When AI Starts Making Decisions

I believe enterprise data leadership has to evolve from primarily delivering information to creating the conditions for consistent, trustworthy decisions.

Historically, data leadership focused on making sure the right data reaches the right people so they can make better decisions. With autonomous and agentic AI, the system may participate in those decisions and execute actions at a much higher speed.

That changes the leadership responsibility. We still need trusted data, but we also need to govern the context around the decision. What business rules apply? What architecture patterns are acceptable? What dependencies exist across domains? When does the system have enough confidence to proceed, and when should it escalate to a human?

For me, that is where decision intelligence becomes important. It connects data, business context, architecture, governance, and execution into a repeatable decision process.

The other important shift is retaining the decision history. If an AI system or a team makes an important decision today, six months later the organization should be able to understand what was decided, why it was decided, what evidence supported it, and what tradeoffs were considered.

So, I see the future role of enterprise data leadership as creating the foundation for better decisions at scale. That still starts with trusted data, but it also includes shared context, clear decision rights, reusable patterns, traceability, and runtime controls.

That allows organizations to gain the speed of AI without allowing speed to create organizational drift.

When AI agents begin making or influencing decisions at scale, the challenge is keeping architecture, governance, and execution connected as those decisions happen.

Traditionally, we've managed these disciplines through separate processes. Architecture establishes the design, governance defines the rules, and delivery teams implement them. That approach works when changes move through predictable review cycles. It becomes harder when agents operate continuously, sometimes crossing systems or business domains where different constraints apply.

I see three areas that deserve attention:

• Establish clear boundaries. Architecture defines the patterns and technical constraints, while governance establishes the policies and decision rights. Those boundaries need to translate into controls that can actually be enforced when an agent acts. Otherwise, we risk having well-defined principles that don't consistently hold in production.
• Bring governance into execution. Traditional reviews happen at defined points in time, while agents can make decisions continuously. We need policies, authorization checks, and escalation rules available during execution, with the level of oversight determined by the consequences of the action.
• Preserve context across handoffs. This is one of the areas I find particularly important. An agent might make a reasonable decision within its own workflow, but that decision could create unintended consequences when another agent or business function acts on it. Preserving the reasoning, applicable constraints, and decision history across those handoffs helps prevent local decisions from creating broader enterprise problems.

Accountability also needs to remain explicit. Agents can recommend or execute within delegated authority, but someone in the organization must own that authority, including the conditions under which human intervention is required.

The real test comes in production. Can we trace why an agent acted, confirm it operated within its authority, and understand the consequences across the broader workflow? That's where architecture, governance, and execution either stay aligned or begin to drift.

Enterprises need to preserve enough context to reconstruct the decision, not just record the final outcome.

That means capturing what problem was being solved, what data and evidence were available at the time, which policies and architecture standards applied, what alternatives were considered, what tradeoffs were accepted, who had decision authority, and what dependencies or constraints influenced the choice.

The timing also matters. A decision may have been completely reasonable based on the information available six months ago, even if the organization would make a different decision today. Without that historical context, AI can easily misinterpret the decision or treat an old choice as a current best practice.

So, I would think about decision context in layers: business intent, evidence, constraints, alternatives, rationale, accountability, and outcome.

If we preserve those elements in a structured and traceable way, AI can do more than retrieve history. It can understand why a decision was made, determine whether the same assumptions still apply, and use that history more intelligently in future decisions.

Institutional knowledge becomes valuable when we can reuse it without having to reconstruct the reasoning behind it every time.

Think about an architecture decision made three years ago. It may have been the right choice given the regulatory requirements, technology limitations, and business priorities at the time. But those conditions may have changed. If all we preserve is the decision itself, someone could follow the same approach today without realizing why it was originally chosen.

I've always found that organizations are reasonably good at documenting what they decided, but much less consistent at preserving why. To make that knowledge reusable, we need to connect decisions to their supporting evidence, business context, ownership, and subsequent changes. That gives people a way to trace the reasoning and understand whether it still applies.

Decisions also get revised, policies change, and sometimes an organization deliberately reverses course. Preserving that history helps us understand how our thinking evolved, including what we learned along the way.

For AI, this opens up an interesting opportunity. We can make institutional knowledge accessible across the enterprise, but retrieval alone isn't enough. AI systems need enough context to distinguish a decision that still applies from one whose assumptions have expired.

Think about how much time organizations spend rediscovering their own reasoning, revisiting old decisions, or tracking down someone who remembers why something was done. Preserving that context allows teams to build on what the organization has already learned rather than starting from scratch each time. That's where I see some of the biggest value.

As AI takes on more decision authority, the accountability model has to become clearer, not weaker. Ultimately, accountability should remain with the organization and with clearly defined human roles.

AI can be given authority to recommend or execute decisions, but the enterprise is still responsible for defining the boundaries within which that authority operates.

That means accountability has to be designed into the operating model. It is important to be clear about ownership. Who owns the decision in the first place? Who decided the agent can operate with that level of autonomy? What can it actually see and do? And if it gets into a situation outside those boundaries, who takes over? If those answers are vague, accountability gets vague very quickly.

I also think accountability should be proportional to consequence. A low-risk, reversible action can operate with greater autonomy. A decision that changes a system of record, moves money, affects a customer, or creates material downstream impact requires much stronger controls and potentially explicit human approval.

The important distinction is that human oversight does not mean a person has to approve every individual action. At scale, that would defeat much of the value of autonomy. The human role shifts higher. We define the boundaries, decide how much authority we are willing to delegate, watch the outcomes, and step in when the system moves outside those boundaries.

That is why traceability is so important. The organization should always be able to explain how the decision was made and why the agent was allowed to make it.

So machines can increasingly participate in decisions and execution, but accountability should remain anchored to named human roles and the enterprise that chose to delegate that authority.