Skip to main content
Back to the Iridium Blog
Read3 minIridium Team

The AI ROI gap is becoming a management problem, not a model problem

Why realised AI value depends on workflow design, adoption and measurement—not model choice alone.

Many scattered AI activity signals resolve into a smaller set of measured business outcomes.

Enterprise leaders have spent the last few years asking whether AI works.

Increasingly, that is not the most useful question.

The more interesting question is why some organisations are generating meaningful returns while others remain stuck in pilots.

Recent research is beginning to reveal a pattern.

Contentstack’s 2026 survey of enterprise digital leaders found that only 48% had clearly defined KPIs and were actively measuring agentic AI against them. Yet among that group, 94% reported measurable positive returns from internal agentic programmes.

Salesforce’s 2026 enterprise research similarly argues that the organisations generating returns are not simply moving faster. They are doing more of the foundational work around trustworthy data, human involvement and guardrails.

Gartner’s analysis points towards specialised agents built around specific business processes.

The implication is important.

The ROI gap is increasingly an operating-model gap.

“Productivity” is too vague

A generic AI assistant can make many employees feel more productive while producing very little measurable enterprise value.

That is not because the benefit is imaginary.

It is because individual productivity is difficult to aggregate.

If someone writes an email ten minutes faster, what happened to the ten minutes?

Did it become additional output?

Was another bottleneck removed?

Did revenue increase?

Did cycle time fall?

This is why workflow-level measurement is so powerful.

A business can measure:

  • average handling time before and after;
  • cases completed per person;
  • approval cycle time;
  • number of manual handoffs;
  • rework rate;
  • error rate;
  • cost per completed transaction.

Now AI has an economic baseline.

The best agent use cases are often boring

Enterprise AI conversations naturally gravitate toward impressive demonstrations.

The biggest returns may come from much less glamorous work.

Checking.

Reconciling.

Classifying.

Searching.

Copying.

Preparing.

Chasing.

Updating.

These tasks consume huge amounts of organisational time precisely because they sit between systems and responsibilities.

They are also ideal for agents because they have observable outcomes.

A good enterprise AI programme should therefore maintain a portfolio of workflows ranked by:

  1. volume;
  2. human time consumed;
  3. process clarity;
  4. information availability;
  5. economic value;
  6. risk;
  7. measurability.

That is a much stronger starting point than collecting AI ideas from across the company.

Value comes from the surrounding system

A model is only one component.

Real returns depend on whether the agent has:

  • reliable organisational knowledge;
  • persistent context;
  • controlled system access;
  • a defined workflow;
  • appropriate approval;
  • measurable outcomes.

This is why enterprises using the same frontier models can achieve very different results.

The differentiator is not necessarily intelligence in the model.

It is intelligence in the implementation.

Sources