# stackpole-mit-sloan-ai-decision-matrix-2026-09-22

## Veille

Article by **Beth Stackpole** in *Ideas Made to Matter* (**MIT Sloan**), published on **September 22, 2026** (~5 min), presenting the *research briefing* « Designing Decision Rights for AI » by **Ina Sebastian**, **Peter Weill**, **Thomas Haskamp** and **Jan vom Brocke** (**MIT CISR**), built from interviews with **30 executives**.

## Titre Article

A framework for determining when AI can make decisions

## Date

2026-09-22

## URL

https://mitsloan.mit.edu/ideas-made-to-matter/a-framework-determining-when-ai-can-make-decisions

## Keywords

decision rights, decision rights, AI Decision Matrix, ambiguity, risk, framing acting learning, framing, acting, learning, routine decision, consequential decision, exploratory decision, strategic decision, human oversight, human in the loop, AI agents, One New Zealand, AI center of excellence, agent owner, portfolio of decisions, MIT CISR, AI governance

## Authors

Beth Stackpole — rédactrice, *MIT Sloan Ideas Made to Matter* ; recherche de Ina Sebastian, Peter Weill, Thomas Haskamp, Jan vom Brocke (MIT CISR).

## Ton

Popularization article of academic research for an executive readership. It opens with a "What you need to know" box, defines the two dimensions and three components, then walks through the four types, each with an example drawn from a single company, One New Zealand. The register is prescriptive without being peremptory: human oversight is maintained across all categories, and AI autonomy expands as trust is built. Researcher quotes frame the piece. The text does not give aggregate figures on the 30 interviews and refers readers to the original briefing for methodology.

## Pense-betes

- **Low ambiguity** = repeatable, predictable decisions; **high** = multiple possible interpretations.
- **Low risk** = limited, reversible impact; **high** = financial, operational or reputational impact, difficult to correct.
- **Framing**: problem, assumptions, stakeholders, constraints, success criteria. **Acting**: gathering, evaluating, recommending, authorizing, executing. **Learning**: tracking outcomes, responding to them, updating the system.
- **Routine**: framing codified in advance, action and learning automated; humans stay close while trust is being built. Example: marketing segmentation into SQL by an agent, **−60%** time, human review before launch.
- **Consequential**: humans on execution, exceptions, learning. Example: customer service; inaccurate pricing information is a *"nonnegotiable constraint"* → human in the loop for anything touching pricing.
- **Exploratory**: framing co-built by humans and AI; governance takes priority as experimentation moves toward riskier cases. Example: marketing content creation.
- **Strategic**: humans lead framing and learning, AI facilitates action. Example: network, **15 task agents** + orchestration agents; human decisions (reliability, customer experience, cost).
- **Practices**: business units frame each case; AI and data center of excellence (unified platform, horizontal orchestration layer, responsible AI policy); human owner named per agent; manage AI as a **portfolio of decisions**.
- Connects to [[chen-reneau-flansburg-clef-decision-models-2026-10-01]] on decision modeling.

## RésuméDe400mots

Which tasks should be delegated to artificial intelligence, and when should humans remain involved? For Beth Stackpole, writing in MIT Sloan Ideas Made to Matter about work from the MIT Center for Information Systems Research, designing these decision rights is one of the major challenges facing companies. Ina Sebastian, Peter Weill, Thomas Haskamp and Jan vom Brocke built the AI Decision Matrix from interviews with 30 executives, on one premise: not all decisions are equal.

The matrix crosses two dimensions. Ambiguity measures the extent to which data determines the answer and the extent to which people agree on it. Risk measures the consequences of an error and the difficulty of correcting it. Each decision has three components: framing (problem, assumptions, constraints, success criteria), acting (gathering, evaluating, recommending, authorizing, executing) and learning (tracking outcomes, responding to them, evolving the system).

The four resulting types call for different roles for AI, with human oversight throughout. Routine decisions lend themselves to automation: at telecom operator One New Zealand, an agent translates marketers' requests into SQL queries to create segments, cutting time spent by 60%, with human review before launch. Consequential decisions require balancing automation and intervention: in customer service, any decision touching pricing keeps a human in the loop, since a pricing error is deemed unacceptable. Exploratory decisions let framing evolve between humans and AI, as with marketing content creation, with governance monitored as experimentation moves toward riskier cases. Strategic decisions require human leadership: fifteen agents analyze network data and orchestration agents coordinate them, but trade-offs between reliability, customer experience and cost remain human.

The article draws four practices from One New Zealand: giving business units ownership of framing for each use case, creating an AI and data center of excellence (unified platform, orchestration layer, responsible AI policy), naming a human owner for each deployed agent, and managing AI as a portfolio of business decisions. Routine and consequential decisions deliver quick efficiency gains, while exploratory and strategic ones deliver longer-term value. The authors summarize the shift: moving from the question "where should AI be used?" to managing the decisions it shapes.

## GrapheDeConnaissance

- Beth Stackpole —publie→ article A framework for determining when AI can make decisions (DOCUMENT, 0.98)
- article A framework for determining when AI can make decisions —est_basé_sur→ Designing Decision Rights for AI (DOCUMENT, 0.96)
- Ina Sebastian —a_créé→ AI Decision Matrix (METHODOLOGIE, 0.96)
- Peter Weill —a_créé→ AI Decision Matrix (METHODOLOGIE, 0.96)
- MIT CISR —publie→ Designing Decision Rights for AI (DOCUMENT, 0.95)
- Peter Weill —dirige→ MIT CISR (ORGANISATION, 0.92)
- MIT CISR —mesure→ AI Decision Matrix construite à partir d'entretiens avec 30 dirigeants (MESURE, 0.92)
- AI Decision Matrix —utilise→ droits de décision (CONCEPT, 0.92)
- AI Decision Matrix —s_applique_à→ agents IA (TECHNOLOGIE, 0.9)
- Ina Sebastian —affirme_que→ la même capacité IA peut être sûre dans un contexte et risquée dans un autre (AFFIRMATION, 0.95)
- AI Decision Matrix —recommande→ automatiser les décisions routinières et garder un leadership humain sur les décisions stratégiques (AFFIRMATION, 0.94)
- One New Zealand —utilise→ agents IA (TECHNOLOGIE, 0.95)
- One New Zealand —mesure→ plus de 50 agents IA en production début 2026 (MESURE, 0.94)
- One New Zealand —mesure→ −60 % de temps de segmentation marketing via un agent text-to-SQL (MESURE, 0.93)
- One New Zealand —observé_dans→ AI Decision Matrix (METHODOLOGIE, 0.88)
- MIT CISR —recommande→ nommer un propriétaire humain pour chaque agent IA déployé (AFFIRMATION, 0.93)
- MIT CISR —recommande→ gérer l'IA comme un portefeuille de décisions métier plutôt que de cas d'usage (AFFIRMATION, 0.93)
- centre d'excellence IA —permet→ droits de décision (CONCEPT, 0.8)

---
Canonical: https://www.thekb.eu/en/fiches/stackpole-mit-sloan-ai-decision-matrix-2026-09-22/