avoidable toil or impose unwanted human contact.
The point is to replace accidental substitution with conscious choice.
The next question
The MIT AI Decision Matrix provides organisations with a powerful answer to a pressing management problem.
How should humans and AI share decision rights as ambiguity and risk change?
Its research and the experience of One New Zealand show how that question can be translated into practical governance rather than remaining an abstract principle.
But increasing AI capability leaves another question unresolved.
Leaders will need to decide not only where machines are capable of operating autonomously, nor only where risk and ambiguity mean that humans must remain.
They will need to decide what kinds of human contribution they still consider valuable.
That means thinking about judgement, responsibility, authorship, relationship, care, expression, capability and lived human experience before those qualities disappear unintentionally from the value chain.
The AI Decision Matrix asks, in effect:
When can AI be trusted with the decision?
Alongside it, we should increasingly ask:
Where should humans continue to matter?
Because automatable is not the same as appropriate to automate.
And an organisation that becomes more intelligent through AI should not have to become less human in the process.
Lúwāra is an emerging framework. Its method and pilot programme are in development, and nothing here describes an operational certification.
Three questions are still open, and I would value help with them. How can an organisation show that human control over an AI-assisted decision is real, rather than a signature on someone else’s recommendation? How should human contribution be judged where it is itself part of the value, such as care, authorship or craft, without treating more human effort as better? How can lived experience, meaning how it actually feels to work somewhere, be evidenced fairly without reducing it to a satisfaction score?
If one of these is your question too, I am holding thirty-minute conversations with people who run AI-enabled work. ray.mead@luwara.org · luwara.org
References
Sebastian, I. M., Weill, P., Haskamp, T. and vom Brocke, J. (2026). Designing Decision Rights for AI. MIT CISR Research Briefing, Vol. XXVI, No. 6. Massachusetts Institute of Technology, Center for Information Systems Research.
Cavalcante Siebert, L., Lupetti, M. L., Aizenberg, E. et al. (2023). “Meaningful human control: actionable properties for AI system development.” AI and Ethics, 3, 241–255.