ARE DECISIONS THE NEW UNIT OF AI TRANSFORMATION?

Most AI transformation programmes are built from the supply side.

They catalogue capabilities, identify use cases, estimate hours that might be saved and gradually assemble a portfolio of assistants, copilots, agents and automations. This is a perfectly reasonable way to deploy technology. It is a less convincing way to describe transformation.

A company can automate thousands of activities without materially changing how it prices a product, allocates capital, accepts risk, chooses suppliers, evaluates talent or decides which customers deserve attention. Conversely, a relatively small AI system can have enormous organisational significance if it changes one of those decisions.

This is not a new observation about organisations. It is, strangely, an old one that the current AI debate has managed to rediscover by taking the scenic route.

Herbert Simon’s Administrative Behavior proposed that organisations could be understood through the decision-making behaviour of their participants; his work on decision-making inside economic organisations later received the 1978 Nobel Prize in Economics. (Simon, 1947/1997).

James Galbraith subsequently described organisational design as an information-processing problem: uncertainty increases the amount of information that must be processed between decision-makers during the execution of work. (Galbraith, 1974).

Both ideas become unusually relevant when the marginal capacity to search, analyse, compare, predict and generate alternatives changes by several orders of magnitude.Current AI transformation thinking is beginning to move in this direction. The main advantage of AI operating models come from redesigning both how work is performed and how decisions are made, rather than simply inserting AI tools into existing processes.I think the implication should be taken further.The decision should not be one component of the operating model discussion.It should become the unit through which the transformation is analysed.Consider pricing.Company A uses an AI system to collect competitor prices, analyse historical sales and draft the weekly pricing pack. The analysts save several hours. The pricing committee still meets on Thursday, considers broadly the same alternatives and changes prices at the same level of aggregation.Company B uses a similar model, but restructures pricing around continuously updated signals. Decisions that were previously made by product category are now made for individual accounts. Some price changes occur automatically within defined parameters; unusual cases are routed to commercial teams.Calling both initiatives “AI for pricing” obscures almost everything that matters.In Company A, the production of information changed.In Company B, the resolution, frequency and locus of the decision changed.Those are different transformations even if the technology underneath them is nearly identical.Tasks are therefore useful implementation units. Workflows are useful process units.Decisions are closer to economic units because they reveal whether new information-processing capacity has changed what the firm actually does.

Organisations contain a surprising number of mechanisms for avoiding decisions.Policies compress many possible situations into one rule.Thresholds avoid reconsidering every case.Hierarchies determine which decisions deserve scarce managerial attention.Committees aggregate information that is distributed across functions.Annual planning converts an effectively continuous strategic problem into something that can be discussed on a calendar.Customer segments allow firms to make one commercial decision for thousands of people who are, inconveniently, not actually identical.These arrangements exist for many reasons, including coordination, consistency, control and legitimacy. But they also reflect a basic historical constraint: making differentiated decisions is costly.Someone has to obtain the information.Someone has to interpret it.Someone has to compare alternatives.Someone has to coordinate with the people affected.Then the organisation has to repeat the exercise the next time the world changes.Simon called attention to precisely this kind of constraint through bounded rationality: decision-makers do not optimise over every possible alternative because information, attention and computational capacity are limited.AI changes the economics of several of those constraints simultaneously.A system can search a large information space, analyse it, generate alternatives and interact with other software without requiring a proportional increase in human labour. Agentic systems now extend that capacity across longer sequences of work rather than isolated predictions or generated artefacts. OpenAI’s 2026 research on Codex, for example, documents users increasingly delegating tasks estimated to span hours of human work and orchestrating several agent activities in parallel.The interesting consequence is not simply that existing decisions become cheaper.Some decisions that were previously uneconomical to make can begin to exist.A retailer can move from nationally determined promotions to store-level decisions.A manufacturer can update replenishment in response to local signals rather than a fixed planning cycle.A bank can reconsider a customer interaction at each relevant event rather than only during periodic reviews.A software company can continuously reprioritise parts of a backlog instead of waiting for the next planning ceremony.In each case, AI increases what I would call decision resolution.The analogy is deliberate.A low-resolution image combines detail into large blocks because representing every distinction is costly. Many organisational decisions work in the same way. Categories, segments, periodic reviews and standard policies compress a more complicated reality into a manageable number of choices.As the cost of processing context declines, the organisation can potentially decide at a finer resolution.This happens along at least two dimensions.The first is granularity. Instead of deciding for a category, the organisation can decide for a case.The second is cadence. Instead of deciding periodically, the organisation can decide when relevant information changes.A third change follows naturally: authority can move.Once the information required for a decision can travel with the system, a choice previously concentrated in a specialist function may become possible at the edge of the organisation or inside software itself.That last step is not an implementation detail.Athey, Bryan and Gans model the allocation of decision authority between human agents and AI and show that the appropriate allocation depends on more than AI reliability. Giving authority to a human can affect whether that person continues to invest effort in learning information that matters to the decision. (Athey, Bryan & Gans, 2020).This is one reason the usual question, “What percentage of this process can be automated?”, is strategically weak.It treats the organisation as a collection of labour inputs.The more interesting question is how the technology changes the economically sensible level at which the organisation chooses.There are good reasons not to maximise decision resolution.A company capable of changing every price for every customer every minute has not thereby discovered that it should.Fine-grained decisions can create inconsistency.Continuous decisions can make causal effects harder to understand.Local optimisation can damage global objectives.Rapid adaptation can make organisational behaviour less predictable to the people expected to rely on it.Some categories exist because computing every case was historically too expensive. Others exist because consistency is itself valuable.AI makes it necessary to distinguish the two.

