Enterprise AI has entered an interesting phase. The technology is becoming more capable. Organisations are deploying more AI. Investment continues to increase. But the financial returns are not necessarily increasing at the same rate.
McKinsey's 2026 State of AI research captures the disconnect. Eighty percent of respondents say AI has improved their individual productivity. Yet only 37% say AI has contributed positively to their organisation's EBIT — essentially unchanged from the previous year. (McKinsey & Company)
At the same time, 60% expect their organisations to increase AI investment over the next year. (McKinsey & Company)
This raises an increasingly important question for enterprise leaders:
If we're deploying more AI, where is the corresponding business value?
Part of the answer may lie in how organisations think about AI economics. Too often, the objective becomes maximising what AI can do: more agents, more use cases, more autonomous steps, more AI embedded across the enterprise.
But maximum AI does not necessarily mean maximum ROI.
The better objective is to design the most effective business process — and use AI only where it improves the economics or outcome of that process.
Productivity isn't the same as enterprise value
AI can produce an impressive productivity result without materially improving the economics of the business.
Imagine an AI agent reduces a 20-minute activity to two minutes. Measured at the task level, that's a significant improvement.
But what if:
o the task represents only 10% of a larger process?
o the output still requires manual review?
o another team needs to re-enter the information into a different system?
o exceptions increase?
o the process still spends two days waiting for an approval?
The agent may have become dramatically more productive while the economics of the end-to-end process barely changed.
This is one reason AI ROI needs to be considered beyond the individual model, agent or task.
McKinsey's latest research reinforces this distinction. While individual productivity gains are widespread, enterprise financial impact remains much more concentrated. Just 6% of respondents qualify as AI "high performers" – organisations attributing at least 5% of EBIT to AI while also reporting significant value from it. (McKinsey & Company)
The question therefore shouldn't simply be: How productive is our AI?
It should be: How much better is the business process because of it?
AI isn't free labour
There is another factor that becomes increasingly important as organisations move from pilots into agentic AI production: operating cost.
An agent performing a task once in a demonstration is very different from an agent performing thousands or millions of actions across an enterprise.
AI introduces consumption-based costs. Models consume tokens. Agents may make multiple model calls to complete a single process. More complex reasoning can require more computation. Enterprise deployments also require infrastructure, integration, monitoring, security, observability and ongoing management.
At small volumes, these costs may appear insignificant. At enterprise scale, they become part of the unit economics of the process.
McKinsey's 2026 survey found that one in five organisations has already constrained its use of AI because of AI-related operating costs, including token costs. (McKinsey & Company)
That doesn't mean AI is too expensive. It means AI has an economic profile that needs to be designed and managed like any other enterprise capability.
And that changes the question from Can AI perform this activity? to Is AI the most valuable way to perform it?
Not every step needs intelligence
Consider this example of a process containing 100 activities:
- 60 are highly predictable. Given the same input, they should produce the same outcome every time.
- 25 require interpretation or reasoning.
- 15 require judgement, approval or accountability from a person.
It may be technically possible to apply AI across most of that process. But economically, it may make little sense.
The predictable activities may be better handled by deterministic automation: fast, repeatable and inexpensive.
Agentic AI may be most valuable for the activities where context, interpretation, reasoning or adaptability genuinely changes the outcome.
And people may need to remain responsible for decisions where judgement, risk or accountability matters.
The optimal process could therefore look something like:
Deterministic automation → Agentic AI → Automation → Human decision → Agentic AI → Automated execution
rather than:
AI agent → AI agent → AI agent → AI agent
The goal isn't to maximise the amount of AI inside the process. The goal is to maximise the performance and economics of the process itself.
The cost of the agent isn't the cost of the process
This distinction also changes how organisations should calculate agentic AI ROI. A narrow calculation might compare the cost of an AI agent with the labour cost associated with the task it performs. That can be useful, but it doesn't tell the whole story.
Enterprise leaders need to consider the economics of the complete process. That includes questions such as:
• Has total cycle time decreased?
• Has throughput increased?
• Has the cost per completed process fallen?
• Has human effort actually been removed or simply moved somewhere else?
• Have exception rates changed?
• Has quality improved?
• Has rework decreased?
• Has customer experience improved?
• What does the AI cost to operate at production volume?
• What additional infrastructure and governance does it require?
• And ultimately, has the process created measurable financial value?
This matters particularly for CFOs and COOs because the apparent efficiency of an AI task can disappear once downstream costs and operational dependencies are included.
A five-minute saving isn't valuable if it creates ten minutes of work elsewhere.
High performers redesign the workflow
The organisations seeing the greatest financial impact from AI also appear to be approaching the problem differently. McKinsey found that nearly three-quarters of AI high performers are fundamentally redesigning workflows because of AI, compared with only around one-quarter of other organisations. (McKinsey & Company)
The difference is not simply how much AI they deploy, but how they redesign work around it. Rather than inserting AI into existing processes, they reconsider the workflow end to end, identifying where automation is most effective, where governed agentic AI[SM6.1] adds value, and where human judgement and accountability should remain.
This shifts the focus from making individual tasks faster to designing a better-performing business process overall.
Recent McKinsey research into AI operating models makes a similar point: organisations capturing value are more likely to focus on end-to-end domains and cross-functional workflows rather than disconnected AI use cases. (McKinsey & Company)
That might mean removing a step entirely.
Eliminating a hand-off.
Automating a predictable decision rather than asking AI to make it.
Changing when a person becomes involved.
Allowing an agent to gather and interpret information but requiring deterministic controls before an action occurs.
Or redesigning the workflow so information moves directly between systems instead of reproducing an existing manual process with AI layered on top.
This is where the economics can change materially.
Measure the outcome, not the amount of AI
There is a temptation during emerging technology cycles to use adoption as evidence of progress:
o Number of AI pilots.
o Number of agents deployed.
o Percentage of employees using AI.
o Number of processes containing AI.
Those measures can tell you whether adoption is occurring. They don't necessarily tell you whether value is being created.
For an enterprise AI program, a better scorecard may look at:
a) Business outcome
Revenue generated, cost removed, risk reduced or customer outcome improved.
b) Process performance
Cycle time, throughput, quality, exceptions and rework.
c) Human effort
How much work has genuinely been removed, augmented or redirected?
d) AI operating cost
What does the capability cost per completed process at production scale?
e) Reliability
How consistently does the process deliver the intended outcome?
f) Control
Can the organisation understand, govern and audit what happens? Effective AI governance should define how agents operate within the broader architecture, including what they can access, what actions they can take and where human oversight is required.
Together, these measures provide a much more useful picture of agentic AI ROI than the performance of the agent alone.
Start with the economics of the process
The next phase of enterprise AI isn't about finding more places to deploy AI. It's about becoming more selective about where AI creates disproportionate value.
Recommended Steps:
1. Start with the process.
2. Understand its current economics.
3. Identify where time, cost, complexity and risk actually sit.
4. Then determine which parts should be automated, which genuinely benefit from agentic reasoning, and which require human judgement.
Only then should the organisation decide how much AI the process needs. Because as agentic AI becomes easier to deploy, the competitive advantage won't come from simply having more agents. It will come from knowing where intelligence creates value… and where it doesn't.
Maximum AI isn't the objective. Maximum business value is.
Is your agentic AI investment delivering measurable business value?
Clear Dynamics works with organisations to identify high-value enterprise processes and determine the right combination of automation, agentic AI and human judgement to improve performance safely and economically.
Join our upcoming webinar: Enterprise AI: why start with the process, not the agent.
