AI Works Better When It Optimizes the Outcome, Not the Intermediate Step

Artificial intelligence initiatives often begin with a familiar question: How can AI perform an existing task faster or more accurately? New research highlighted by Kellogg Insight suggests leaders may benefit from asking a more fundamental question: Does AI need to perform that task at all?

In “How a Different AI Approach Can Cut the Queue,” writer Emily Stone examines research by Simrita Singh, Itai Gurvich, and Jan A. Van Mieghem on the use of machine learning in hospital triage. Their work challenges a common assumption about AI implementation: that technology should replicate the intermediate steps humans traditionally use to reach a decision.

The Overarching Theme

The researchers compared two approaches to prioritizing patients from chest X-rays. In the conventional diagnosis-first approach, AI first predicts a patient’s condition and then translates that diagnosis into a priority level. Their alternative direct-to-queue model skips the preliminary diagnosis and uses features of the X-ray to determine the patient’s priority directly.

The business lesson extends well beyond healthcare: when deploying AI, optimize the model for the outcome the organization actually values—not necessarily for the process humans have historically followed to get there.

Major Takeaways for Business Leaders

1. Don’t automatically teach AI to imitate the existing workflow.
Organizations frequently digitize or automate processes step by step. But some intermediate steps may exist because humans needed them—not because they are essential to the final outcome. The research demonstrates the potential value of redesigning a process around AI rather than simply inserting AI into the existing process.

2. Define the real objective before selecting the AI metric.
For hospital triage, perfect diagnostic classification wasn’t the ultimate goal. The operational goal was getting the right patients treated at the right time while minimizing the consequences of waiting. As the researchers emphasize, operational context should shape how machine-learning systems are designed.

3. Removing an intermediate prediction can improve performance.
The researchers’ theoretical analysis found that the diagnosis-first and direct approaches perform equally only when the disease type can be predicted with 100 percent accuracy. Otherwise, the direct approach performs better at minimizing total waiting costs.

4. The improvement was substantial.
The team tested its approach using 112,120 anonymized chest X-rays, sorting cases into four urgency levels. In the simulated radiology setting, the direct-to-queue approach reduced total waiting costs by more than 30 percent compared with the diagnosis-first approach.

5. The principle applies beyond healthcare.
The authors point to customer service and bank call-center queues as potential applications. More broadly, the concept could be relevant anywhere organizations use predictions, classifications, scores, or customer segments as intermediate steps before making an operational decision.

Leadership Talking Points

• Are we using AI to automate our current process, or to improve our desired outcome?
• Which steps in our workflows exist because humans historically needed them?
• Are our AI teams optimizing technical measures such as prediction accuracy when the business actually cares about cost, speed, risk, retention, or customer outcomes?
• Could eliminating an intermediate classification or decision improve both speed and performance?
• What would our AI architecture look like if we started with the business objective and worked backward?

Reflection Questions

1. What business outcome are we ultimately trying to optimize with AI?
2. Are our current AI performance metrics directly connected to that outcome?
3. Which intermediate steps in our processes might be unnecessary in an AI-enabled workflow?
4. Where could errors compound because one AI prediction feeds another decision?
5. Do our AI teams have enough operational and domain expertise to understand the consequences of optimizing the wrong objective?
6. Where in our organization could a “direct-to-outcome” experiment be safely tested?

Potential Action Items

Map one high-volume decision process. Identify its inputs, intermediate classifications, decisions, outputs, and measurable business outcome.

Separate process metrics from outcome metrics. For example, distinguish “classification accuracy” from measures such as customer waiting time, revenue, cost-to-serve, risk, or service-level performance.

Challenge every intermediate step. Ask whether it is genuinely necessary to achieve the outcome or merely inherited from the traditional human workflow.

Run a controlled comparison. Where appropriate and responsibly governed, compare the existing multistep approach with an AI model optimized more directly for the desired outcome.

Bring operations leaders into AI design. This research reinforces the idea that AI implementation isn’t solely a data science problem. Business leaders and domain experts need to help define what the model should optimize.

Maintain appropriate human oversight. Particularly in healthcare and other high-stakes settings, operational efficiency should be considered alongside safety, fairness, explainability, clinical or professional judgment, and regulatory requirements.

Why This Matters

Perhaps the most important implication is that AI transformation and process automation are not the same thing.

Automating a flawed or unnecessarily complicated workflow can simply produce the same decisions faster. AI creates a more interesting opportunity: reconsidering the architecture of the decision itself.

That shifts the leadership conversation from “Where can we add AI?” to “What outcome are we actually trying to achieve, and what is the simplest responsible path to it?”

Recommended Reading

For leaders who want to explore the idea further, several related Kellogg Insight pieces are particularly relevant:

“Do You Really Need All That Data?” explores a complementary principle: more data isn’t automatically better. Research highlighted by Kellogg shows how algorithms can identify the information actually required to reach an optimal decision, potentially reducing both computation and data-collection costs.

“The Vicious Cycle of Long Wait Times” examines queue dynamics from another angle, showing how longer waits can change customer behavior and make congestion worse—an important reminder that operational systems can contain feedback loops that aren’t obvious from individual transactions.

“Podcast: To Thrive Alongside AI, Find the Bottlenecks” explores Kellogg economist Benjamin Jones’s argument that the limiting factor in a system often determines its overall performance. That makes identifying bottlenecks potentially more valuable than simply improving tasks AI already performs well.

“Podcast: Automation, Answers, and Advice—a Playbook for AI Adoption” offers a broader leadership perspective on moving beyond experimentation toward more effective organizational use of AI.