AI Creates Value Only When the Operating Model Is Ready

What the operating model includes

An operating model is the practical answer to how work gets done. For any process AI might touch, it covers:

• Process: the steps, handoffs and points where work waits.
• Roles and decision rights: who does each step and who can approve, override or escalate.
• Data ownership: who is responsible for the quality and meaning of the data the process depends on.
• Controls: the checks that keep the work accurate, compliant and safe.
• Measures: how the business knows the process is working.
• Support: who keeps the tools running and improving after launch.

When these are clear, AI has a defined place to fit. When they are not, AI inherits the confusion and often makes it more visible.

Where AI meets the operating model

Every AI use raises the same operating questions, whether it drafts documents, classifies requests or recommends a next action:

• Where exactly in the process does the AI output appear, and what happens next?
• Who is accountable for a decision that uses the output?
• When should a person review, correct or reject it?
• What happens when the AI is unsure, wrong or unavailable?
• Who owns the data it reads and the records it updates?

If nobody can answer these, the initiative is not ready, however good the technology looks.

Signs the operating model is not ready

• The process the AI will support is not documented, or is done differently by each team.
• Nobody in the business owns the outcome; the initiative belongs to IT or a vendor.
• Key data lives in spreadsheets or in systems that disagree with each other.
• There is no agreed way to handle exceptions today, so there will not be one for AI either.
• Success is described as "launch the tool" rather than a change in a business measure.

Five tests before an AI pilot

1. 1Business outcome: Define the workflow, baseline time/cost/error rate and decision to improve.
2. 2Ownership: Identify process owner, system owner, data steward and accountable human decision-maker.
3. 3Readiness: Assess source quality, integration, permissions, privacy and exception volume.
4. 4Controls: Document model evaluation, human review, logging, escalation and fallback to non-AI operation.
5. 5Economics: Compare implementation, ongoing review and failure-management costs against plausible benefits.

Getting the operating model ready

Preparing the ground does not have to be a large program. For a single use case, it usually means:

1. 1Map the current process as it actually runs, including workarounds.
2. 2Simplify it before adding AI. Remove steps that exist only because of old problems.
3. 3Name a business owner for the outcome and agree the measures, with a baseline.
4. 4Define the points where people review, approve or override the AI output.
5. 5Confirm who owns the data and fix the most important quality gaps.
6. 6Plan support, monitoring and improvement after launch, not only the launch itself.

This work often delivers value on its own. A clearer process with better data can remove part of the problem before any AI is involved, and makes the AI that follows easier to trust.

An illustrative example

A distribution team wants an AI assistant to classify customer inquiries. Before choosing a model, map the intake workflow, identify common exception classes, set response-quality acceptance tests and determine which cases require employee review. Pilot against the existing process; compare total handling time and rework, not just the speed of classification.

What to measure

Useful measures include:

• Cycle time
• Accuracy against approved labels
• Human override rate
• Exception rate
• Cost per completed case
• Customer impact

Questions for leadership

• For each AI initiative, which process does it change and who owns that process?
• Have we agreed how we will measure the result, and do we know today's baseline?
• Where will people stay in the loop, and is that written into how the work is done?
• Who will own the AI capability after the project team moves on?