Within the next three years, most large organizations will have incorporated AI agents into their operations. Some will have gained a genuine competitive edge. Most will have improved their efficiency.

The difference between one outcome and the other will not depend on which technology they chose. It will depend on whether they redesigned how they make decisions.

Decisions are the mechanism through which an organization converts its strategy into results. Not processes, not data, not models — decisions. And AI agents are not simply a more efficient automation tool. They represent a redistribution of decision-making capacity. That requires deliberate redesign. The organizations that understand this today will build an advantage that will be very difficult to replicate — not because the technology is inimitable, but because the ability to decide well cannot be copied: it must be built.

Most organizations incorporating AI today do so for an understandable reason: efficiency. They automate tasks, reduce timelines, and free up working hours. It is a real benefit, and it makes sense to pursue it.

The problem is stopping there.

Automating tasks is not the same as redesigning decisions. And the difference between the two is not incremental — it is an order of magnitude.

When an organization improves efficiency, it reduces costs and frees up capacity. That impacts margins. When it improves the quality of its decisions, it directly impacts the outcomes that matter most: revenue, market share, long-term margins, competitive positioning. And also objectives that are not always measured in financial terms: emissions reduction, reputation, talent retention.

Decisions are the mechanism through which an organization converts its strategy into results. Better decisions do not do the same things faster: they do different things, at the right moment, with the right criteria. The cumulative impact of that over time far exceeds that of any operational efficiency program.

An organization can automate hundreds of processes and still lose in the market because it continues making exactly the same decisions as before — just with less friction.

What we see in consumer goods

In many consumer goods organizations, Revenue Management teams produce sophisticated analysis: pricing models, promotional simulations, mix recommendations. The problem is not the quality of the analysis. It is that this analysis rarely reaches the decision.

Why? Because marketing looks at brand health over the medium term, sales looks at quarterly revenue, and finance looks at margin. Each area optimizes its own function. If the model was not designed to capture and articulate those three perspectives, the output ends up as an exercise that no one truly engages with. Decision-makers act under short-term pressure, often without ever reviewing the recommendations.

To be clear: there is nothing wrong with a decision-maker, having reviewed the recommendation, consciously choosing a different course. That is judgment. What is a problem is when this happens systematically, without genuine dialogue between the parties, and without learning — when incentives are misaligned and no one is optimizing the right function.

At a dairy company we work with, 38 people produce this type of analysis, mostly manually. The impact on business decisions is marginal. Not because the work is poor. But because the system connecting that work to the actual decisions was never designed.

AI agents are not a more powerful version of automation. They are something qualitatively different: systems that can observe context, generate alternatives, recommend courses of action and, in certain domains, execute decisions without human intervention.

That is not efficiency. It is a redistribution of decision-making capacity within the organization.

When an organization incorporates agents, it is inevitably answering — whether it realizes it or not — questions of architecture: What kinds of decisions can an agent make autonomously? Which ones require human validation? Who defines the criteria by which the agent evaluates alternatives? Who is accountable when the agent decides poorly?

If these questions are not answered explicitly and deliberately, the organization answers them anyway — by default. And default answers are rarely the best ones.

A sound decision architecture does not start with technology. It starts by mapping which decisions drive business results, and from there determining what role each actor — human or artificial — should play in each type of decision.

There are decisions that agents can make better than humans: fast, repetitive ones, based on recognizable patterns and with clear criteria. Automating them is not a risk — it is an obligation. Continuing to invest human judgment in those decisions is, in itself, a poor decision.

There are decisions that agents can enhance without replacing: the agent processes information, identifies alternatives, and quantifies trade-offs, while the human integrates context, assumes risk, and decides. This is the most powerful combination — and also the most underutilized. Across every function of an organization, there are people making decisions with incomplete information, undetected biases, and implicit criteria that no one has questioned. A well-designed agent does not replace that judgment: it elevates it.

And there are decisions that remain exclusively human: those that require judgment about values, those with irreversible consequences, those that define the direction of the organization. Not because agents cannot offer a perspective, but because accountability cannot be delegated to a system.

The work of architecture is knowing which is which. And technology does not resolve that.

What we see in Oil and Gas

Exploration and production decisions in the oil industry involve enormous sums and a factor that tends to be underestimated: the cost of time. Every month of delay in a production decision is money left in the ground. And unlike other industries, windows close: market conditions shift, assets depreciate, opportunities vanish. Deciding late in oil and gas is not just being slow — it means permanently leaving money on the table.

What we observe in several organizations in the sector is that analytical models have improved considerably. They have more data, better algorithms, and greater processing capacity. The problem is that feeding those models remains a slow, manual process: multiple areas produce estimates independently, with different criteria, on different timelines, with no clear design for how those inputs are integrated.

The bottleneck is not the model. It is the architecture of the decision-making process across functions. The result is paradoxical: organizations with sophisticated analytical capabilities that continue making decisions at the pace of twenty years ago.

The risks and opportunities of this moment are many. But there are some patterns worth naming because they are already happening.

When an organization incorporates agents without redesigning how it decides, something predictable tends to emerge: teams begin to follow the system’s recommendations without questioning them. At first it looks like efficiency. Over time, no one can explain why certain decisions are being made, on what criteria, or with what accepted margin of error. That is not a technology problem. It is an architecture problem that technology has made visible.

But that same phenomenon points to an opportunity. Across every function of an organization, there are different criteria operating in parallel, decisions made with undetected biases, implicit judgments that no one has questioned. This already exists with humans — it is simply invisible because it is distributed across individuals rather than systems. Well-designed agents do not create that problem: they surface it and enable it to be corrected. An explicit decision architecture turns what is invisible into something that can be measured, improved, and scaled.

What the most advanced organizations are doing

Some organizations are already further along this path. In the pharmaceutical industry, for example, companies have deployed AI systems specifically aimed at reducing cognitive biases in research portfolio decisions — one of the most costly errors in the sector is continuing to invest in projects that the data indicates will not succeed, simply because a great deal has already been spent on them. The systems do not replace the decision: they equip leaders to redirect resources with greater clarity and less inertia.

The result is not that AI decides. It is that humans decide better — with less noise and with more explicit criteria. That is exactly the right direction.

Consider algorithmic trading. When the first funds adopted algorithms to operate, they generated extraordinary advantages over those who continued to decide manually. Over time, algorithms became widespread, advantages compressed, and the game changed permanently: today, not having algorithms is not a competitive disadvantage — it simply means not being able to play.

Something analogous is happening with decision architecture and agents, but with an important difference: we are still in the early stages. Organizations that invest today in redesigning how they decide — incorporating agents deliberately rather than reactively — are building an advantage that will be very difficult to replicate later. Not because the technology is inimitable, but because the organizational capability to decide well cannot be copied: it must be built. And it takes time.

The window is open. It will not be open forever. Technology can be bought. The ability to decide well must be designed.

Ernesto Weissmann
Partner at Tandem
ew@tandemsd.com

Tandem works with complex organizations to design decision architectures: who decides what, with what information, at what point, and on what criteria. If this topic is relevant for your organization, we would be glad to have a conversation.