Skip to main content

AI Is Everywhere. So Why Is Business Value Still So Hard to Capture?

on
AI Decision Architecture Flowchart

AI Is Everywhere. So Why Is Business Value Still So Hard to Capture?

AI is everywhere now.

Marketing teams use it.

Developers use it.

Finance teams use it.

Customer support uses it.

Agents are beginning to do more than answer questions—they are starting to take actions.

Investment keeps climbing.

Models keep improving.

And yet something still does not quite add up.

McKinsey’s research points to a stubborn gap:

Companies are adopting AI much faster than they are capturing measurable business value from it.

That is the part I find most interesting.

Because if the technology is getting better, adoption is rising, and investment is increasing, then where is the value?

The obvious answer would be:

Give it time.

Maybe.

But I am not sure that explanation is enough.

The deeper problem may be that enterprise AI is moving into a new phase.

The bottleneck may no longer be intelligence.

It may be management.

AI capability is becoming abundant

For the first few years of the generative AI boom, the central question was easy to understand:

What can AI do?

Companies experimented with everything:

  • chatbots
  • copilots
  • document summarization
  • code generation
  • retrieval systems
  • forecasting
  • content creation
  • customer support
  • workflow automation

That made sense.

The technology was changing so quickly that experimentation itself had value.

But the environment is different now.

Powerful AI is becoming available across an expanding universe of vendors, models, platforms, and open-source ecosystems.

McKinsey estimates that AI attracted roughly $124 billion in equity investment in 2024, while agentic AI experienced the fastest percentage growth among the technology trends it tracked.

Access to capable AI is becoming easier.

And that changes the competitive question.

When everyone can buy intelligence, intelligence itself becomes less differentiating.

The harder question becomes:

What should the organization actually do with it?

Here is where the story gets strange

McKinsey’s broader research shows that more than three-quarters of organizations were already using AI in at least one business function.

Yet more than 80% were not reporting tangible enterprise-level EBIT impact from generative AI.

And only about 1% of executives described their deployments as mature.

That is quite a gap.

AI can be working at the task level while the economics remain stubbornly unimpressed.

  • A sales team saves time.
  • A developer writes code faster.
  • A support agent handles more requests.
  • A marketing team generates more content.

Useful?

Absolutely.

But useful is not the same as profitable.

And that distinction is where a lot of AI enthusiasm starts to run into operating reality.

Adoption is not value

Imagine a company that looks highly advanced on paper.

Marketing uses generative AI. Finance uses forecasting models. Sales has copilots. Engineering uses coding assistants. Operations launches autonomous agents. Customer support deploys conversational AI.

The AI activity is everywhere.

Fantastic.

Now imagine the CEO asks:

Which of these systems is actually creating value?

Things get quieter.

Then:

  • Which ones deserve more investment?
  • Which ones are getting too expensive?
  • Which workflows need redesign?
  • Which systems are creating unacceptable risk?
  • Which decisions should AI make?
  • Which should remain human?
  • Which projects should stop?

A dashboard showing monthly AI usage will not answer those questions.

Usage tells you that people are using AI. It does not tell you whether the organization is becoming better because of it.

That is a management problem.

The bottleneck is moving

I think enterprise AI is moving through three broad stages.

Stage 1 — AI Capability

The question was: What can AI do?

This was the era of models, prompts, copilots, proofs of concept, and experimentation.

Stage 2 — AI Workflow

The question became: How can AI perform real work?

Now we get agents, tool use, automation, workflow integration, and human-AI collaboration.

This is where much of the market is today.

Stage 3 — AI Management

And now a third question is emerging: How do we manage hundreds of AI-enabled decisions and workflows?

That means deciding:

  • what to automate
  • what AI may recommend
  • what AI may execute
  • what humans must approve
  • what data can be trusted
  • what the workflow should cost
  • what constitutes acceptable risk
  • what success looks like
  • what deserves more capital
  • what should be redesigned
  • what should stop

That is much bigger than model selection. It is decision architecture.

