Start with the Decision not the Stack

When a company sets out to be “data-driven,” the instinct is often to start with the tech stack. But in my view, the organizations that struggle most have not clearly answered a more fundamental question: “What decisions are we making badly today, and could AI help us make those decisions better?”

In practice, the goal is usually not better data and instead comes down to wanting to make faster decisions with less risk, surface patterns human observation would likely miss or remove bias from resource allocation. Companies that have not defined the decisions they are seeking to address can find themselves with data capabilities that sit adjacent to the business rather than shaping its direction.

The Hidden Cost of Cheap Analysis

In many organizations, I’ve seen AI lower the cost of generating analysis, making clarity around decision definition even more important. Teams can now produce dashboards, scenario models and market research in a fraction of the time it once took, and that is powerful. But, in my view, it has helped introduce a subtle and counterproductive failure mode: analysis paralysis at scale.

When generating one more analysis adds little cost, there is often a reason to run one more scenario, build one more deck or pull one more dataset before committing to a decision. Under these conditions, the volume of available insight can begin to outpace an organization’s capacity to act on it. Leadership teams can find themselves struggling with AI-generated reports that each tell a slightly different story, creating the illusion of rigor while delaying the decisions that matter.

The irony is sharp: a tool designed to accelerate decision-making can end up doing the opposite when there is no discipline around how it is used. More analysis that is more cheaply produced does not automatically mean better decisions. It often means slower ones.

Moving from Analysis to Action

Avoiding these traps, in my view, requires more intentional analysis and discipline in how that analysis is tied to decisions. Below are four key practices that, in my experience, help teams stay focused on translating AI-driven insight into action:

• Define the decision before commissioning the analysis. Analytical effort should begin with a clear statement: “We are trying to decide X by Y date, and we need to understand Z to make that call.” If the analysis is not tied to a specific, time-bound decision, it should not be started. This eliminates a surprising amount of performative analytics.
• Establish a “good enough” threshold. Not every decision should require 95% confidence. A pricing change in a single market may not need the same analytical rigor as a platform migration. Being explicit about the level of certainty needed for different decision categories can help teams avoid over-investing in low-consequence choices.
• Separate decision deadlines from analysis timelines. When the deadline arrives, the team decides with whatever information is available. This can help force prioritization – teams quickly learn to focus their energy on the two or three questions that may move the needle, rather than producing exhaustive but unfocused analysis.
• Designate a single decision owner. This does not mean deciding alone. It means one person is accountable for driving the process, synthesizing the evidence and ensuring a decision gets made. The goal is not to exclude perspectives, but to ensure someone is responsible for the conclusion.
  ‍

Where to Automate, and Where Not To

The same focus is important when determining how and when to deploy AI within your organization. In my opinion, the framework is straightforward: automate processes whose value comes from efficiency, protect those whose value comes from experience.

We see AI excel at tasks with clear inputs and measurable outputs, such as data cleansing, report generation, anomaly flagging and routine customer queries. These are high-volume, low-judgement tasks where speed and consistency matter more than nuance. Automating them can free human attention for higher-value, more meaningful work.

But many high-value activities in a business are not purely efficient. They are often interpretive, relational and context-dependent: the consultative sales conversation, the creative brainstorm or the moment of genuine customer insight. These often live in areas where friction is a feature, not a bug. Importantly, that friction often contains information. A deal that stalls for reasons no one can articulate or a customer complaint that doesn’t fit the taxonomy – these are signals. Automating these moments away may not resolve the underlying issue and can, in some cases, obscure it. Protecting space for human judgement in these instances is not a failure in AI adoption. In my experience, it is a sign of organizational maturity.