AI is an operating-model question first
Why the winners are redesigning decision rights before they deploy models.
Most organisations began their AI programmes the same way: a portfolio of pilots, a centre of excellence and a steady stream of promising demonstrations. Far fewer can point to a material change in their cost base, their growth or the quality of their decisions.
The gap is rarely the technology. The models available today are capable enough for most of the use cases leaders care about. What stalls value is the organisation around the model: who is allowed to act on its output, how work is redesigned around it, and who is accountable when it is wrong.
Pilots prove the model, not the business
A pilot answers a narrow question: can the model do the task? It seldom answers the questions that decide whether value reaches the profit and loss account. Will the underwriter trust the recommendation? Will the planner change the order? Will a regional manager accept that a decision once made locally is now made centrally, or the other way round?
When those questions go unasked, the model is layered on top of existing work. People review its output, then do what they would have done anyway. The organisation pays for the technology and keeps the old costs.
Start with decision rights
The organisations getting the most from AI begin by mapping the decisions that matter, not the tasks that can be automated. For each decision they ask who makes it today, with what information, how often and at what cost when it is wrong. Only then do they ask where a model could make that decision faster, more consistently or closer to the customer.
This reframing usually leads to changes that have little to do with technology. Approval limits move. Layers of review disappear. Some decisions move to the front line because the model gives people the information they previously had to escalate for. Others move to the centre because consistency now matters more than local judgement.
- Model recommends
- Analyst reviews
- Manager approves
- Decision, as before
Cost of the technology is added; old costs stay.
- Model recommends
- Front line decides within clear limits
- Overrides logged and fed back
- Named owner monitors performance
Fewer layers, faster decisions, value reaches the P&L.
Redesign the work, then the roles
Once decision rights are clear, the work can be redesigned around them. That means defining when a person must review the model, when they may override it, and how those overrides feed back into improving it. It also means being honest about which roles change, which grow and which are no longer needed, and investing early in the skills people will need to work alongside the new tools.
Govern for trust
Finally, trust has to be designed in rather than hoped for. Leaders need clear measures of model performance, a named owner for each model in production and a simple route for raising concerns. Boards need to understand where AI is making or shaping consequential decisions, and what safeguards apply.
Technology determines what is possible. The operating model determines what is captured.
Three questions for the executive team
Which five decisions, if made better or faster, would move our results most? Who owns each of them today, and would that change if a model were involved? What would we need to believe about accuracy and accountability to let the model act without a person in the loop? The answers are the foundation of an AI strategy that pays for itself.
What would change our view
If companies that deployed AI without changing decision rights captured as much value as those that redesigned them, the operating model would matter less than we argue.
This piece is RavenArc analysis. It draws on established management practice rather than new data, and it cites no specific figures.
RavenArc tests every decision against six questions. See the RavenArc Decision Method.
One considered idea, straight to your inbox.
A short note when we publish something worth a leader’s time. No noise, and you can leave with one click.