There is another reason to use decisions rather than tasks as the unit of analysis: it becomes much harder to pretend that “human in the loop” has settled anything.Organisations often locate authority at the final visible action.The manager approves, the recruiter selects, the executive decides.But the final selection is only one part of how a decision is produced.Someone determined which information was relevant.Someone defined the objective.Someone excluded alternatives that did not satisfy a threshold and ranked the remaining options.By the time a person receives the recommendation, much of the practical decision may already have occurred.Research on organisational AI has been making this point for some time. Shrestha, Ben-Menahem and von Krogh distinguish between full human-to-AI delegation, sequential human-AI structures and arrangements in which human and machine decisions are aggregated. They show that the appropriate structure varies with properties including interpretability, speed, replicability and the size of the relevant search space. (Shrestha, Ben-Menahem & von Krogh, 2019).More recent work makes the sequencing problem even sharper. Zhong models multilayered decision processes in which both humans and technologies can correct existing errors while introducing new ones, showing that the location of a technology within the sequence can matter independently of its standalone quality. (Zhong, 2026).The practical consequence is quite important.Two companies can both truthfully claim that a human retains final authority while giving that human radically different decisions to make.In one organisation, the professional receives evidence, uncertainty estimates and several plausible alternatives.In another, the professional receives one recommendation, generated after the system has already filtered the evidence and ranked the possibilities.The signature is human in both cases.The cognitive architecture is not.This is where workflow diagrams tend to become unhelpful. They are good at representing sequence: activity A moves to activity B, which triggers activity C.A decision requires a different description.For a consequential choice, I would want to know what is being decided and at what level of granularity; which information can alter the choice; who defines the criterion of success; where AI enters the reasoning; who can change or reject the recommendation; how reversible the action is; and how the organisation will discover whether the decision was good.That description is closer to a decision architecture.It also reveals why the same model can have completely different organisational consequences depending on where it is placed.Raisch and Krakowski’s work on the automation-augmentation paradox is useful here. They argue that automation and augmentation are interdependent rather than clean alternatives, and that focusing excessively on either can create reinforcing organisational effects. (Raisch & Krakowski, 2021).The boundary becomes even less stable with generative AI.An employee may appear to be augmented because the final decision remains theirs, while the activities through which they once developed the judgement required to make that decision are progressively automated.The arrangement works perfectly until the machine encounters something outside the pattern it has learned and the supposedly supervising human discovers that supervision was also a skill.Acemoglu, Kong and Ozdaglar model a related long-run problem. Their 2026 work shows how agentic AI can improve present decision-making while reducing incentives for people to acquire context-specific knowledge, potentially weakening the human knowledge environment on which future decisions depend. (Acemoglu, Kong & Ozdaglar, 2026).This does not imply that humans should continue performing obsolete work for educational reasons.It means that the design of a decision has a temporal dimension.The organisation should care about today’s answer and about whether it is preserving the capability to recognise when tomorrow’s problem is different.

