MIT Framework Maps When AI Should Decide vs. When Humans Must

Determining which decisions to delegate to artificial intelligence and which require human judgment has become a critical governance challenge as companies deploy autonomous agents at scale. Researchers at the MIT Center for Information Systems Research have developed a practical framework to help business leaders navigate this boundary.

The AI Decision Matrix, created by Ina Sebastian, Peter Weill, Thomas Haskamp, and Jan vom Brocke, classifies decisions along two dimensions: ambiguity and risk. The framework emerged from interviews with 30 executives and recognizes that identical AI capabilities can be appropriate in one context and dangerous in another.

Why it matters

As companies move beyond pilot projects to deploy dozens or hundreds of AI agents, the question shifts from "Can AI do this?" to "Should AI do this?" Without clear decision rights, organizations risk either over-automating high-stakes choices or under-utilizing AI where it could safely accelerate operations. This framework provides a systematic method to make those distinctions before problems arise.

Two dimensions that determine decision rights

The matrix evaluates decisions based on ambiguity—how clearly data determines the answer and whether people agree on interpretation—and risk, meaning the consequences of error and difficulty of reversal. Low-ambiguity decisions are repeatable and predictable; high-ambiguity decisions allow multiple interpretations. Low-risk decisions have minimal impact if wrong; high-risk decisions carry significant financial, operational, or reputational consequences.

Every decision involves three components: framing (defining the problem and success criteria), acting (gathering information and executing), and learning (monitoring outcomes and updating the system). The framework prescribes different roles for humans and AI across these components depending on where a decision falls in the matrix.

Four decision types with distinct automation strategies

Routine decisions (low ambiguity, low risk) are strong automation candidates. Companies can codify framing in advance and automate action and learning while humans monitor performance. One New Zealand, a telecommunications provider operating more than 50 AI agents, uses automation to create audience segments for marketing campaigns, cutting segmentation time by 60 percent while marketers review plans before launch.

Consequential decisions (low ambiguity, high risk) require balancing automation with human intervention. One New Zealand deploys AI agents for customer service plan upgrades but mandates human involvement for all pricing decisions, which the company identified as a nonnegotiable constraint. Agents are introduced gradually with outcome validation before scaling.

Exploratory decisions (high ambiguity, low risk) involve interpretation and creativity where errors have limited impact. Humans oversee actions and learning while framing evolves through interaction. One New Zealand's marketing team uses agents for content creation, maintaining human oversight to revise mistakes before they reach customers while accelerating experimentation.

Strategic decisions (high ambiguity, high risk) demand human leadership in framing and learning while AI facilitates action. One New Zealand uses 15 task-based agents to analyze network data during power outages, with orchestration agents providing analysis to support human decision-making. Actual decisions remain human-led because they involve trade-offs between reliability, customer experience, cost, and infrastructure priorities.

Implementation guidance from the field

One New Zealand's approach offers lessons for other organizations. The company empowered business units with responsibility for framing each AI use case, ensuring agents address genuine business needs. It established an AI center of excellence with ownership of data foundations and architecture, including a unified data platform and responsible AI policy. Every deployed agent has a named human owner accountable for monitoring performance, refining data, and improving the agent over time.

The researchers emphasize that leaders should manage AI as a portfolio of business decisions rather than isolated use cases, looking beyond individual implementations to consider how AI affects core business processes. Companies that design decision rights systematically will move faster, reduce risk, and build trust in their AI deployments, according to the MIT Center for Information Systems Research team that developed the framework.