WHY I’M THINKING ABOUT THIS

Aperture AI started from a simple frustration: organizations can count AI usage far more easily than they can explain AI value. I wanted to make the decision layer visible — what deserves to scale, what needs fixing, and what should stop.

THE SIGNAL

Enterprise AI adoption has moved rapidly. McKinsey’s 2025 State of AI survey reported that 78% of respondents said their organizations used AI in at least one business function, and 71% reported regular use of generative AI in at least one function.

But only 21% of respondents reporting gen-AI use said their organizations had fundamentally redesigned at least some workflows.

That gap matters.

Buying AI is not the same thing as redesigning work around AI.

McKinsey also found that workflow redesign had the strongest relationship, among the attributes it tested, with seeing EBIT impact from generative AI.

THE QUESTION

When a leadership team asks:

“Is our AI investment working?”

what evidence should actually answer that question?

Tokens do not answer it.
Seats do not answer it.
Agent runs do not answer it.
Even daily active users do not answer it.

Those are activity signals.

The business question is whether the activity produced a better outcome.

THE OBVIOUS ANSWER

Most dashboards start with what is easy to collect:

These metrics are useful for operations and cost management.

They are not ROI.

An AI copilot can have high usage because employees are forced to use it. An agent can complete thousands of runs while producing low-quality outputs that humans silently repair. A product can reduce task time while creating downstream review cost.

THE TENSION

The enterprise needs a chain of evidence:

AI asset → use case → owner → usage → quality → risk → cost → business outcome

If one part is missing, the interpretation becomes weak.

I would go one step further and require every material AI product to have an Outcome Contract:

Outcome Contract field Example
Business problem Support resolution is too slow
Baseline 18 minutes median handling time
AI intervention Grounded support copilot
Expected outcome Reduce median handling time without increasing repeat contact
Quality guardrail Unsupported-claim rate below release threshold
Owner VP Support
Review window 60 days
Evidence source Ticketing + eval + finance data

Now the organization can ask something meaningful:

Did this product improve the outcome it was funded to improve?

Usage is not ROI supporting data visual
Secondary-research visual. Source and interpretation are stated inside the chart and article.
Usage is not ROI governance supporting visual
Supporting enterprise governance evidence. Populations differ; the article keeps those distinctions explicit.

MY PRODUCT TAKE

I think AI governance and AI economics are converging.

As enterprises add copilots, agents, APIs and embedded models, the operating problem becomes less about “Do we have AI?” and more about:

IBM’s 2025 breach research adds another reason the inventory matters: 63% of surveyed organizations lacked AI governance policies, and among organizations reporting an AI-related security incident, 97% lacked proper AI access controls. Those are different populations, but both point to the cost of letting an AI estate grow faster than visibility and control.

WHAT I WOULD TEST

For one enterprise, I would start with only five to ten AI assets.

For each one, reconcile:

vendor bill + internal usage + eval evidence + owner + business baseline

Then ask leadership to make a scale/fix/stop decision.

If the data does not materially improve the decision, we are collecting the wrong data.

WHAT WOULD CHANGE MY MIND

If organizations can make equally good investment decisions by joining their existing FinOps, observability and BI tools with minimal effort, a new AI decision layer may not deserve to exist.

That is why I would start with decision quality—not with a dashboard.

Secondary research