The usual enterprise AI portfolio contains things such as customer-service copilots, contract-review tools, coding assistants, finance agents and knowledge assistants.These categories identify systems.They tell us surprisingly little about the strategic importance of the choices being changed.A decision-centred portfolio would start from a different inventory. Take pricing again.The useful questions are not initially about models.How often is the decision made?How much economic value moves when it is made well rather than badly?How much unexplained variation exists in current human decisions?How quickly does the outcome become observable?Can a bad decision be reversed?Does better information genuinely improve the choice, or is the underlying problem disagreement over objectives?Is consistency valuable?Does the organisation learn something important by having people make the decision themselves?Only then does model capability become interesting.This approach also changes measurement.Usage is a weak metric for transformation because it measures interaction with technology.Hours saved are better, but still mainly measure production.A decision provides a unit that can be connected to outcomes.For a credit decision, that might mean default, approval quality, margin and fairness across relevant groups.For pricing, margin, volume, churn and the speed with which an incorrect assumption is detected.For procurement, total cost, resilience, supplier performance and concentration risk.For hiring, later performance, retention and selection quality rather than the number of CVs processed per hour.There is no universal metric because decisions have different objectives.That is precisely the point.A transformation portfolio should be able to explain which organisational decisions improved, on what evidence, and at what cost.Experimental research suggests that the answer will not always be “give the AI recommendation to a human”.Agarwal, Moehring and Wolitzky find in a fact-checking experiment that people changed their own effort in response to AI information; in their setting, the optimal policy selectively automated some decisions while delegating others to humans with access to the AI prediction. The result is useful because it treats human behaviour as endogenous to the system rather than assuming that adding a person automatically improves it. (Agarwal, Moehring & Wolitzky, 2025).This is the level at which AI transformation starts to become an organisational design problem rather than a deployment programme.Some decisions should become faster.Some should become more granular.Some should move closer to the customer.Some can be delegated almost entirely to machines because outcomes are observable and mistakes are inexpensive to reverse.Some deserve extensive machine analysis but intentionally slow human commitment.Others may remain coarse by design because consistency, legitimacy or strategic coherence matters more than local optimisation.The objective is not to create the maximum number of AI-mediated decisions.It is to change the right decisions at the right resolution.That is also where I think the current conversation about agentic organisations is still too focused on work.Agents clearly matter because they expand the range of activities that can be delegated. However, firms need to redesign operating models around this shift rather than layering agents over existing workflows.But the ultimate object of that redesign is not the workflow.A workflow is only valuable because something at the end of it matters.A customer is served.A loan is approved.A machine is repaired.Capital moves.A product ships.A person is hired.A risk is accepted.Once that becomes the unit of analysis, some large AI projects become strategically small and some apparently narrow systems become extremely important.That is useful information for anyone deciding where the next euro of transformation investment should go.

There is a broader implication here.For decades, organisations have been shaped by the cost of acquiring information, moving it through hierarchies and applying scarce expertise to a manageable number of decisions.

Many familiar structures are partly solutions to those constraints.

Galbraith’s information-processing account of organisational design was written in 1974, but the underlying question remains remarkably current: organisational structures reflect the information required to perform work under uncertainty.AI is changing that information-processing constraint.The important question is therefore not simply how many human activities a model can perform.It is what happens to the organisation when the economics of deciding change.Some firms will use AI to make existing decisions more efficiently.The more interesting ones will notice that decisions themselves can be redesigned: made at a different level, on a different cadence, with different information and a different allocation of authority.That shift is much harder to see in a use-case catalogue.It is also much closer to transformation.A useful test for an AI programme is therefore brutally simple:Which important decisions does this programme allow the organisation to make differently?If the answer is unclear, the company may still have an excellent automation programme.It may simply not have changed very much yet.