McKinsey’s emphasis on agentic AI makes this shift especially important because agents do not merely produce outputs. They can plan actions, use tools, interact with enterprise systems, and execute workflows.

Once AI starts acting, “Was the answer good?” is no longer enough.

Now we need to ask: Was the action permitted?

Smarter AI does not automatically create better decisions

Consider an autonomous marketing agent.

It analyzes conversion rates, customer segments, ad performance, and competitor activity. Then it recommends: Increase campaign spending by 30%.

Maybe it is right.

But before anyone moves the money, a few inconvenient questions appear.

  • Is ROI above the required threshold?
  • Is the campaign already over budget?
  • Is the underlying data reliable?
  • Is the recommendation consistent with company policy?
  • Does a 30% budget increase require human approval?
  • Has performance been improving—or quietly deteriorating?

The AI may be excellent at analysis. It still does not get to invent the company’s spending authority.

That has to come from somewhere else: business rules, thresholds, permissions, ownership, escalation logic, governance.

The more autonomous AI becomes, the more valuable those boundaries become.

Intelligence is one thing. Authority is another.

This is the distinction I think matters most. AI can tell you what it thinks should happen. The organization still needs a system for deciding whether that recommendation becomes action.

One way to think about the architecture is:

Data
↓
AI / Models / Agents
↓
Decision System
↓
Metrics → Rules → Thresholds → Risk → Cost → ROI → Confidence
↓
Decision
↓
Scale | Continue | Govern | Redesign | Route | Pause | Stop
↓
Action
↓
Results return to the system

AI supplies intelligence. The Decision System supplies context, boundaries, and accountability. That is how probabilistic intelligence becomes governed action.

McKinsey’s research keeps pointing in this direction

One of McKinsey’s more revealing findings is that the adoption practice most strongly associated with bottom-line impact was tracking well-defined KPIs for generative AI solutions.

That sounds almost boring. Which is precisely why it matters.

The glamorous part of AI is the model. The less glamorous part is asking:

  • What outcome are we trying to improve?
  • What was the baseline?
  • What is the KPI?
  • What does success look like?
  • What is the system costing us?
  • When do we scale?
  • When do we stop?

The model can be impressive while the management system around it remains immature.

And McKinsey’s emphasis on workflow redesign points to the same issue: organizations cannot simply lay AI over existing processes and expect transformation to appear automatically.

A faster bad workflow is still a bad workflow. Sometimes it is just a bad workflow moving at impressive speed.

Agents make the problem harder

A chatbot generates information. An agent acts. That is a much bigger jump than it sounds.

Once AI starts taking actions, organizations need answers to questions like:

  • How much can it spend?
  • What systems may it access?
  • What may it change?
  • What confidence level is required?
  • When does a human have to step in?
  • What happens when the system behaves unexpectedly?

Governance cannot live only in a PDF somewhere. It increasingly has to live inside the workflow itself.

This is the layer I have been exploring through Decision Systems

At Decision Systems AI, I have been building what I call Decision Systems Orchestrators around this gap.

The systems combine:

  • business metrics
  • data validation
  • deterministic rules
  • thresholds
  • risk
  • cost
  • ROI
  • confidence
  • escalation logic
  • human review

The point is not to make the model smarter. The point is to make the decision around the model explicit and accountable.

That becomes much more important when an organization moves from a handful of pilots to dozens—or eventually hundreds—of AI-enabled workflows.

The next competitive advantage may be management

The first phase of enterprise AI rewarded organizations that experimented quickly. The next phase may reward something less glamorous but more valuable: operating discipline.

Knowing:

  • where AI is creating value
  • where it is not
  • where automation makes sense
  • where human judgment still matters
  • where cost is getting away from you
  • where risk is rising
  • where workflows need redesign
  • where capital should move

As AI capability becomes cheaper and more widely available, access to intelligence becomes less special. What becomes more valuable is the ability to allocate, govern, and act on that intelligence well.

Which leaves executives with a different question.

Not: How much AI are we using?

But: Do we have a system for deciding whether our AI is actually creating value—and what we should do next?

That may be where the next phase of enterprise AI really begins.